OpenAccountants
How AI answers tax questions, which sources it can reach, and where professional judgement matters.
Read the methods, findings and limitations of our studies. Each report reflects its stated study period; results have not been re-tested for this release. References to our platform describe it at that time.
Research report
13 August 2026
OA-RR-2026-01
An investigation of automated access to official tax sources: what can be reached, what blocks access, and what that means for checking an AI answer.
356 sources · 133 jurisdictions
Access was measured from one network. A refusal in this study does not establish that a source is inaccessible to every system.
Research report
13 August 2026
OA-RR-2026-02
Accountants' reviews of AI-drafted tax facts, followed by a benchmark comparing answers with and without web search. The report examines stale figures, missing conditions and confidence.
2,713 facts reviewed · 60 benchmark questions
Draft review and benchmark accuracy are different measures. The sample does not establish an accuracy rate for every model or tax question.
Working paper
Study period: 17 August–3 September 2026
The same tax questions put to chat products, coding agents and direct model interfaces, examining differences in jurisdiction assumptions, web searches and answers.
262 analysed runs · 9 constructed questions
This study measures disagreement and behaviour, not correctness. Many comparisons use small samples and need further testing.
These are research publications, not current tax advice. Tax examples belong to the periods studied and need checking against current official sources before use.
Found an error or want to discuss the method? Contact the team.