I welcome everybody. We are fortunate today that we are joined by a number of organisations that can look at how public services interact with AI and that have been quite progressive in their approaches to technology. Certainly, we are looking forward to hearing from their experience and any lessons that can be learned more broadly across the public services. We have apologies from Senators Dee Ryan and Scahill this morning.

As experienced public servants, I am sure our guests are all very aware of the constitutional requirements of appearing before Oireachtas committees. I welcome them to our first public meeting in 2026. As everyone will understand, the role of our committee is to explore the opportunities and challenges presented by artificial intelligence. We are operating in a modular form, and we are looking at the deployment of AI within the public sector. Within the witnesses' organisations, they have looked at how AI can be used, so it is to be hoped they will be able to both share that experience plus make recommendations to us on the broader public service and how we, as legislators, can take responsibility for the roll-out of AI programmes.

I welcome from the Citizens Information Board, Graham Long, chief executive, and Elsa Flanagan, head of division for ICT, digital and data; from the Office of the Revenue Commissioners, Ruth Kennedy, Revenue Commissioner, and Keith Redmond, chief technology officer; and from the Department of Foreign Affairs and Trade, Síle Maguire, director general of corporate services division, Siobhán Byrne, director of the passport service; Eimear Friel, director of policy co-ordination and international technology, and Eddie McGinn, director of ICT. I thank them all for coming before us today.

I now invite Graham Long of the Citizens Information Board to deliver his opening statement.

On behalf of the Citizens Information Board, CIB, I thank the Chair and members for the invitation to address the committee on the role and impact of AI in the public sector. This is a topic that is shaping the future of how we deliver services and engage with members of the public.

CIB celebrated its 25th anniversary last June, while the Citizens Information Services that we fund and support celebrated their 50th anniversary in 2025. As a State agency, CIB is supported by the Department of Social Protection. Our vision is an Ireland where everyone has access to a trusted, independent source of information, advice and advocacy. This remit is not just a function, it is a cornerstone of democracy and social inclusion. Our work spans critical areas such as information on social welfare entitlements, employment rights and health services.

CIB’s reach is extensive and is very much rooted in communities. We fund and support a network of vital services. Alongside the Citizens Information Services, there is the Money Advice and Budgeting Service, the National Advocacy Service for People with Disabilities, the Sign Language Interpreting Service and the Register of Irish Sign Language Interpreters. These organisations can be lifelines for people facing challenges, making decisions and seeking clarity in moments that matter.

To give a sense of the level of activity, the Citizens Information Services assisted more than 384,000 people last year, with just under 709,000 queries. The citizens information phone service answered more than 111,000 calls. CIB runs the citizensinformation.ie website directly. Last year, the website welcomed more than 8.6 million users and recorded more than 27.4 million page views. CIB plays a pivotal role in providing independent, trusted information.

As we developed our current digital strategy, one message was very clear, which is that choice matters. People want the convenience of online services but not at the expense of face-to-face support. Our board is committed to preserving that human connection while embracing new and improved technologies. Our digital strategy reflects this balance, seeking to enhance services without losing the empathy and trust that define our work. One element of this is collaborating on an initiative from the Citizens Information Services to launch a digital help service, designed to help users access online public services confidently, while building digital literacy.

However, we face new challenges, particularly as AI tools become integrated into people’s digital lives. Alongside the use of AI to generate online content, the rapid rise of tools such as Google’s AI Overviews has already impacted traffic to citizensinformation.ie, with an almost 19% decline last year. These tools often extract and mix CIB’s content with other sources, some of which may be outdated or inaccurate. In response, we have been working with the Office of the Government Chief Information Officer to develop an AI tool called AskCI.

At the end of last month, we set AskCI live for users for a soft launch. This followed the establishment of a strong governance framework and a considerable amount of assessment and testing, including with the Citizens Information Services. Unlike generic AI systems, AskCI draws only from our verified content on citizensinformation.ie. It uses natural language processing to deliver clear, conversational answers, acting as a translator of trusted information. AskCI works in multiple languages and helps to make information more accessible for our users.

As of 13 January, AskCI had recorded 1,345 sessions and 2,922 questions. Areas of interest for users are broadly in line with the website, covering social welfare, money and tax, employment, moving country and housing.

CIB's digital content team is monitoring the accuracy and appropriateness of the responses generated by AskCI. We will be tracking AskCI over the next few months to see how it gains traction with users and how it sits alongside our website content. Future options for AskCI include a voice interface, the use of a pop-up overlay on the website or repurposing it as an app.

While the early results are encouraging, AskCI is simply a new way to explore our website content. It is not a substitute for in-depth research or for in-person and personalised information, advice or advocacy. The AI landscape for us is complex, with AI both a powerful way to generate misinformation and a powerful tool to interact with information generally. Like most organisations, we will be watching developments in this area very closely over the coming months and years.

I thank Mr. Long. I now invite Ms Kennedy to make an opening statement on behalf of the Revenue Commissioners.

I thank the Cathaoirleach for the opportunity to make an opening statement on behalf of Revenue. I understand today's meeting is to focus on Revenue's use of artificial intelligence. I am accompanied by my colleague Mr. Keith Redmond, Revenue's chief technology officer.

Revenue's mission is to serve the community by fairly and efficiently collecting taxes and duties and implementing customs controls. Its vision is to be a leading tax and customs administration, trusted by the community, and an employer of choice. Revenue has a long history of using technology and innovation to support the delivery of its mission and vision. As far back as 1963, Revenue purchased the first modern computer in the Irish Civil Service. This was the second computer ever installed in an Irish organisation.

More than 25 years ago, Revenue's online service, ROS, was launched, allowing taxpayers to file returns and make payments online. Since then, Revenue has introduced a wide range of digital services, from myAccount to electronic customs declarations and real-time payroll reporting. Last year, more than 87 million transactions were processed by Revenue's digital services. The availability of these services reduces the administrative burden on taxpayers and makes it as easy as possible to be tax compliant.

Artificial intelligence is frequently considered a new technology. While the new tools now available no doubt provide greater speed, scale and flexibility, the use of AI is well established in Revenue and in tax and customs administrations globally. While Revenue does not use AI to make tax-related decisions or carry out its statutory functions, its use is very much part of what Revenue does and how it operates. For many years, Revenue has leveraged such technologies to assess tax risk and detect potential non-compliant behaviours. This enables Revenue to allocate compliance resources more effectively. Revenue's PAYE real-time risk system, which uses a combination of predictive analytics and rules-based models to detect risk in PAYE refund claims, received a Taoiseach's award back in 2012.

While Revenue has been using AI for a long time, technological advances over the past few years, particularly in respect of large language models, mean the current focus is on the application of AI to deliver improvements to both external-facing service provided to taxpayers and their agents as well as internal-facing operational efficiencies. For example, Revenue has deployed AI models to assist in more accurately directing taxpayer queries to the appropriate Revenue caseworker, assist staff in accessing all relevant information to allow them promptly and accurately to respond to taxpayer queries, and generate initial drafts of tax and duty manuals, which are then rigorously verified and refined by subject matter experts before publication. The commonality across these models is that they are designed to support human decision-making with a clear focus on gaining efficiencies. More information on these AI solutions is included in the briefing provided to the committee today.

Revenue owns and manages a very large and highly integrated set of ICT services. Revenue has long-established, embedded and robust governance arrangements in place to oversee their development and deployment and to ensure value for money is being achieved in relation to the associated expenditure. Revenue's IT executive has oversight of all ICT expenditure and each system development project has a tailored governance structure supporting its delivery.

This approach ensures that risks are managed and mitigated and that project objectives and benefits are realised within budget and delivery timelines. Additional governance arrangements are in place in respect of the development and deployment of AI solutions, including that developments are in line with both the EU AI Act and the national guidelines for the responsible use of AI in the public sector. Revenue’s effectiveness as a tax and customs administration depends on the public's trust and confidence in its fairness and integrity. As such, robust internal processes that prioritise ethical standards, data protection and meaningful human oversight in the design of all AI solutions are in place.

Revenue’s approach to leveraging technology, including AI, is a direct continuation of a decades-long strategy to invest in ICT in a manner that aligns with its corporate priorities and enables the provision of quality digital services. This ensures that Revenue’s critical core business is maintained in a way that adds value and drives efficiency. The development of transformative and innovative business solutions, some of which are AI led, supports Revenue in fulfilling its mission to better serve the community. Every innovation is carefully adopted and governed, ensuring Revenue remains a leading, secure and modern administration for the years to come. My colleague and I are happy to answer any questions the committee may have in relation to Revenue’s use of AI.

We welcome the opportunity to brief the committee as part of its examination of the use of artificial intelligence in the public sector. I will begin with a few words about the Department of Foreign Affairs and Trade, a global organisation with offices in Ireland and at over 100 locations around the world. The Department works to protect and promote peace on our island, support our diaspora communities, ensure active and principled membership of the UN, the EU and other international forums, deliver a high-quality international development programme and expand Ireland’s presence and influence abroad to underpin prosperity and security at home.

The Department also provides a first-class service to Irish citizens. The Passport Service, delivered by the Department, provides passports to 1 million Irish citizens at home and overseas each year. The service is at the forefront of the Government’s programme for the delivery of online services to citizens. As the committee is aware, the rapid acceleration of AI and in particular advancements in generative AI opens new opportunities but also brings unprecedented challenges. Ireland’s national AI strategy takes a human-centric approach to governance, emphasising fairness, transparency and the building of public trust in AI. These principles are at the core of our Department’s approach to AI domestically and in our global engagement.

The Department’s Statement of Strategy 2025–2028 outlines our ambition to build our capacity to innovate as a Department that consistently delivers excellent public service and is agile and responsive to national and global developments. Part of delivering on this ambition to build our capacity in emerging technologies, in particular AI, can be seen in the competition currently under way to recruit, at assistant secretary level, a chief information officer for the Department. This will be a new role on our management board, responsible for leading digital transformation across the Department including the Passport Service and our global network of diplomatic missions. AI offers many new possibilities for public service transformation. As with all Departments, we are working to avail of the possibilities while mitigating the risks. The Department is already active in exploring the potential of AI to enhance our work. In 2023, the Department developed an interim policy on generative AI for all staff of the Department. This policy, designed from a cybersecurity perspective, outlines that public versions of generative AI services must not be used for business purposes. This was to mitigate the risk of official or sensitive data being submitted to external sources beyond the Department’s control.

As the committee is aware, the guidelines on the responsible use of AI in the public service were published in May 2025 and provide us with a comprehensive framework for AI adoption. The Department’s approach is now rooted in this framework, balancing AI innovation with responsible governance, emphasising human oversight and public trust as central considerations in AI adoption. In line with these guidelines and the Department’s goal to promote and champion a culture of innovation and transformation, work is under way to update the 2023 policy to reflect the continuing advancements and developments happening within the Department and wider afield.

The updated policy will provide a refreshed framework for the appropriate use of AI technologies so as to identify the right use cases for AI and enhance the quality of our public services, efficiency and productivity, and in so doing, uphold ethical standards, fundamental rights, legal compliance and public trust.

Members will be aware that the EU AI Act classifies AI systems based on risk, setting specific principles for high-risk AI applications while imposing bans on unacceptable risk AI practices. The Department is integrating this risk-based approach to ensure safety, respect for human rights, data privacy and accountability, and to enhance public trust. The Department is committed to appropriate ongoing oversight of AI deployment. We are currently reviewing the structures necessary within the Department to review high-impact AI projects and ensure adherence to a risk-based approach.

There is currently no deployment of AI technology within the online services that the Department provides to the public. As a Department, the safeguarding of people's health, safety and fundamental rights is front and centre of our service provision. Data protection and quality are key considerations in the deployment of AI, with the Department committed to protecting personal and business data in line with robust cybersecurity and ethical data handling practices.

In order to examine the technical feasibility of establishing an in-house large language model service and chatbot, research has been undertaken by the Department's ICT unit since 2024. This research operates in a sandbox environment on the Department’s own business data. This data is not personal and is held on premises under the terms of the Department's ICT security policy, which itself is based on National Cyber Security Centre guidelines. This research is in the low-risk category. The sample use cases of this research are all solely in the area of knowledge management, using AI tools to improve our own personal productivity and the Department's operational efficiency. None of the use cases involve the direct delivery of public services.

One use case is now implemented as part of the design and development of the Passport Service’s new processing system. A digital assistant, which is based on generative AI and named Ask Finn, was introduced during 2025 for internal use by officials of the Passport Service. This digital assistant was developed internally and is based on premises, meaning it is low-risk, extremely secure, does not leverage commercial or cloud-based systems and does not have any capability to access or retrieve personal information. The system will allow officers to ask questions related to the processing of passports, which Finn will answer by pulling from the Passport Service’s policies and procedural documents. The Department intends to develop other appropriate AI use cases following a period of robust testing.

It is already clear that the current wave of technological innovation and advancement will have a profound impact on ways of working. While the opportunities offered by AI to potentially improve our public services are significant, we are highly aware that so too are the risks. The Department of Foreign Affairs and Trade intends to harness these opportunities while mitigating the risks, innovating and improving our service delivery to the benefit of citizens, while at the same time protecting our infrastructure, building public trust and safeguarding fundamental rights.

I thank the committee for its time. My colleagues and I welcome any questions.

Thank you. I thank all the witnesses for the opening statements. We now move to the members.

I thank the witnesses for their statements and the additional documentation provided to the committee. I will start with Revenue but this question also applies to the other witnesses. Does Revenue conduct a risk assessment before any new service goes live, either internally or to the public? If so, what does that risk assessment look like or entail?

When we look at technology in general, we first need to understand the technology assessment and understand whether certain technology is the right technology to solve a problem we may have. For example, AI may be chosen as the right technology to perform that assessment or solve that problem.

We will then do a risk assessment on the use case and how we are using that technology. As members have seen from our briefing information, we have a number of different solutions that use AI. Depending on the risk associated, we put in different levels of mitigation. We also need to make sure that it is a business fit as well. We need to make sure that it provides us with a solution to the business problem we are trying to solve. In the documentation that we provided, we outlined some of the risks, such as data protection, and our role in taxpayer confidentiality and the tax data we hold on individuals. To mitigate risks in that aspect, we will always look at encrypting data. Where we use cloud services, we encrypt data in transit. This means that, as it goes to the cloud service, it is encrypted and only Revenue holds the decryption key. In certain high-risk models, we would not persist or store any data in the cloud, so it does not stay there, and it is then encrypted on the way back down. Again, only Revenue holds the key to that encryption. They are the kinds of things we would look at. We would do risk assessments on all technology, not just AI, and based on that risk assessment, we would put in place mitigating factors.

Is there an established framework in terms of risk assessments that Revenue is using or is it developing its own?

As members will know, the EU AI Act entered into force on 2 August 2024. There are a number of key milestones in the AI Act. The Deputy may be aware that, on 2 August 2026, with the establishment of a new AI office, there will be various templates for risk assessment of high-risk models that are in Annex III of the Act. While that is not yet in place, we have developed our own risk assessments. We worked very closely with Department of public expenditure and reform on the guidelines for use of AI in the public service. We are also engaged at an EU level and at OECD level with tax and customs administrations globally. Together, we have learned from best practice globally in developing our risk assessments. We have used international research and best practice and have worked with the Department of public expenditure and reform. When the AI office comes into effect, we will work with it to ensure that we follow the risk assessment that is outlined, which it must provide-----

As a practitioner, will Revenue get an opportunity to feed into that in terms of its experience in developing and implementing risk assessments?

Absolutely. Our chief information officer would be involved in the digital leaders' group. We would share experiences across the public sector on risk assessment across all technology. We were highly involved in the creation of the public service guidelines and we worked very closely with colleagues from the Department of public expenditure and reform, the Office of the Government Chief Information Officer, the Department of agriculture, and others who were more advanced in the use of AI. So, we do feed that in. We also feed in at a European level, at a World Customs Organization level and at an OECD level.

What of the Department of Foreign Affairs and Trade?

In the Department of Foreign Affairs and Trade, the deployment to date of AI is in the low-risk category but even there our Department risk structures play a role. We have two committees, both chaired at deputy secretary level. One is the passport oversight board and the other is the digital transformation board. The Department's risk management structures apply to AI as they do to other parts of our work. As we move forward with AI, and like Revenue, we are guided and informed by the public service guidelines, particularly the responsible use of AI canvas that has been developed in that. At management board level, we have discussed that future use business cases made by individuals or by work areas would be asked to workshop the responsible AI canvas, which looks at the risks under the various headings at that initiation stage. We certainly have a robust approach to risk. That is being rolled out with regard to AI developments.

In my last minute, I have a separate question for the Citizens Information Board. Mr. Long made the point that people still liked having access to face-to-face support and gave his sense of the importance of that to people.

I sense that as well, and the need for it into the future. We could envisage a scenario where a huge amount of work is done through AI and where some people might get left behind. Will Mr. Long give his experience in relation to that from the perspective of the CIB?

It is really critical. When we were developing our digital strategy, by far the preferred channel for people who came back to us on our digital strategy was a drop-in, in-person service. We have the evidence on it from the user base who were feeding back on our survey. I do not think AI can in any way replace an in-person, face-to-face interaction. Our board is very committed to it. An individual has a frame of reference that the AI does not have. Individuals have the ability to advocate and personalise and they can empathise with the person who is front of them. AI can do none of that. We are working on a project to look at impact metrics, the impact on people, of the services we provide and I do not think AI could in any way compare to what we can get back from a face-to-face and an in-person interaction. We are, therefore, very committed to in-person interaction in the CIB.

