AI for UK Tax Advisers: Research Notes, Client Letters and Checks

A UK tax adviser can use AI safely for research notes, client letters and explanations, provided each output has its own line. Research notes need every cited source opened and read. Client letters need the adviser's conclusion to come first. Computations should stay in tax software or a tested spreadsheet, with AI limited to explaining the result.

The useful question is not "should we use AI?" but "which output, and what check?" A practice that answers per output type avoids both extremes: banning everything, which pushes use underground onto personal accounts, and allowing everything, which is how a wrong rate reaches a client letter. Each line below comes with a prompt, a sample and the check that makes it hold.

Follow me on Instagram@sagnikteaches

Three output types, three lines

Write that table into your practice procedures, give each line an owner, and review it when your tools or your professional body's guidance change. The reasoning behind each line follows.

Connect on LinkedInSagnik Bhattacharya

The confidentiality and data-protection floor

Whatever the output, three rules apply first. Client information is confidential, and a client's tax affairs are also personal data under UK data-protection law (the UK GDPR and the Data Protection Act 2018), so you need a lawful basis, a clear purpose, minimal data and a supplier you can account for. In practice:

Subscribe on YouTube@sagnikteaches

• Anonymise. Replace names, addresses, UTRs, National Insurance numbers, company numbers and anything distinctive with labels. Keep the key on your own systems. The tutorial on whether ChatGPT is safe with client data sets out the account and settings questions.
• Use a business plan. ChatGPT Business, Claude Team and Microsoft 365 Copilot (now called Microsoft Copilot in Microsoft's documentation) state on their own pages that they do not train on business content by default (checked 4 November 2026). Personal accounts are for nothing client-related.
• Record it. Add the tool, its purpose and the data allowed to your data-protection records and to your engagement letter's wording. Telling clients how you work is simpler than explaining later. The engagement-letter tutorial shows suitable phrasing.

If something identifiable does go into a tool by mistake, treat it as a data incident: delete the chat if the tool allows, note what was entered and into which account, and tell the person responsible for data protection the same day.

Research notes: use AI to find leads, never to settle the rule

Tax research starts from legislation and from HMRC's own guidance: the manuals collection on GOV.UK (HMRC technical guidance, last updated on 6 October 2026 when I read it) and the legislation on legislation.gov.uk. Case law and tribunal decisions complete the picture. The manuals collection is candid about its limits: it says the guidance is not comprehensive, does not give a definitive answer in every case and is based on the law as it stood when published. A note is useful if it points to these sources accurately and says which kind of source each one is.

The results you get from AI here depend on the tool. A general chatbot answers from training data that can be out of date and may invent references. A vendor research tool built for tax typically searches a defined body of material and shows its sources, which is better, but the answer is still a summary of what it retrieved, and the retrieval can miss the relevant provision or an amendment. Treat both as a fast way to find where to look.

A research prompt that asks for leads, not conclusions

I am a UK tax adviser. Question (anonymised): does a company
director who lends money to their own company and later has the
loan repaid in instalments face a tax charge on the repayments?
List: (1) the areas of tax law that could be relevant, by name;
(2) for each, what kind of source I should open (legislation,
HMRC manual, tribunal decision); (3) key facts that would change
the answer; (4) any recent changes I should check for. Do not state
a conclusion. Do not quote section numbers unless you are sure,
and mark any you give [VERIFY].

A realistic output lists several areas (the loan account treatment, any charge on repayments in particular circumstances, anti-avoidance considerations), tells you to look in the relevant HMRC manual and the legislation, and asks about the loan's terms. Notice what you did not ask for: an answer. You then go to the sources, read them, and write the note.

The verification routine for every cited source

1. Open the source at legislation.gov.uk, in the HMRC manuals collection or on the tribunal's site. Not through a summary site or the tool.
2. Check it exists and says what the note says it says.
3. Check the version in force for the dates that matter. On legislation.gov.uk, the default text is the latest available revised version; use the timeline of changes to see earlier points in time, and read the "Changes to legislation" box, which lists amendments the editors have not yet applied to the text.
4. Remember that HMRC manuals describe HMRC's view, which is not law. A note should say which is which: "the legislation provides" versus "HMRC guidance states".
5. Save the quoted wording and the date you read it in the working paper.

