Transforming the delivery of services with AI
Revenue Commissioner Ruth Kennedy discusses how Revenue has integrated artificial intelligence (AI) into its operations.
In 2012, Revenue won a Public Service Excellence Award from the Government for its PAYE real-time risk system which used predictive analytics to detect fraud in PAYE tax refund claims.
“In Revenue, our mission is to serve the community by fairly and efficiently collecting taxes and duties, and implementing customs control,” says Kennedy, who asserts that an organisation’s AI use must align with its mission.
The Commissioner says trust is essential for Revenue to deliver on this mission as its operations hinge on a culture of ‘voluntary compliance’.
Governance
Kennedy outlines Revenue’s bespoke governance framework underpinning its use of AI. The framework is shaped by the EU AI Act and public sector guidelines. It is designed to ensure there is oversight of AI use, and safeguard human agency and data privacy.
“You can take those overarching guidelines, but you have to localise them and you have to build them into your own corporate governance,” explains Kennedy.
The Commissioner explains why the technology “is actually not intelligent”. She states that AI “does not understand the words you put in”, adding that “each word is merely a token”. The technology then “comes back with a set of tokens that it thinks represent the answer”.
“It does not understand the question nor does it understand the answer,” says Kennedy.
Organisation’s must understand how large language models (LLMs) work, the guardrails they require, and the skills staff need to use them. Decision-making and verification skills must be built into staff to ensure they can use LLMs.
“With the governance and the training in place, we have gone from the pilot stage to rolling things out enterprise-wide,” says Kennedy. “Pilots are quite easy to do. When you need to scale up, that is where it is much trickier.”
Implementation
Revenue has implemented multiple AI solutions such as RevAssist which was rolled out in June 2024. RevAssist is an LLM used by staff to answer questions on the organisation’s tax and duty manuals (TDMs), of which there are approximately 1,500. It is trained exclusively on these manuals which ensures the data is high quality as it is not built on unverified sources.
RevAssist answers also include links to the sources of the manuals used to create the answer, and staff are obliged to verify the answer provided. A feedback loop is in place for staff to log when RevAssist provides an incorrect answer to ensure it is continuously improving.
“It is extremely useful because if you are answering a phone, you could get a question on absolutely anything,” says Kennedy, who explains that it enables staff to answer queries quickly.
She states that Revenue “took that to the next level” by building an agentic AI model to create or update TDMs. Every year, Revenue must create new and update existing manuals on foot of the Finance Bill which underpins changes to tax legislation following the Budget.
The agentic AI model is trained on Revenue’s processes and guides for creating or updating manuals. It produces draft manuals which Kennedy says “will get you 70 per cent of the way” to a final document.
The same governance applies in that staff must review these drafts, and TDMs are only published after review at senior level. The AI generated draft significantly reduces the time it takes for guidance on new legislation to be created.
As the capabilities of AI evolve, Kennedy says a future development that could be explored is removing the need for TDMs, as AI assistants trained on Irish tax legislation would answer questions without a need for a TDM.
Furthermore, Revenue is using LLMs to identify how users’ enquiries should be directed, analyse documents, and craft legislation. The organisation is developing an AI-enabled self-service hub to include these activities. Crucially, these will be trained using curated data and will provide links to the sources for answers.
Revenue is also assessing how AI can improve every facet of the customer service journey. Currently, clerical staff must summarise each customer call and cannot answer the phone while doing so. AI can be used to complete summaries, freeing up staff.
Considerations
Kennedy asserts that there is a discipline required in using any technology. This discipline must be exerted during analysis, testing, operations, and ongoing maintenance of technology.
Organisations must also commit to technological developments which require “ongoing investment”. Kennedy continues: “The total cost of ownership is about 20 per cent to implement and 80 per cent to maintain.”
Committing to these developments is also made difficult by their rapid rate of change. Revenue has implemented models in some cases which subsequently needed to be changed “because the underlying vendor has changed what they are doing”.
AI is mostly cloud-based and Kennedy asserts that organisations must be comfortable with their data being processed in the cloud. Because AI use is on a pay-per transaction basis, organisations must budget carefully to ensure continuity of service. Kennedy asserts that providing staff with the requisite skills is critical. “You must ensure that you do not outsource AI. You need to start building the skills internally.”
Concluding, Kennedy states that organisations must ask themselves: “How can we transform public service to be different? If we were designing public service from the get-go using AI, how would we design it?
“We would not design it the way we have it now, so do not automate what we have now using AI because then we are missing a trick.”