Apologies for missing the first contribution. A risk-based approach has been mentioned a number of times. I will address the green or environmental aspect of that shortly. Revenue and the Department of Foreign Affairs and Trade will be best placed to answer my first question. We are aware that the omnibus legislation will seek to erode that very aspect of the EU AI Act, namely, risk assessment. What are the witnesses' thoughts on the risk assessment as it stands? I am sure most people would further develop, rather than erode, the risk assessment in a few areas. We know the omnibus legislation is coming down the line. If that erodes risk assessment, are the witnesses committed to the risk assessment as it current stands or will they reduce their commitment to it, regardless of how the omnibus Bill impacts the AI Act?

I thank the Senator for that question. As I said in my opening statement, Revenue relies on the trust and confidence of Irish citizens and taxpayers in order to allow us to carry out our functions and to have high levels of voluntary compliance, which we do have. Risk assessment is crucial in terms of the public having trust in us and us carrying out our statutory function. Yes, it is enshrined in the EU AI Act but we do not just do it because it is enshrined in the Act. As I have said, we have had large numbers of ICT projects. We are a highly digital organisation, also recognising, as the CIB has said, that there is a need for in-person communication as well. However, we would always carry out risk assessments on any technology we use. There is a lot of talk about AI at the moment. We have been using things like predictive analytics models all the way back to 2012 and would have carried out risk assessments at that time too. We do a risk assessment any time we are doing a procurement for any technology.

How does Revenue identify the risk at that level when it is doing the procurement process? How is risk identified?

If we are talking about ICT risks, we would look across a broad spectrum of ICT risks that would be well understood within the industry. We might be looking at things like cybersecurity risks. Again, that is crucial in terms of data protection. We might be looking at data protection risks, as I spoke about earlier. Vendor lock-in is a really important risk we need to consider in all our ICT spend. We would look at that as well. We also need to look at how our staff will interact with services and how we can provide those services in a way that is accessible to staff. All our digital services have to be accessible. In the ICT industry, there is a broad range of risks and we would look at those across all our technology developments, not just AI. Trust is crucial in the work we do and we do risk assessment because it is right thing to do, rather than necessarily just because it is enshrined in the EU AI Act. We have always done risk assessments on all our technology developments to date.

Would Revenue have concerns about the impact of the use of AI on the environment? Is that worked into the risk assessment beyond how people engage with it in order to take in the wider risks?

There has been a lot of discussion on the impact on the environment and the amount of power needed to build what are the foundational models. The larger hyperscalers - the big enterprises that build the foundational models - use a huge amount of power to build those models. Once they are built, there is obviously an environmental impact to the use per go as well, that is, every time it is used.

In Revenue, we have a carbon reduction strategy and we report on it every year to the Sustainable Energy Authority of Ireland, SEAI. We have a large data centre and a back-up data centre. We house more than 50 public sector bodies within our data hall and we have been working over the last number of years to make carbon reductions. In that respect, I think we hit 5% in the last two years and indicative figures for 2025 look like us getting nearer 10%. That is done by replacing old hardware and using cloud-type technologies that allow us to virtualise our servers and containerise things. It is really important for Revenue and society as a whole that we look at that carbon reduction strategy. Then we need to look at how we use AI and what models we use but the creation of those foundational models, regardless of whether we use them, is what uses up a huge amount of power.

Our approach is a measured one. We are probably not operating at the scale of the Revenue Commissioners but as with the Revenue Commissioners, this commitment to maintaining public trust is very much at the heart of what we do. For example, protecting the personal data of people is extremely important and this is very relevant to the work under way in the passport service.

Generally, our approach to risk is that every new project coming out in the Department is subject to risk. In our project oversight board, for example, and in developing capital projects abroad, we have a risk matrix.

I am sorry to interrupt but I will be specific. As Ms Maguire said, the use of AI is not at a certain scale within the Department of Foreign Affairs and Trade, but I wonder about the use of AI - maybe it falls under the Department of justice - in relation to the asylum process, International Protection Accommodation Services, IPAS, centres and stuff like that. Is there not an intersection somewhere with the Department of Foreign Affairs and Trade with regard to other Departments' use of AI?

There is not at the moment. Our use of AI and the use of it in the passport service is contained to a tool, Ask Finn, which accesses policies and procedures and is used by officials of the passport service in helping them assess passport applications. There is no link between that system and others.

AI is not used to monitor passports. I am aware of an AI that is being used in regard to identifying whether documentation is real. Would that not be within the Department of Foreign Affairs and Trade?

I will pass over to Ms Siobhán Byrne, who is director of our passport reform.

As Ms Maguire said, our only use of AI within the Department is related to passport service processing but it is very much a tool used by officers in helping them determine the eligibility of somebody in their entitlement to a passport. You must be an Irish citizen to hold a passport.

In relation to the documentation that is submitted, we still obtain physical documentation from all of our customers. We have a highly skilled workforce that assesses those documents. We receive documents from all around the world and all jurisdictions. We have tools in place that allow us to identify the security features that must be seen on documents from countries all around the world. That is very much a person looking at those documents and using the tools available to them to assess them. The only use of AI within the passport service is very much as a tool that looks at the policies and the guidelines to assist the officer in making a decision as to the person's entitlement to hold the passport or, in the case of a child, to ensure all guardianship requirements are being adhered to and so on.

It is, I suppose, a starting point for us. We have a limited use and we are certainly not at the scale of Revenue, as Ms Maguire has mentioned. It is quite a different process.

I thank the witnesses for attending. I will refer to each opening statement. Mr. Long said something interesting about how traffic to the citizens information website has declined. He attributed this to the growth in AI, which is an aggregator of information, and to people taking their information from the various AI tools with which we are all familiar. He flagged one of those. Could he speak about that a little more? Was that the genesis for the citizens information service taking the decision to develop an AI rival with more authentic and verified information?

We were already working on it prior to the obvious reduction in our statistics, which was apparent from March onwards. It coincided with the turning on of search engines and an AI insight overview tool. We were able to track a reduction in our statistics and match an increase in the number of times we were referenced in search impressions on the same tools. For us, it was very clear that people were embracing AI. Whether knowingly or not, they used AI overviews and did not click through. As a result, we were getting zero click searches. From our point of view, that was a major driver in terms of putting the project forward. It was apparent that we needed to provide a tool that people would want to use. Given the speed of decline, it was apparent that people wanted to use AI to investigate information. It was such a powerful way to do it that our users were doing it with or without us.

Mr. Long referred to Google, which is where people carry out most of their searches. He said the challenge is that when it provides an AI summary through its search function on matters that have been taken from the citizens information website, it is not necessarily the case the information is wrong now but that it is dated information. Is that the issue?

The information could be perfectly accurate, but we are trying to limit it to one set of sources such that we can stand over the verification processes. We limit our AI to our website content. It does not have access to any other material. That way we feel like we can stand over the information in the same way we stand over our website. Other tools are aggregating information across multiple different sources and, of course, we cannot stand over other sources that are being mixed in.

I may have picked it up wrong, but I thought Mr. Long had indicated that some of the information being presented in summary form might have been taken from an earlier source and was out of date. The citizens information website is updated constantly. I refer, for example, to the updates happen following tax and welfare changes in the budget.

The other sources can absolutely be outdated and inaccurate. We would observe that from time to time on other AI tools.

It is not necessarily the case that AI is taking information from the citizens information website that is out of date. Is it taking information from an older version? I am trying to get to the bottom of-----

It would be from elsewhere. From somewhere other than our website.

It is from elsewhere. It is not the case that AI search tools are necessarily taking information from a wrong version of the citizens information website. Rather, they are going to other websites that might be inaccurate.

Typically, they use multiple different sources to generate one response. Ours is often one of those sources, which we know from the search impression statistics. They are using other sources, and we do not have any control over that.

We are still in an early enough phase whereby people are using search engines which give them AI summaries. We all know that from the various apps we have. Obviously, the citizens information service is doing its own work, which seems very laudable. Is there a way for the service to engage with companies which provide search forms to ensure that they extract the true information that the board is presenting and work hand in glove, even if it impacts on the traffic? The citizens information service does not have a profit-making website; it simply wants people to get the right information. Could there be greater synergies?

That would be really worthwhile. It is something we have considered. We have not pursued it as much as we should. We have made contact with Google on the matter.

Ms Maguire referred to the sandbox approach the Department of Foreign Affairs and Trade is taking.

It seems to be something that is working, whatever is the first one the Department has gone with. She spoke about how the Department will do more of these. Could she give us an indication of the areas the Department is looking at doing more of?

We are looking at knowledge management and areas that can improve efficiencies in the Department. It could be in the area of policies and procedures. Accessing that information is what we are principally looking at. Our key citizenship public-facing areas are the passport service and the consular service. We are not looking at replacing the human interaction at that level yet. It is really in the knowledge management area through the use of large language models.

What does "knowledge management" mean in plain English? Could that cover diplomatic cables from various embassies across the world back to the Department of foreign affairs?

I might ask Mr. McGinn to elaborate on the area we are researching at the moment. It is coming out of our ICT unit and I will ask Mr. McGinn to elaborate on the areas we are currently testing, apart from the passport policies and guidelines.