A short, honest line on the file is enough: "Leads generated with AI assistance on anonymised facts; all authorities opened and read on [date] by [name]".

Test a research tool before you trust it

If you are considering a vendor research tool, run a proper trial. Write twelve questions to which you already know the answer from your own work: four easy, four needing an anti-avoidance or exception point, two on recent changes, two where the correct answer is "it depends". Score each response as right, partly right or wrong, and whether the cited sources said what the tool claimed. Record the results and run the test again after any major product update. If a tool cannot handle the "it depends" questions without fabricating a firm answer, it is a lead generator at best. Do not rely on a vendor's own accuracy claims; your questions are more relevant than their benchmark.

Client letters: AI drafts the words, you own the advice

Letters are where AI helps most, and where plain English matters most. The safe pattern is that you decide the advice first, then ask AI to draft the letter from it.

Draft a client letter in plain English (reading age about 14, no
jargon without a one-line explanation). Use ONLY the points below.
Do not add advice, reassurance or deadlines that I have not
written. Structure: what we looked at; what we found; what it
means for you; what you need to do and by when; what we will do.

POINTS: [your anonymised conclusion and facts, with figures]

Here is a before and after (illustrative). The tool's first draft of a section on a loan-account repayment says: "You can safely take this money out of the company without tax implications." You never wrote that. What you concluded was that repayments up to the balance of the director's loan account are generally not income, provided the account is genuinely in credit and records support it. The corrected letter says: "Repayments of what you have lent to the company, up to the credit balance on your loan account, should not be taxed as income, as long as the records show the money was lent. Please send us the bank statements showing the original loans so we can keep them on file." The difference is the condition, the action and the absence of "safely".

Three further checks on every letter: every number comes from the computation, not from the tool; any date is checked against the source; and the letter does not promise an outcome that depends on HMRC or a tribunal. For explaining outcomes in plain language at greater length, the tutorial on explaining tax to clients in plain English has more sample wording.

Computations: why the line is red, and what AI may do

Language models predict words. They do not run arithmetic reliably, and they tend to apply the wrong band, the wrong year's allowance or a rule that does not apply, while sounding sure. A computation that is a few per cent out is worse than one that is obviously broken, because nobody questions it.

Keep computations in your tax software or a spreadsheet you built and tested, with the year's rates and allowances typed in from the official source and dated. A second person re-performs the key figures. AI has two legitimate supporting roles here:

• Explaining the result. You paste the result (anonymised figures) and ask for a two-paragraph plain-English explanation for the client, which you then check against the computation.
• Proposing test cases. You describe the computation's rules and ask for edge cases (income exactly at a threshold, a loss carried forward, a part-year) that your spreadsheet should handle. You run them and compare with software.

A quick illustration: you describe a small spreadsheet that works out a director's extraction of salary and dividends and ask for five edge cases. The tool proposes zero dividend, salary at the exact threshold, a loss year, a mid-year start and a rate change part-way through the year. You run them. The mid-year start exposes a formula that divided by twelve regardless of the start month. A real bug, found by AI-suggested tests, not AI arithmetic. For help with the spreadsheet side, the ChatGPT and Excel guide shows safe ways to ask for formulas.

What a research note looks like before and after your edit

A first AI-assisted note often reads well and says little you can rely on. Here is a short illustration of the edit that turns a lead into a usable note.

Tool's lead (illustrative): "There may be a tax charge where a director withdraws funds from an overdrawn loan account, and HMRC guidance covers this in detail."

Your note after opening the sources: "Question: whether the director's overdrawn loan account at the year end gives rise to a charge on the company. Legislation: [section reference, read at legislation.gov.uk on the date shown]; operative wording quoted below. HMRC guidance: [manual reference], read on [date], which states HMRC's view on the timing of repayment. Facts relied on: balance at year end, repayment date, no write-off. Conclusion: a charge arises unless the balance is cleared within the period the legislation allows; the repayment on the stated date falls outside it. Confidence: medium, pending evidence of the repayment date."

