AI tax research evaluation checklist

Sources#

1. Open three citations

   Each should land on the exact section, from the official publisher, not a summary or a search page.

2. Ask about a past year

   The answer should use the law and rates in force then, and say which year it used.

3. Ask about the wrong country

   A Canadian question should never come back with US rules in it.

4. Check the date of the source

   You should be able to see when each source was published or last checked.

Refusals#

1. Ask something it cannot know

   It should say it does not know, and why, rather than produce a confident answer.

2. Ask for a figure without the facts

   It should ask for the missing fact, not assume it.

3. Ask about an unsupported country

   It should say the country is not covered.

4. Change one input

   The figure should move the way the law says it should.

Review and deliverable#

1. Hand it to a reviewer

   Can they see every source and decide on each finding?

2. Look for the trail

   Who asked, who reviewed, what changed and when should all be recorded.

3. Check data use

   Your client data should not be used beyond your work without your consent.

4. Export the result

   It should leave as a workpaper or client pack with its sources attached.

Related guides, and why

• Guide

  Transfer pricing year-end documentation checklist 2026

  What to have in place at year-end for each controlled transaction: the delineation, the functional analysis, the method and its reason, and the evidence a local file rests on.

  Also about recordkeeping

• Guide

  UK MTD quarterly update review checklist 2026

  Who must use Making Tax Digital for Income Tax from April 2026, the quarterly update deadlines, and what to check before each update is sent.

  Also about recordkeeping