The first intention is to build a model that is repeatable. The implementation of the passport service digital assistant has proven the sandbox, the model - the LLM - the associated technologies and our capability to run them is something we can do. We first wanted to ascertain that and that we are on a solid footing where we could use this kind of tool. I see the implementation of it being in areas where people need reference material to make decisions or do their job. Having access to that reference material previously meant trawling through a very large number of documents, whether they were statutes, international agreements or previous cases, to use a legal example. The AI models now deliver that at a person's fingertips in a conversational, easy-to-query way. This is where we see it.

For use cases, it might be for people who work in finance areas, which are heavily controlled by guidelines, so they would have access to those guidelines and could query them instantaneously. It might even be in HR where there is a certain amount of procedure and guidelines to be referenced to do the work. A user would have quick access to those kinds of things. That amounts to knowledge management. It is not necessarily for communication or emails, like the Deputy alluded to there.

It is not in the space of - intelligence-gathering is the wrong phrase - how embassies are communicating to one another, accessing information and engaging in diplomatic affairs. Is that an area that has been touched by AI yet?

Not specifically. We are still looking at a new policy and future use cases. The more comfortable one becomes with these tools and the more accessible they are, the further uses they will have. I have had not had specific discussions on those cases Deputy Geoghegan mentioned.

I thank all the witnesses for being here. It has been really interesting to read through the material and hear the witnesses' statements today. It is really heartening to see public sector bodies going above and beyond what is required at the moment in the regulatory framework. It is so crucial, as Ms Kennedy mentioned, for that public trust in our public services to remain intact. It highlights that there are weaknesses in that regulatory framework which we all need to address.

I will start on something specific. If I have time to broaden out in this round of questions, I will. Otherwise, I will come back to it. The specific area I want to talk about is the Irish language and I will come to the Citizens Information Board first, please. Can the AI tool which is being trialled be used in Irish? I am mindful that these tools are essentially a new form of public service, which are subject to the Official Languages Act.

The question can be input in Irish and the tool will respond in Irish. A number of our team internally are Gaeilgeoirí and are Irish speaking. They have reviewed it and it needs future development. We will continue that. It was only released on 23 December.

It is also capable of translating to 50 other languages. At the moment, one has to type in the question. If it is typed in of those 50 languages, the tool will respond in that language. We will add a voice utility and functionality in order that a person will be able to speak into it and it will respond.

Given the reliance of these systems on a pool of data, does that exist? Is there that bulk of language data available to the tool to offer the same service as Gaeilge as it does as Béarla?

We have a very similar number of Irish pages and English pages in the citizens information content.

The model is drawing from that data.

Yes, exactly.

I thank Ms Flanagan.

My next question is for the witnesses from Revenue. Regarding the current exploration of AI call handling, how does Revenue deal with the difficulties presented not only by the different dialects in Irish but also the different regional accents people might have? Is there still an option to revert to pressing 1 for X service? Secondary to that, how feasible is voice recognition in the context of our duties under the Official Languages Act, given regional accents have huge variance, particularly in the Irish language, and where there is limited training data available?

In terms of our use of AI in the call centres, as outlined in our briefing documentation, we are engaged in an initial proof-of-concept process. When a taxpayer rang one of our call centres and talked to a customer representative or one of our Revenue officers, the process at the end of that call was always for the Revenue officer to summarise what happened during the call and store it on our system to ensure that if the person rang back or there were further communications, the next Revenue officer who answered the phone was aware of what had happened previously. What we are trialling at the moment is allowing the AI to summarise that call. We are currently at the proof-of-concept stage whereby our operators are reviewing the AI summaries, can edit them if they are incorrect and can provide a feedback loop in terms of the development of that AI model.

We have call lines as Gaeilge and as Béarla and the model is being trialled on both. It is very much still in the proof-of-concept stage and the Deputy is correct that there are different dialects. We are working in conjunction with Údarás na Gaeltachta, which is looking at the provision of AI services as Gaeilge. We are providing an tÚdarás with some data to help train its Gaeilge model. From my engagements at EU level, I know the Swedes are looking to develop a Swedish large language model. That is a national initiative in which they are engaged because the majority of large language models are developed as Béarla. The Swedes are looking at a national level at a Swedish large language model. Our efforts in this area are a work in progress. Údarás na Gaeltachta is also doing a lot of work in the space of Gaeilge and AI.

Is Revenue able to feed into Údarás na Gaeltachta in terms of what we need as a State to ensure we keep tightly to the Official Languages Act and all of that?

We have good connections there. Revenue's rannóg na Gaeilge, which provides all our Irish language services, has been in contact with Údarás na Gaeltachta and they are doing some work together to try to prove out how it might work. With a lot of these types of solutions, a proof of concept must be done first to understand how the solution might work before going into production. That is the phase we are in at the moment.

Ms Kennedy talked about a multivendor strategy. In this regard, Revenue seems to be streets ahead of other areas of the public sector where there seems to be a real reliance, for example, on particular office suites. How has that been implemented and could there be a wider application of it across the public sector? Does Revenue have the opportunity to help other organisations, including Departments, to consider having a broader suite of tools available?

As I said earlier, with AI, as with any other technology, we need to be aware of the vendor landscape and be careful of vendor lock-in. Sometimes, there is only a single vendor providing a service. Where there is a multivendor landscape, procurement can be done on a framework-type of initiative, with a number of vendors qualifying. In terms of the AI models, particularly with the large hyperscalers, we found that some providers were better at some tasks than others. When it comes to the technology assessment, looking at the business problem that needs to be solved and trialling various models, we find some models perform better at, say, a question-and-answer process based on curated documentation while others may perform better on reasoning.

In our ICT architecture, we use what are called broker layer technologies, whereby we disaggregate our revenue services from the cloud provider services in order to be able to swap in and out. We have a large ICT estate. We sometimes find that some vendors stop providing a service we were using, which means having to move.

This has been going on since time immemorial in terms of ICT. We built a service a long time ago using Abode Flash, if I remember correctly. Suddenly, we had zero-day attacks, that was discontinued and we had to move on. In making technology strategy and choices, you have to be aware. Our former CEO liked to say ICT applications were for life and not for Christmas. It is not that you build it and then it is done; usually the total cost of ownership in things like Gartner is 20% build and 80% run. It is important to look at that. Revenue is involved with other public sector bodies. Our chief information officer and chief technology officer sit on various forums and we share all our experiences because we have a mature application of ICT. Others are starting and we are very willing to share our experience with other public sector bodies.

Are there any open-source technologies in Revenue's suite of vendors?

There are across our entire ICT estate but not in our AI space. Open source does not mean free and it is important to understand that. Open source has pros and cons. We evaluate every technology appropriately.

I thank the witnesses for the services they provide. Some of them are excellent. I have a couple of general questions for all the witnesses. How do they identify bias in their codes and programmes regarding, let us say, people with disabilities, vulnerable people and people who are deaf or blind?

One would need to look at the use cases for the technology. In line with the responsible use of AI and the committee's recommendations published just before Christmas, everything we do involves a human or a Revenue officer in the loop. None of our statutory functions are performed and none of our tax decisions are made by an AI system. The examples we have provided are where we use AI, maybe routing an inquiry to the appropriate Revenue official. The officials deals with the inquiry, rather than AI technology. As I said to Deputy Gibney, AI may summarise a call but there is a Revenue official who reviews the summary before it is stored. We do not have a use case where a human is not in the loop at the moment. That is how we ensure there is not bias built into the processing.

The Ask Finn model now deployed in the passport service is an internal model meant for staff of the service, regardless of accessibility issues. More broadly, the point on accessibility cuts through the priorities of the Department. The passport service went online in 2017, I believe. At that point, 100% of applications were made on paper; now it is 8%. We are concerned to ensure the experience of that 8% is as inclusive as possible. I ask Ms Byrne to talk about the passport assist service.

The passport assist service is about digital inclusion. We fully recognise not everybody wants to use a digital channel but we know it is the most efficient way in which we can process a passport application and give the best service to everybody. Members are probably familiar with the green form and the post passport service that has been in place for quite a long time. The change we implemented in December was to allow us to improve the service for those who wish to continue to use paper. They contact us and speak to a person who identifies the best way forward for the applicant to obtain their passport. If that is the paper route, we generate their form for them, populating fields relevant to them, and send them the form. They adhere a photograph to the application and bring it to the post office, the same way they would have brought their green form to the post office. It is trying to mimic the process so people who wish to continue with paper can do so in a way familiar to them while, at the back end, we have improved our service to them.

They were getting their passports in eight weeks. The current service is now six weeks because we have been able to improve the efficiencies in the back end. For us, it is very much focused on that direct interaction with the customer and identifying the best way for them to apply. In the context of that digital inclusion piece, we are ensuring that we assist those who want to go digital and, for those who do not wish to do so, that we have a channel that is more modern and beneficial, both for them and us.

AskCI has been developed to solely answer questions and queries from a user input based on the content from the citizens information service that we have been designing, developing and writing for many years now. This provides, we hope, non-biased answers and responses that are helpful, accurate and clear.

