evidoria

← Back to browse

Good practice Imported

RevAssist — Irish Revenue's Internal AI Assistant, Live Since June 2024

Ireland · Dublin · See the Ireland profile · See the Dublin profile

Evidence: Observational / pre–post Top 51% 60/100 · Ask Evidence Copilot about this practice

Ireland's Revenue Commissioners built RevAssist, an internal AI assistant trained only on its 1,500 Tax and Duty Manuals. Query-routing accuracy on its MyEnquiries portal rose from about 70% to 98%, at national call volume.

45-65%
MyEnquiries correct-routing accuracy (pre-automation baseline)
70% to as high as 98%
MyEnquiries routing accuracy (post-NLP) (2024-2026)
94-97%
Query-processing success rate (2024-2026)
1,500
Tax and Duty Manuals searchable via RevAssist
~2.5 million
Annual phone calls handled by Revenue
~1.5 million
Annual electronic queries handled by Revenue
RevAssist — Irish Revenue's Internal AI Assistant, Live Since June 2024

Details

Maturity
Established
Promoter
Office of the Revenue Commissioners
Period
2024-present
Region (NUTS)
IE06
Keywords
taxation, public administration, internal operations

Context

Revenue's public-facing 'MyEnquiries' online query system was, in staff's own description, confusing for taxpayers: before automation, customer queries reached the correct back-office team only 45-65% of the time. The agency separately handles roughly 2.5 million phone calls and 1.5 million electronic queries a year.

Activities

Since June 2024, Revenue has used natural-language processing to classify and route MyEnquiries submissions. In parallel it launched RevAssist, an internal generative-AI assistant letting caseworkers search all 1,500 of Revenue's Tax and Duty Manuals in plain language, with answers citing and linking to the source manual and requiring staff verification before use. A feedback loop lets staff flag incorrect answers to refine the model. A related pilot for AI-assisted drafting of legislation-based documents began in Q4 2024, with every output still requiring human finalisation.

Results

Routing accuracy is reported to have risen from roughly 70% to as high as 98%, and query-processing success to 94-97%. Revenue's CTO Vincent Duffy has stressed the tool's deliberately narrow scope, restricted to Revenue's own guidance rather than external sources like HMRC or IRS.

Conclusions

As of mid-2026 these tools remain in production for internal research and query-routing only. Revenue has signalled interest in extending AI to phone-based security questions and automatic call summaries, but has not published an independent, third-party evaluation of the accuracy figures it reports.

Implementation
Indicative cost
Medium (€50k–€500k) — No public cost figure disclosed for RevAssist or the MyEnquiries routing NLP.
Time to results
Medium (1–3 years) — NLP query-routing and RevAssist live since June 2024; related AI-assisted drafting pilot began Q4 2024.
Staffing & skills
Revenue caseworkers/staff using RevAssist to search Tax and Duty Manuals, Revenue IT team led by CTO Vincent Duffy, Back-office teams receiving routed MyEnquiries submissions

Conditions for success

  • Deliberately narrow scope — restricted to Revenue's own 1,500 Tax and Duty Manuals rather than external tax guidance (HMRC/IRS)
  • Answers cite and link to source manuals; staff must verify before use
  • Feedback loop lets staff flag incorrect answers to refine the model

Common failure modes

  • No independent third-party evaluation of the reported accuracy figures has been published
  • Related AI-assisted legislative-drafting pilot still requires full human finalisation of every output

Where it fits

Governance type
national tax administration
Scale
national (Ireland), ~2.5M phone calls & 1.5M e-queries/year
Income level
high-income (EU/Ireland)

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful