AI for Tax Advisory: Drafting Client Advice Memos Safely

A tax advisory practice can draft client advice memos with AI safely if it sends the tool only anonymised facts, uses a business plan that does not train on its content, treats every authority the draft cites as unverified until someone opens the primary source, and keeps a partner's review before anything leaves the building.

Most firms get the order wrong. They start by asking AI to find the law, which is the riskiest job on the list, and only later worry about what they pasted in. Start the other way round: use AI for structure and plain wording on facts you have already cleaned, and bring it near research last, with a verification step that cannot be skipped.

Follow me on Instagram@sagnikteaches

Where a memo can go wrong before the AI has typed a word

An advice memo is a small, structured argument. It states the question, sets out the facts the advice relies on, identifies the rules that apply, applies them, gives a conclusion with a confidence level, and lists what would change the answer. Every one of those parts can fail when AI is involved, and each fails differently.

Connect on LinkedInSagnik Bhattacharya

• Facts fail through leakage. The prompt carried a name, an address, a tax reference or a company number that did not need to leave your systems.
• Rules fail through invention. The draft cites a provision, case or guidance note that does not say what the draft claims, or that never existed.
• Application fails through smoothing. The draft glosses a fact that does not fit, because the fluent version reads better than the honest one.
• Conclusion fails through overconfidence. A hedged professional judgement comes back as a flat statement.

The workflow below puts a control against each of those four failures. None of it needs special software; it needs a written routine and someone who follows it.

Subscribe on YouTube@sagnikteaches

Choose where the drafting happens

Where you run the prompt matters more than which model sits behind it. A free or personal account may use conversations to improve the product unless a setting says otherwise, and the settings change. Business plans are different: ChatGPT Business and Claude Team (both need at least two seats) 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; I checked all three on 4 November 2026. Check the current terms page for the plan you buy, and keep a dated copy of what it said in the practice's tool register.

A sensible rule for a small practice: only the approved business account is used for client work, and personal accounts are for nothing client-related at all. If you have not written that down, do it this week. The safety tutorial for accountants using ChatGPT covers the account questions in more depth.

Two settings are worth confirming on whichever plan you choose: whether chat history is retained and for how long, and who in your organisation can see shared projects. A shared project that holds the memo template should hold no client material at all; keep facts in individual chats you can delete once the memo is final.

Even on a business plan, minimise. The tool does not need to know who the client is to help you write a paragraph about the facts, so do not tell it.

Step 1: Turn the file into a fact sheet with labels, not names

Before any prompt, write a one-page fact sheet by hand, from the file. This is the step that makes the whole workflow safe, and it also improves the memo, because a short list of facts exposes gaps that a long file hides.

Replace identifiers with role labels and keep what matters to the tax analysis: amounts, dates, percentages, entity types, residence status, relationships between parties, and the sequence of events. Drop what does not: names, addresses, tax references, account numbers, company numbers, employer names, the client's occupation if it is unusual enough to identify them, and any free-text notes from emails.

A filled-in fact sheet (illustrative)

QUESTION: Can Client A treat a transfer of shares to Client B
(spouse) as no-gain-no-loss, and what is the exposure if the
shares are later sold?

FACTS
- Client A: individual, owner-director of Company X (trading company).
- Client B: spouse of A, lives with A, both resident throughout.
- Company X: 100 ordinary shares, A holds all 100, incorporated 2014.
- Proposed: A transfers 40 shares to B in the current tax year,
  for no consideration.
- Market value of the 40 shares: approx. 220,000 (valuation letter
  on the file, dated within the last 6 months).
- Intended later sale: Company X sale discussed for 18-24 months' time;
  no heads of terms signed.
- B has no other income this year except a part-time salary of 14,000.

NOT KNOWN: whether B will take an active role in Company X; whether any
shareholder agreement restricts transfers.

Notice the last line. Writing down what you do not know tells the model, and the reviewer, where the memo must hedge. Without it the draft will usually assume the convenient answer.

A second check on the fact sheet catches what a find-and-replace misses. Read it as a stranger would: could anyone who knew the client work out who this is from the combination of facts alone? A 2014 trading company, an owner-director, a spouse on a part-time salary and a sale in two years is generic. A rare product, a village name or an unusual family structure is not. Generalise it ("a manufacturing business", "a regional chain") or leave it out if it does not change the tax answer.

The tutorial on anonymising client data before AI has a longer method, including how to keep a private key that maps labels back to the real client so the finished memo can be re-personalised inside your own systems.

Step 2: Use a memo template the AI has to fill, not invent

If you ask for "a memo on this", you get a generic essay in the model's house style. If you hand over a template and your fact sheet, you get a draft shaped like your firm's memos. Keep the template in a shared project or a saved prompt so everyone uses the same one.

You are helping a tax adviser draft an advice memo.
Use ONLY the fact sheet below. Do not add facts. If a fact you need is
missing, list it under "Information needed" instead of assuming it.