The first version is a pointer. The second states the question, separates the law from HMRC's view, records when you read each source, lists the facts you relied on and gives a confidence level. That is what a reviewer needs, and what you will want to find in the file two years later if the position is questioned.

Choosing and vetting a vendor tool

Several vendors sell AI research assistants aimed at tax professionals. The tools change quickly, and any list of names would be out of date within months, so ask these questions instead. The comparison of AI tax research tools on citations, cost and checks goes further on how to judge citation quality. Put them to the vendor in writing and keep the answers.

1. Where does the data go? Are prompts and uploaded files stored, for how long, in which region, and can you delete them?
2. Is your content used to train models? Get the answer in the contract or terms, not on a sales call.
3. What is the source base? Which legislation, manuals and decisions does it search, and how often is it updated? What happens when a provision is amended?
4. Does it show sources you can open? A tool that cites a page you can click is far more useful than one that gives a bare answer.
5. What does it do when it does not know? Ask a question with no clear answer; a good tool says so.
6. What is the exit? Can you export your saved work, and what is deleted when you cancel?
7. Who is the supplier? Where is the company based, who are its sub-processors, and does it hold an independent security certification you can check?

Then run the twelve-question trial. Only after both should a tool join the approved list, and its entry should carry a review date. If the vendor will not answer the first two questions in writing, that is your answer.

Deadlines and dates: a special case

Dates deserve their own rule because a single wrong date can cost a client money and cost you a complaint. AI can list the filing and payment dates it believes apply, but those come from the official calendar for the tax year in question, copied by a person, and checked by a second. If a client letter includes a deadline, the person signing the letter confirms it on the official page the day it is sent, and records the source. The same applies to claim windows, amendment periods and anything counted in days from an event, where chat answers regularly miscount weekends and the start day.

A sensible habit is to keep a short, dated table of the dates your practice quotes most often, owned by one person and reviewed at the start of each tax year. Letters draw from that table, not from a prompt.

Internal notes and checklists: the easy green

Tidying meeting notes, turning a call recording's transcript (with consent) into an action list, producing a pre-meeting checklist from your own standard list, and drafting file notes are low-risk, because you supply the content and you read the result. They also take time away from the work you are paid for, so they are a good place to start. Keep the confidentiality floor: anonymise, business account, no identifiers.

A worked example: one new client, four outputs

Say a three-adviser practice takes on a marketing agency owner who has moved from sole trader to limited company and asks three things: how to take money out, whether a particular vehicle purchase works, and whether a loan to the business can be repaid. The adviser sorts the work by output type.

First, research: she uses the research prompt to list the areas and source types for each question, opens the legislation and manuals, and writes a two-page note quoting the operative text. One lead in the tool's list turns out to be irrelevant, one useful point (a rule she had forgotten applied) comes from the list. Second, computations: she builds the salary-and-dividend comparison in the practice's tested spreadsheet with this year's figures typed from the official pages and has a colleague re-perform two cells. Third, the letter: she writes her conclusions as six bullet points, asks AI to turn them into a plain-English letter, and corrects two sentences. Fourth, the file note and next-steps checklist: the tool tidies her call notes into an action list, which she reads and files. The tool's list held nine leads; five were dead ends, three confirmed what she expected and one was the forgotten rule, so her two-page note quotes four authorities, each opened and dated. The letter grew from six bullets to about 400 words, and one of the two sentences she corrected had added "you will save" wording she never wrote. Total AI use: four prompts; every substantive check stays human. The time saved is real but modest, mostly on drafting; the benefit that surprised her was the research-lead prompt, which made her read the legislation more carefully than she otherwise would have, not less.