In respect of accessibility, the citizens information website had a recent review from the National Disability Authority, NDA, and was marked at 94%. It also recently reviewed AskCI. It gave it a 93.3% rating in terms of accessibility and deemed it to be aligned with standard 2.1 of the web content accessibility guidelines, WCAG.

As all of the witnesses’ organisations are using AI, do they see it having an effect on the reduction of the workforce?

At the moment, when it comes to the use cases and the uses we have made of AI and technology in general, it can absolutely be seen that efficiencies are being gained, particularly with regard to quicker speed and our time of processing. We are very much in the early stages of it at the moment. From a Revenue point of view, in the context of our role in collecting taxes and duties fairly and efficiently, we have roles in the areas of compliance and customer service. We may see a change potentially to where staff are deployed. Maybe more staff can be deployed to compliance-type activities if some tasks are able to be automated, but that will all take time to understand.

We have all heard of the ransomware attacks. How do the witnesses’ organisations defend against them? What systems do they have in place? Do they connect with other countries or systems? What way do they work?

It is a very serious concern of ours. We are in step with the National Cyber Security Centre in the way we operate for security. We have a large global network in the Department of foreign affairs. We have a very high priority at the centre to keep that safe and we do that through layered security, different approaches and technologies and external security services. Having that secure infrastructure and striving to have that sort of infrastructure allows us to work in these AI areas. It is a main concern for the Department.

I thank all the witnesses for being here today. As public representatives, the witnesses’ organisations are probably three of the organisations that we engage with the most, among others. I engage with all of them a lot and I appreciate the work they do. What we have been discussing mostly in terms of the software they are using is knowledge-based management. Essentially, the information they have is being shared, either internally or externally. I will go through each organisation as the opening statements came. First is the Citizens Information Board, CIB. Is AskCI, which was developed in partnership with the Office of the Government Chief Information Officer, OGCIO, essentially a large language model, LLM, based on the knowledge that the CIB has inputted onto the website?

Yes. It is basically using an LLM and the retrieval of augmented generation to process the queries passing through an LLM and then responding using the content from the website.

I might have missed it but what is the accuracy, in percentage terms, of the return?

The return of accuracy is the response rate, that is, what we are getting back from it. There is no knowledge base. There is no external content being used at all. The responses that come back are all from the website.

That is brilliant. What software is being used to host all of the articles and information of the citizens information website in general?

The content is being stored on Cosmos DB. We are using Azure-based technology that is being hosted by the OGCIO and we are using others. The LLM is OpenAI. We have done a lot of prompt engineering around retrieval-augmented generation. It takes the query and the query goes through. There is a Python and Django, both front end and back end. Numerous different technologies are used.

There is a lot of complexity to it. It is not as straightforward as people might think. I commend the Citizens Information Board on that.

I am conscious of time and I have a lot of questions, so I will move on to Revenue. I commend it on the adoption approach. It is probably one of the most detailed risk assessments we have seen in relation to different aspects of AI. It is from March of last year. It is certainly a model we can all look towards. It is not perfect but things are changing consistently and, for its time, it is very forward-looking. I also commend Revenue on avoiding vendor lock-in. I worked in the tech sector for a long time. While the sector would love public services and clients to be locked in, it is in the public interest that there is no lock-in. I commend Revenue on that. I think about six or seven various models were outlined. MyEnquiries, classification and routing, RevAssist AI assistant, RevAssist document generation AI assistant were all referred to in the briefing document. Are there individual vendors for all of those specific projects?

We have a range of vendors under a framework agreement, which we concluded a number of years ago. There are different technologies in use. We have a large ICT staff within Revenue and we develop all of these solutions internally. We use the APIs available from the vendors to allow us to develop the solutions in-house. We do not use ChatGPT or whatnot but we are using the models behind them. As I said, we found model performance for different tasks to be different, depending on the vendor. Again, it is about assessing the technology and fit when deploying the solutions.

I note the Revenue highlights the routing accuracy is between 92% and 98%, which is great, but there is no demographic performance analysis, looking at different demographics. There has been none of that. To ensure the model performance is equitable, would that be something Revenue would look towards? Deputy Mythen alluded to it as well in terms of people from disadvantaged backgrounds. Should that be considered as part of the routing analysis?

What the routing does is that taxpayers who have queries and use our digital services are able to raise that query digitally with us. Before we used these models, we would have asked taxpayers to categorise their query. We had a category and a subcategory. As many people will know from computer systems, the dropdowns became extremely long and most people choose the first one or "other". The reason we need to classify it is to make sure we route the query to the most appropriate Revenue officer. As can be imagined, Revenue is a large organisation and some of our staff work in, for example, PAYE, and are experts in that area, while others are experts in debt collection or in the customs area. Therefore, it is not just that customers can be routed to any Revenue official. The idea of the model is that we have our staff, we understand skills and areas of work and it enables us to route the query. When people are raising queries with Revenue, they want to say they want to ask about tax; they do not really care about our internal structures. Depending on the level of detail customers put in their inquiry, it will probably mean we have less to work with in terms of routing. Ultimately, however, it will go to a caseworker.

I thank Ms Kennedy for that. The Department of foreign affairs mentioned Ask Finn on a number of occasions. Is it true that Ask Finn was developed by a successful Irish company called Intercom?

No. Ask Finn was developed by the Department. It was developed by Mr. McGinn and his colleagues, working with Ms Byrne.

It is just as well I asked that question. There is another system, Fin AI, which was developed by Intercom and I had some questions about the appropriateness of using that. However, Ms Maguire has clarified that, which I appreciate. At the moment, Ask Finn is used for queries in relation to passports.

Will it get to a stage where it will do the likes of recognition of images, for example, when somebody submits an online image of themselves for passports, etc?

We use very different technology for photographs. We have a tool that assesses photographs on the way into us and then we have another tool, which is a facial recognition tool, that assesses photographs against all the photographs within our database to ensure that the right passport is being given to the right person. It would be very much best practice among passport services globally to have a facial recognition system. It is, however, very much an independent system. It maps an algorithm of the face in the most recent photographs sent in against the most recent passport photograph we have on the last passport. It does a comparison. It is very much a stand-alone tool and it is in its own space. For every passport that is issued, however, the photograph is still assessed by an officer before that passport goes out the door. It is an automated process to a degree but every photograph is still assessed by an officer before that passport would be issued.

I would say my photograph from ten years ago is very different to now.

Now it falls to me to ask a couple of questions. As public service organisations, I applaud the work the witnesses are doing. As a committee, we want to see innovation and new technology used with the necessary safeguards in place. The witnesses' organisations are kind of role models within the public service in that regard.

I will start with Revenue in terms of looking at where we are moving. I have no doubt that the use of agentic AI is already happening. Maybe Revenue is already dealing with it, and the witnesses can talk about it in a moment, but I fully expect that companies are probably using agentic AI to interrogate questions such as how do they deal with the tax treatment of a particular invoice or possibly what is the most tax efficient way to deal with a particular set of business options that it is doing. Will the witnesses talk to us about whether Revenue is already using agentic AI within its offices - I will ask the others as well - and how is Revenue going to respond to companies or tax advisers who are using AI agents?

"Agentic AI" is definitely the phrase of the day at the moment. It is important to clarify the difference between AI workflow and agentic AI. With AI workflow, we have a number of models following a similar process every time. In agentic AI, one enters the world of AI reasoning. In agentic AI, one sets AI agents a goal rather than a direction of travel-----

For example, my goal is "How do I deal with this in the most tax efficient manner?"

"How do I maximise?" Absolutely. The agents then go off and use the tools available to them, whether it is the Internet, some legal resources, case law or tax law, and can come back with an answer for the taxpayer. Ultimately, we have a system of self-assessment in tax and the taxpayer is responsible for declaring the correct tax at any point in time. Our compliance staff will use risk-based approaches to find where we want to intervene and to understand whether the correct tax treatment has occurred on a transaction, an income or whatever. If someone takes advice from an agentic AI solution, the person has to be sure that it is correct and that he or she is following the law. Some of the big four and some of the legal firms are developing their own AI to support their staff internally. One of the AI assistants that we have built internally is trained on our tax and duty manuals and on tax legislation. We have found in using it that it is very good at providing answers to very straightforward questions but if one gets into complex tax transactions across multiple corporates or international tax, it is not a replacement for a tax adviser at the moment. It is a case of buyer beware in some instances. It is a model of self-assessment and taxpayers need to ensure that they have correct advice when they are filing their tax returns.

Is Revenue seeing it being used to any great extent at present?

It is hard to know. We see the end result, which is the tax return. With regard to how they got advice for how they may have declared or categorised certain transactions in their return, we have not yet really seen someone who said he or she relied on ChatGPT or any sort of AI tool for that.

We have seen anecdotally, in places like the WRC, where people have relied on it in case law, for case law that does not exist per se.

We also see things. For example, AI is now able to credibly recreate documents to make them look correct. Revenue has always had to challenge the veracity of claims or details provided by taxpayers in tax returns. We have seen things like really credible receipts created and submitted, but when we go back to the third party who is meant to have issued them, we can verify that they have not been issued by them.

I am looking at the point made about the big four, that, ultimately, we will start to see tax advisers using agentic AI and, ultimately, they will. So, KPMG or whoever will have to take responsibility for it.

The taxpayer ultimately does, which is where it comes. An interesting thing we have done in terms of reasoning in an AI model is that one of the use cases we have is where we take the primary legislation, as enacted by or amended by the Finance Act, and we then generate tax and duty manuals, which really are the guidance for taxpayers and agents. We have created an AI tax and duty manual generator. If a new clerical officer were to arrive in, you would tell them how you create tax and duty manuals, and the type of English, the structure and the examples used. We have trained an agentic AI process to follow that and provide a first draft for our subject matter experts then to refine. In part of that, what we have done is asked the AI to follow a process that you often see in cybersecurity, which is a red team and blue team, so there is a defender and an attacker of the system. We have turned it into a defender of the legislation and an attacker, and we have asked the two agents to have a conversation to see if they can find any weaknesses in the legislation. That is something we would do generally anyhow. You would have a bunch of legislative people sit around and try to pick holes and find words and things that mean there is a weakness. Now, we can use the reasoning in AI to assist us in doing that as well. That is somewhere where we have seen the benefit of that type of reasoning within AI.

I am conscious of time but I have a quick question for the Department of Foreign Affairs and Trade. Ask Finn is a digital system. It sounds really interesting in terms of internal processes and obviously it is about whether it delivers efficiencies and so on. What has been the initial feedback? Has it succeeded in its goal? Has it delivered efficiencies? Has it helped?

Ask Finn was only launched in December so we are still at the very early stages. It was launched in tandem with our new core passport processing system, which replaced a system that was installed in 2004 and was at the end of life. We are very much at the early stages of Ask Finn. At the moment it is being used by our most proficient users of the system to ensure that the information that is being shown to them is correct before we roll it out further. We are taking a very cautious approach to the roll-out of it because it involves the issuing of a passport, an extremely important document. Certain areas of it are quite complex when it comes to citizenship and guardianship. We take those issues very seriously. Before we roll out Ask Finn further, we want to ensure that the information being given out is correct and that is being validated by expert users within the passport service at the moment.

I agree with Deputy Gibney in saying that this debate has been really refreshing to listen to in the sense that all of the organisations are very forward thinking. As a group of witnesses, and this is one of the most comprehensive group of witnesses we have listened to, they have all been extremely knowledgeable of their briefs.

I think Deputy Mythen has already asked my question on bias but I will take it a step forward by asking a collective question. Some of the organisations are at the early stages of building AI. The Revenue Commissioners are much further along. Have there been elements of co-design? I have heard phrases such as the hope that there is not much bias or there was 92% or 93% but they seemed to look back after something was developed. We have had a lot of groups in here and before the Joint Committee on Disability Matters who have talked about needing to be involved in the co-design phase rather than after something has been designed.

I will start with the Citizens Information Board and work around.

We have not worked with disability groups other than the accessibility test done by the NDA that Ms Flanagan referred to. We did testing with the citizens information services to get a wider perspective on how the tool was working before soft-launching it to the public but we did not have a formal co-design process in producing our chatbot.

The policies and procedures that Ask Finn works off are managed by the policy and legal division in the passport service. We have invested a lot of attention in making sure those documents are without bias and are based on the legislative right to citizenship. Ms Byrne mentioned citizenship being tricky. There are all sorts of issues around guardianship. There are various issues we need to be very careful about. We put a lot of work into the core documents the digital assistant works from being correct. The governance arrangements over clearance of those documents are robust, so that Ask Finn is checking with up-to-date information that is verified.

That is correct. It is important to note Finn is not looking at any personal data belonging to passport applicants; it is only looking at the policies and guidelines according to which an officer will make decisions. There is no personal data within the tool.

At present, all our AI solutions are deployed internally within Revenue. In a lot of our digital service, we have engaged in co-design with various groups, including the National Adult Literacy Agency, NALA, on adult literacy and Age Action on older people in general on our digital services.

On the first AI assistant we deployed internally, we worked with about 20 clerical officers on our PAYE helplines. They were the individuals who answer the phone calls from taxpayers every day. We worked closely with them on the design and testing of the assistant we had provided to support them in their work. We got good feedback from them and they drove the process. They had first-hand knowledge of what taxpayers asked them every day. They understand the type of queries they got. They were really excited by the ability to speed up the queries. When you are in a call centre, you can always see how many calls are in the queue waiting for you to answer them. The people working on our calls try to provide the best service possible. They were drivers of what the AI could do to make them more efficient in their jobs and provide the best customer service possible.

The more co-design we can have, the better, especially for front-facing AI tools. Moving away from bias in decisions and towards use, if we can bring in people with disabilities from the beginning, we will not have to make changes to things after they have been put in place.

Sticking with the Citizens Information Board, Mr. Long spoke about decreased traffic to the website and said that helped to speed up the birth of its AI tool. Has the board learned much from Google searches and its own chatbot about what people are asking? Has that helped it make changes to its webpages? Ms Kennedy said people get in touch with Revenue and say they need help with tax but might not know which tax category so Revenue uses its AI tool for that. I often find that with citizens information as well. Somebody came into my office yesterday who we thought in the beginning was looking for help with finances but it ended up being a totally different query. Has the Citizens Information Board learned things from the use of AI that changed its webpage?

We have. Interestingly, we might be looking at more and more tweaking of the website to provide good content for AI - not just our AI, but other AIs, possibly. For example, we are learning about completeness of tables. Where you reference something outside the table in another paragraph, the AI may not pick that up. We go back to our content and look to revise that. On the odd occasion, we would see something on our website that has been referred out through the AI. We would go back and update the content then on the webpage in order to refine and tweak the answer AI is giving.

We are working through that iterative process. As we see the questions and the answers we are looking to refine the content to provide better responses.

I thank Mr. Long. I have one last question while I have time. I am thinking of people who might find the citizens information website who need simpler language or might have a disability. Do they get notice they are talking to an AI chatbot? Is there an avenue for them to talk to a human or how easy is that process?

The notice is on every AI chat bubble as well as on the front page of AskCI that it is an AI tool you are interacting with. In terms of engaging with a service, you cannot at the moment type in "I'd like to speak to a person" but you can get contact details for the citizens information centres or for the citizens information phone service and it will sometimes provide those where it has essentially run out of context.

Has citizens information plans to build that in, in time?

We would love to do that, just to keep it as seamless as possible.

I thank the witnesses for the presentations. I will start with citizens information. I congratulate it on its website. We use it all the time when people come to us looking for information. Generally the information is accurate and a bit more simplified than you might get in the original place. It has basically been translated into more readable English for people like me and for the public. It is a little concerning, therefore, that the representatives are saying people are going online searching for information about, for example, the carer's allowance. You might be in receipt of carer's allowance and you wake up hearing the news that may be subject to tax. You go online to search for this and historically I assume the first result would have been from citizens information. People click that, go to the citizens information site and get the accurate information. The citizens information representatives are saying that now, when you go through Google or whatever search engine, there is a decent chance the first result will be an AI answer which will have taken, potentially, some information from citizens information which is presumably accurate and potentially information from Revenue that is also accurate but it may have mangled it together with information that is inaccurate. As a result of Google and other search engines front-loading AI and pushing it at people, people may be getting less accurate information than they previously would have through search engines. Is that a fair summation?

They are perhaps getting information from a search engine for the first time whereas before they would have got links and clicked them. We would have been able to track that click-through in terms of how we are appearing in search results, what types of things people are searching for and what they are clicking through for. What is happening to us now is we do not have any real insight into that, other than that we have appeared in a search result. We do not have a sense of what people are asking Google, potentially getting results from us and not clicking through. We have a bit of an information gap and I guess that what we also have, from the scale of the decrease, is the sense people are evidently satisfied with the AI overview. They are using AI. They want it as a method to interrogate information and that is part of the reason we have pushed ahead with this project. We think it is something people find really useful and we wanted to try to limit the sources and see how it would work if it was limited to our website.

Yes, but in a sense AskCI is a different project, versus Google, because what citizens information has done by-----

It absolutely is, yes.

------parcelling off the AI into the citizens information website is guaranteeing it is only getting information from sources that are accurate. The problem is the first thing the average person who goes online to check about the potential tax implications of receiving carer's allowance does is go to Google and not citizens information. That is probably 90% or more of cases and there is an increasing danger people are not being funnelled to any of the witnesses' websites, which I trust all have accurate information, and instead are getting garbled AI stuff. People may be satisfied because they got the information but nobody tells you the information you got is inaccurate. You may only discover that later on when it turns out you have a tax bill you did not expect. Is this something citizens information, as an organisation that is about providing people with information, is concerned about? I do not mean so much in terms of the organisation's own AI usage but in terms of the wider information landscape around people's entitlements, and so on.

We are concerned about it. The zero click search is probably a concern for most website owners at this point. It is a fairly widespread view among anyone who is running a public-facing website that traffic is not coming through to the same degree it did before. Somebody mentioned earlier that we are not a commercial organisation so we are not trying to benefit from people moving through to our website.