Writing the lines into practice procedures

Put the table above in the practice manual and add: the approved tools and accounts; the anonymisation rule; the verification routine and where the log is kept; who may approve a new tool; how clients are told; and an incident procedure. Add a requirement that anyone who signs a letter or note has read the sources it relies on. Review the procedure at least twice a year and whenever a vendor changes its data terms, because plan terms and tool behaviour change quickly. The tutorial for clients on choosing an accountant who uses AI well shows what clients will now expect you to say, which is a good test of your own procedures.

Where it goes wrong

• A confident but outdated rule. A tool quotes last year's relief or rate. Show-up: legislation.gov.uk or the manual says otherwise. Fix: rates and thresholds always typed in from the official page, with a date.
• Manual treated as law. A note says "the law requires" when it cites HMRC guidance. Show-up: a reviewer asks for the statute. Fix: separate "legislation provides" from "HMRC guidance states" in every note.
• Letter reassurance. "Safely", "no problem", "guaranteed" slip in. Show-up: a client quotes your letter back to you. Fix: a find-and-check for those words before sending.
• Computation drift. A pasted figure from a chat. Show-up: the software disagrees by a small amount. Fix: figures only from the computation.
• Identifiers pasted in a hurry. Show-up: a UTR in a prompt. Fix: delete, report, and make the anonymising step part of the template.

A fortnight to put the lines in place

Week one: write the table of lines, choose the approved account and set the anonymisation rule. Run the research-tool test with twelve questions if you are considering one. Week two: use AI for internal notes and one client letter, with the checks above, and compare time and corrections with your last similar letter. After a month, decide whether to widen into research leads. Expect the green uses to settle in quickly; the amber ones need the discipline; the red one stays red. If you want to map the lines with someone, that is what my AI implementation consultation covers, and you can begin at the contact page.

UK tax advisers and AI: common questions

Can I rely on an AI tax research tool's answer?

Not on its own. Vendor research tools usually cite sources, which helps, but you still open the cited legislation or manual page and read the wording. Test any tool on a dozen questions where you already know the answer and keep a record of how often it was right, partly right or wrong.

Is putting client details into AI a data-protection issue?

It can be. Client names, income and family details are personal data under UK data-protection law, and confidentiality duties apply as well. Anonymise first, use a business plan whose terms exclude training on your content, record the tool in your data-protection documents and be able to explain the processing to a client who asks.

Which outputs are safest for AI to draft?

Client letters that explain a conclusion you have already reached, internal summaries of your own notes, and checklists. The riskiest are research conclusions and computations, because an error there changes the advice. Draft where you supply the substance; verify hard where the tool supplies it.

Should I let AI do tax computations?

No, not without a check. Language models make arithmetic and rule-application slips that look plausible. Build the computation in your tax software or a tested spreadsheet, and use AI only to explain the result in words or to propose test cases that your computation should handle.

Further reads

• How Accountants Use AI to Spot Errors in Client Books — AI as a second reader on client figures.
• Tax Software With AI: What a Small Business Should Check — What to check in tax software that has AI features.
• How Small Tax Practices Use AI Through the Busy Season — Fitting AI into a small practice's peak weeks.
• Which Legal Tasks Should a Small Firm Never Hand to AI? — A parallel list of tasks professionals keep human.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check the practice is ready before widening AI use.
• Can AI Do My Business Taxes? What It Can Safely Prepare — What AI can safely prepare on tax work.
• Can AI Draft an HMRC Penalty Appeal? Reasonable Excuse, Checked — AI can structure a penalty appeal from the client's own facts. The accountant's pre-send checks on dates, excuse, evidence and tone are what make it safe.
• Can AI Help Prepare an R&D Tax Credit Claim? Narratives and Costs — How to use AI to draft R&D claim narratives from project notes and cross-check cost schedules, with eligibility left to the specialist.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: GOV.UK's HMRC manuals collection (last updated 6 Oct 2026, read 4 Nov 2026), legislation.gov.uk version and changes features (read 4 Nov 2026), vendor privacy and plan pages for ChatGPT Business, Claude Team and Microsoft 365 Copilot (checked 4 Nov 2026), and UK data-protection principles; confirm anything you rely on at the primary source.