We want to be able to track it and be able to stand over the accuracy of the information that we are providing to people. We are essentially taking what we can control, which is our own content, and trying to provide a new tool for people to interact with it. We hope that they find that useful, but where they do not, they can keep clicking through to the website or they can go to our phone service or in to one of our centres.

Related to that, an issue that came up for me during the general election campaign on a number of occasions was people being complimentary about citizens information on the one hand but on the other hand saying that they used to have a local service in their community centre but that was no longer available. I dug into a particular case and found that there had been closures of some information centres, including, for example, in Rathmines and other locations. The witnesses are saying to us that that process, which is presumably linked to funding and which makes it harder for people to go to their local centre and get information, has nothing to do with the process of going down the AI route but people will be concerned about that. The kind of people who go into a citizens information centre will, I presume, have an issue that is quite complex. They may be the kind of people who might come to us, as TDs, for assistance or they may have challenges in terms of literacy. Some cannot just go on to the website and accurately get the information they need from it. It seems to me that the cohort who would be looking to speak to people in an office and the cohort who might benefit from an AI tool are quite dissimilar.

If people have a complex or nuanced query, or a very personalised query, they are not going to get an answer from AskCI. They might spend quite a lot of time on the website trying to personalise it to themselves, in which case they are going to need to call the phone service or drop into a centre to get somebody to look through all of the different aspects. As was mentioned earlier, sometimes a query that starts in one place moves on and turns out to have several other facets to it. That is really the benefit that comes from visiting a centre. We absolutely support and promote the face-to-face service and we do not see AskCI as being anything remotely comparable to it. It is really very important to go and see somebody who has the expertise and the experience to look at it in the round and to move from facet A to facet B and to realise that, actually, this is a more nuanced problem than even the person involved might realise.

I have a quick question for Revenue. As I understand it from the opening statement, if I as customer contact Revenue, I do not interface directly with AI but there may be AI operating in the background to route me in a certain way. Can I choose to opt out of that? Can I say, when I first make contact, that I do not want anything to do with AI and do not want AI to direct my query?

The current way we use AI to classify queries is through our MyEnquiries services, which is available through the Revenue online service or myAccount. It allows customers to raise a query electronically and then it is dealt with by a Revenue officer. All of those queries get routed through AI purely to classify them, to determine whether the query is, for example, in respect of PAYE, and if so, we will send it to a person who has skills in that area. If it is something to do with importing a car, we will send it to someone in our customs division. All the AI does is make sure the query gets to the correct Revenue officer as quickly as possible. There is not an opt-out on it.

Are there any plans to use it for any customer-facing services in the future?

We have built an assistant that is trained on our information on our website. We have 1,400 tax and duty manuals which are guidance on the Taxes Consolidation Act and, obviously, all of the primary legislation. Internally, we have deployed a version of that. It is trained on all of that information and it allows a user to engage in a question and answer type scenario. We are testing it internally and would hope, in the future, to be able to release that externally. It is very interesting to go through the testing process on tools like that. We tested it with our own Revenue staff, who all have a good understanding of tax, so the way they would write their questions is very different from how a member of the public might do it. When you are testing ICT systems, you are generally doing binodal and trinodal testing and you know the boundaries, but there are no boundaries with this. We are currently refining that model but we hope to release it publicly once we are satisfied with the guardrails around it and that incorrect information will not be not provided to a member of the public.

I thank the witnesses. I will continue with Mr. Long, as he raised issues I have not heard being raised in public before. The Citizens Information Board provides information and it does so successfully. Like Deputy Ó Murchú, and I imagine every Member of the Oireachtas, I use its website on practically a daily basis. The information provided is a credit to the CIB. There is a real concern that fewer people are getting access to accurate information at a time when the whole world is concerned about accurate information. It was mentioned that the CIB could have more engagement with the likes of Google and so on. Is there a strategy beyond the bot it is generating, which may be challenging, with the number of people CIB is losing to Google searches and artificial intelligence? Ms Flanagan said the CIB is already working with OpenAI. I do not know what the figures are. In practical terms, probably 70% or 80% of people use Google to search. We are probably talking about one company on the search side. I do not know what the figures are for AI apps. I presume OpenAI is the one the highest number of people in Ireland are using. We are therefore talking about two companies. Is there a way for the CIB to enter this brave new world, in the way it did with its website, by working with these entities so people can get access to accurate information, even if they are getting it through the lens of a different entity than a website? Is that too simplistic?

No, the Deputy is describing a sort of dream state for us. That would be ideal. We have not pursued it very far yet, but it is something we have been thinking about. The other thing we have been thinking about is, if people do not find us in a Google search, how else they might find us. Last year, for the first time in many years, we ran small advertising campaigns to promote the website, citizensinformation.ie, in particular, and to remind people that it is there and is a source of information that can be used in many different circumstances. We are approaching it from a few different angles, but we are still getting to grips with it. For us, it is quite new. The issue is less than a year old and it happened very quickly.

That was my next question. I am glad to hear the CIB is doing an advertising campaign. We are holding an Oireachtas committee meeting right now. I hope it is being covered and maybe these issues will be covered more generally. It would be an exaggeration to call this an existential issue for the CIB website and its information, but if the figures are where they are now in such a short period, I can only imagine what the decline will look like. It is probably only heading in one direction. From a publicity point of view, it will be important to communicate to the public at large that the AI aggregators take information not only from CIB but also from elsewhere and that the CIB is still the only accurate source. It is the land registry of information, essentially. It is where people can get the information. Is it worth looking at something like that?

Yes, it would be worthwhile to explore that.

There is no problem if the witnesses want to plug citizensinformation.ie here. They can do so as often as they want.

I thank everyone. One of the things that ties all this together is proper testing, sandboxing and transparency. Once again, it is great to hear that the witnesses are cognisant of all that and being, as it sounds to me, quite tentative in their venturing into it, particularly in the Department of Foreign Affairs and Trade. I am sure it is tempting to go into other areas with AI. However, making sure to be ready every step of the way is hugely important. We have had incidents where that was not the case when chatbots were used. For example, there were issues in the Department of justice. Have the witnesses identified in the guidelines any limitations that allowed that example to occur? Are they familiar with any other examples? I hope they have heard about the Department of justice chatbot. Is anyone familiar with it? Perhaps Ms Maguire will answer first.

It is a point on a journey. In the development of Ask Finn, we benefited from the guidelines for the public service being in place from May last year. We also benefited from the fact that our Department brought in the policy I mentioned, which comes from the cybersecurity perspective, and that some might say was limiting. We limited ourselves in that sense and took a security approach to it.

Mr. McGinn spoke about the footprint we have all around the world, or the awareness we have of cyber risk.

I wish to return to the guidelines. Even if Ms Maguire has not heard of the example, she should note that the chatbot was essentially being deployed although it was not known to the user that it was a chatbot. Given Ms Maguire’s familiarity with the guidelines, where does she believe the weaknesses are?

By having a cross-departmental working group examining the various issues, people will be asking questions of procurement and personal data at the right time. That is how I would see it for us.

Is there anything that Ms Kennedy or Mr. Long wish to add?

We worked very closely with the Department of public expenditure and reform and the Office of the Government Chief Information Officer on the development of the guidelines. They were based on our own experience of doing work. Technology is moving extremely fast and we need to make sure we review constantly in line with technological changes. We are moving into agentic spaces, reasoning and all of that, and we need to understand how the use of reasoning might be covered. However, there are some very good broad principles – the recommendations the committee published before Christmas also have them – on the need for what we would call "the human in the loop", so it is not just about automated processing by AI only. Once you do that responsibly and have the multidisciplinary teams we spoke about, you bring all the perspectives in. However, it cannot be a case of a once-and-done exercise and saying it is on the shelf now. Things are moving so quickly in this space that constant reviewing is needed.

I apologise as I wish to return to the subject of citizens information. I have a question on safeguards. Last year in Ireland, 10% of adults started using AI in a romantic way. Many are using it as a virtual assistant. We are also hearing about suicidal ideation among teenagers and wondering where all this relates to citizens information. Have people started to talk to citizens information in the way they might talk to a friend online, by asking, for example, whether they are at risk of homelessness or in a domestic violence relationship? Have the witnesses seen anything like that? If not, it is probably going to come. Will safeguards be built in to redirect people who might not be aware the forum is not the right one in which to ask the kinds of questions I have referred to?

Yes. We have done a lot of work on prompt engineering. The types of queries and questions that can be answered are purely based on the content, and it is about public services. However, we have put in safeguards so that if someone states he or she is suicidal, he or she will be directed to the appropriate channel. We have done a significant amount of testing around that. We have been live only since 23 December. As of yet, we have not seen what has been mentioned. We have technology included within our administration portal that can flag a query not related to the public sector so we can check on it. We have no way to follow up or do anything because we are not collecting any personal information, but we do direct people to the appropriate channel.

It is a difficult one. Sometimes an AI tool is great as it gives a quick summary. Within the Oireachtas, we have Copilot. Obviously, there are some open AI sources. It is often great to get a quick summary. Yesterday, I put in the term "housing assistance". We were thinking about getting a full list of the grants available but the tool gave us the housing assistance payment, the rental accommodation scheme and that kind of thing. To return to what Ms Kennedy said, people sometimes do not know the questions they should be asking. Prompt engineering is, therefore, important. I am referring to what occurs without taking personal information, such as on whether you are a single retired person, for example. I thank the witnesses. My contribution was not a question.

I would have been surprised if the Deputy had directed a question at the Revenue Commissioners on AI being used in a romantic way.

What are the tax implications?

Has the Department of Foreign Affairs and Trade costed the use of Ask Finn?

I believe the cost for 2024 was about €25,000 and that the cost for last year was €30,000. This involved an enhancement of existing computer infrastructure like processing power, memory and physical servers.

On top of that, there were non-cost elements. For example, open source large language models were used as part of the service. A figure could be put on the time of the people who worked on the exercise as well.

Does the actual processing of it happen on servers in the Department or does it happen in a data centre?

It happens on servers in the Department, in its private data centres. No aspect of the processing of a query touches the Internet or Internet sources. The large language model knows only of the document environment it is exposed to.

It is similar to the Revenue and CIB ones, in the sense they are trained only on this. They cannot get information from elsewhere.

Exactly.

Has the Department done a cost-benefit analysis of it?

As my colleague, Ms Byrne, said earlier, it has only been up and running since December. We will do something in that regard. One of the benefits we are seeing is that the tool is integrated into the passport processing system so people are not hopping around from one piece of IT to another in their work. The benefits will not only be in time and efficiency but also an improvement in the user experience of their desktop in the passport service.

If there are efficiencies - if it works - is the implication that it will replace staff? In theory, a smaller number of staff could do the same amount of work as a bigger group of staff.

Maybe not directly. It might release people to do stuff of higher value than the straightforward processing they are doing. On the floor of the passport service, there are people who have been there for many years and have the entire organisational memory. When they go, this kind of thing will come into its own, to outline precedent rules and guidelines those people may have known.

Are there any plans to use this in a direct customer-facing way in the future?

Not to my knowledge. We prefer people to come to us for their entitlements to a passport. We are still website-driven on that at the moment. We maintain a large webchat and there are people at the end of it. Our focus at the moment is to maintain that human touch because when people are at points of distress in their passport journey, they do not want to be wound up even more.

I can imagine being wound up by Microsoft Copilot or some such.

We therefore deal with customers with people as much as possible.

Does the Department develop its own coding and programming? We spoke about language being important. Much of the languages and codes are coming in from sources abroad, for instance America. The language they use is different and their influences in certain areas would be different. How does the Department cope with that?

From my experience, these models are trained elsewhere and developed in ways we may not know about exactly. However, we assess and test each model as we use it. I am talking about large language models. They have strengths and weaknesses in different areas. That is the only thing outside our control. I do not think there is a cost-benefit to developing our own models and we do not even have the skill set to do those things. This is why there is an emphasis on open source large language models.

There was a difference between eastern and western AI. There is a difference in the systems.

I am not aware of how that pans out from what I experienced.

Some are based on social and others are based on different aspects.

As Deputy Murphy alluded to, the CIB said it developed its own AskCI tool and that Google has an AI overview in which there are some inaccuracies and outdated material. However, it was also said that the impact on the traffic to the CIB website is that there is about 19% less. Is there a danger that system could overtake the CIB's systems?

With that and plenty of other AI tools, there is a distinct possibility that is how people will interact with information in the future. That is part of the reason we are looking to explore our own tool to see how it might work. It seems likely that AI will be people's first port of call for information queries, so I think the Deputy is right.

On a funny note, talking about the chatbot and romance, there was a guy here last week who wanted to marry his chatbot.

What are the tax implications of that? Are they treated as a couple?

I thank everyone once again. I want to touch on one thing in relation to risk. One thing I have not seen to date in the risk analysis that has been done is the political risk. What I mean by that is that we are seeing that tech bros in the US, etc., are becoming far more politically involved. Thankfully, I asked the question if it was Ask Finn or Fin AI because Fin AI was developed by Intercom and the chief executive has come out with very anti-government and anti-immigration stances. My fear was that the Department of foreign affairs was contracted to such an organisation under those circumstances and, thankfully, that is not the case. There is that and then there is the likes of X and Grok recently, though that is different. It is more of a B to C type of product than a B to B or a B to G. However, in terms of risk management and assessing vendors, is the Department now looking at the political risk of software? I do not think it has been as prevalent in history as it is now. It is probably something we would not really have considered perhaps five, ten, 15 or 20 years ago but now it is very pertinent. Is that included as part of the Department's risk register and risk assessment or will it be moving forward? We can take answers from Revenue, the Department of foreign affairs and then from the CIB.

Yes. We would risk-assess vendors but we have to operate within a public procurement perspective and we would not be able to say someone could not tender because of associations per se. What we have to do is manage very closely our contractual arrangements with our vendors. On our cloud services, we would ensure all data is held within the EEA. We would ensure all data at rest is encrypted and that it is containerised so that any access is only to the underlying hardware rather than to the data. We would also have logging and auditing of all access to the data.

In that circumstance, let us take as an example - I know it is not the case - that Revenue was contracted to Grok through the public procurement service and what came out came out. Are there contractual grounds for the termination of that contract, based on what came out, in terms of the risk assessment at the moment and the procurement?

There are always termination clauses within contracts. We would work with our Revenue solicitors division to ensure our contracts were robust and would allow us terminate for various reasons. I cannot say categorically. It would have to stand up in a court of law so there would have to be evidence as to why we were terminating. A lot of this is pay-as-you-go-type services and that allows us to move in a multi-vendor framework. You are not contractually tied to using a particular service plus extensions solely for a three-year period like you would have potentially been with other types of on-premises-type technology. If we have multi-vendors, we can switch off the use of service A and move to service B and start paying as we go for that. However, we will continually have to look at our contractual arrangement. When you get into cloud service providers, contract management is a huge part of it. You move away from some of your operations within your data centre and you move into much more of a contract management role.

I take it the answer is probably the same or similar from-----

I would add to that to say that IT and AI are bringing discussions to the management board of our Department and to that of every Department far more than was the case, as regards these kind of trade-offs to which the Deputy refers. We would have a similar approach to Revenue with the moving between providers. Those are very real concerns.

When we were talking about using prompt engineering, I know in my own case that if an issue comes in about tax and I want to inform myself, I will ask, say, Copilot, using the consolidated tax Acts and including budget 2026, what the legislation says about something. I certainly find, even in what I am doing, that it can be quite useful.

I do not necessarily rely on it but it gives me general information. My final question, because of the time, is to the Department of foreign affairs on other use cases. I have been at a number of things recently, including a thing I was at last week where Amazon Web Services had partnered with a number of organisations, including GOAL, where there was discussion about using AI to be able to respond to crisis situations and in the delivery of humanitarian and overseas aid, and I was struck by the fact that they clearly identified it as an option. Have there been discussions in that area, particularly on crisis response, the use of AI in gathering data and information, and decisions being made on overseas development aid?

At the moment, the use of it is that knowledge management piece rather than as a tool to be involved in decision-making. The great thing about Irish NGOs is that many of them are early-adopters. They push the boat out and try various things from nutrition through to humanitarian aid and all of the responses. We are not there yet.

To the CIB, obviously our committee is looking at a number of recommendations, and particularly in this House we are looking at the roll-out of public services. However, part of it is also the delivery of public services to citizens, and we are going to be using AI to a far greater extent with the safeguards put in place. If the CIB were to look to us to make particular recommendations on the roll-out of AI across the public services, what should they be from its experience, given that it covers a broad range of services?

I am not sure we have enough experience yet to respond. We are about a month - not quite - into our pilot but it is something that we would be really pleased to come back to the committee on if we can share the learning we have in the next three to four months. We would be delighted to do that.

Okay. Unless anyone has any final thoughts-----

The only thing I would say on the adoption of technology, and we saw this when e-government was the word of the day, I would urge that we consider transformation, not just speeding up current processes with AI. We often talk about transformation being a caterpillar into a butterfly, not a faster caterpillar. If public services were being designed from scratch in a world where AI exists, we would need to look at how they would be designed. I am constantly trying to battle in my own head what a tax return is in the age of AI. It used to be a paper-based thing where people had to fill in numbers and aggregate it. Then we put it online and made it available but we are still aggregating numbers. Now we are moving towards integrating with natural systems of record. Ultimately, what do we need? We are always trying to ensure we collect the right tax and reduce the administrative burden so that businesses and taxpayers can go about their daily lives. I would urge looking at transformation and not just faster caterpillars.

Yes. We all want our public services to be like butterflies. I thank Ms Kennedy. That is an appropriate note on which to end. I thank all of our witnesses for their contributions and for feeding into this. We know it will influence Government policy in this area. This meeting is now concluded. We will have a short private meeting after this.