Structure:
1. Question (one sentence)
2. Short answer (max 80 words, with confidence: high / medium / low)
3. Facts relied on (bullets, copied from the sheet)
4. Issues (numbered; for each, state the rule in general terms and
   mark any specific citation as [CITE TO CHECK])
5. Analysis (apply each issue to the facts)
6. Risks and what would change the answer
7. Recommended next steps
8. Information needed

Do not cite sections, cases or guidance numbers from memory. Where you
think a source exists, describe it ("the rule on transfers between
spouses") and mark [CITE TO CHECK]. Plain English, short paragraphs.

FACT SHEET:
[paste the anonymised sheet]

Two instructions in that prompt do most of the work. "Use only the fact sheet" cuts down invented facts. "Mark specific citations as [CITE TO CHECK]" turns the most dangerous output into a visible to-do list instead of a confident-looking reference that someone might trust.

What the draft looks like, and what you fix (illustrative)

A typical first draft on the spouse-transfer question comes back with a short answer along the lines of "Yes, the transfer should be treated as made at no gain and no loss, subject to the spouses living together throughout the year, so the exposure arises on a later sale". That is a decent skeleton. Three things need fixing every time:

1. It states the conclusion as "high confidence" because nothing in the sheet contradicted it. You downgrade it to medium until the shareholder agreement and B's role are known.
2. It treats the later sale as a simple disposal by B and skips the question of whether the arrangement is genuinely a gift or a device to use B's lower rate band. It has no facts on that, so it says nothing. You add the point, because a good adviser would raise it.
3. It writes "[CITE TO CHECK]" three times, correctly. You will resolve each in the next step.

The short answer, before and after your edit (illustrative)

Draft: "Yes. The transfer will take place at no gain and no loss, and the later sale will be taxed at B's lower rates."

After the adviser: "Probably yes, on the facts we hold. A transfer between spouses who live together is normally treated as made at no gain and no loss. Whether the later sale is taxed on B depends on the transfer being an outright gift with no understanding that the proceeds come back to A, and on the shareholder agreement, which we have not seen. Confidence: medium."

The edit adds the two conditions the facts leave open, names the document still missing, and replaces "will" with a stated confidence level. A partner reading the second version knows at once what would change the advice.

The draft saved you a blank page and gave you a checklist. It did not give you the advice. That distinction is the safe-use rule for the whole practice.

Step 3: Verify every authority against primary sources

This is the control that protects the firm. Generative models produce text that looks like a citation because they have seen many citations, not because they looked anything up. A professional body's generative-AI guidance for solicitors (dated 30 June 2026) calls this hallucination and cites a 2023 case in which a litigant in person put fictitious, AI-generated case citations before a tax tribunal. A tax memo is exposed to exactly the same problem.

Make citation checking a defined task with a log, not a vague intention to "have a look".

1. List every authority in the draft: each statute section, regulation, case, ruling, manual reference and guidance note, including any you added yourself.
2. Open the primary source from the official publisher: the legislation database, the tax authority's own manuals and guidance, the court or tribunal's published decision. Do not accept a summary site, a commentary page or another AI answer as the source.
3. Check four things for each: it exists; it says what the memo says it says; it was in force on the relevant dates; and it has not been amended, superseded or overturned since.
4. Record the result in a small table at the end of the working paper (reference, source opened, date checked, checked by, note).
5. Quote, do not paraphrase, any wording the advice depends on, so the reader can see the operative text.

That third row is the realistic one. A draft that names a case you cannot find is not an unlucky glitch; it is the failure mode you built this step to catch. When it happens, delete the reference, tell the partner, and re-read the paragraph it supported, because the argument may have been resting on it.

Using AI to help with checking, not replace it

You can ask a model to extract every citation from a draft into a list, which saves ten minutes of highlighting. You can also ask it to compare a quoted passage you pasted from the official source with the sentence in the memo and say whether they differ. What you cannot do is ask it whether the citation is correct. It will say yes.

Step 4: Stress-test the analysis with a second prompt

Once the draft and the citations are settled, run a different prompt that argues against it. This is cheap and catches the smoothing problem.

Act as a sceptical reviewing partner. Below is a draft memo and the
fact sheet it relies on. Do not rewrite it. List:
1. Any statement in the memo that is not supported by the fact sheet.
2. Any fact in the sheet that the analysis ignores.
3. Any conclusion stated with more confidence than the facts justify.
4. Three questions a client's other adviser might ask that the memo
does not answer.

DRAFT: [paste]
FACT SHEET: [paste]

Treat the output as a list of prompts for your own judgement, not as findings. Keep the reply in the working paper next to the draft, with a line noting which points you accepted and which you rejected and why, so a later reviewer can see the challenge happened. Some points will be wrong. The value is that two or three will be right and you had not spotted them.

A typical reply on the spouse-transfer draft (illustrative) lists seven points. Three earn their place: the analysis assumes B will still own all 40 shares when the company is sold, which the sheet does not say; it treats the 220,000 valuation as current although the letter is up to six months old and a sale is now in view; and it never asks whether the shareholder agreement, listed as not known, could block the transfer altogether. Two are generic ("consider the clients' wider estate planning") and two repeat points the memo already makes. You accept three, reject four, and write one line on each.

Step 5: Partner review, and what the partner should actually look at

The reviewing partner is the last control, and also the one most at risk of being worn down by volume. Faster drafting means more memos reaching the partner's desk, with polished prose that invites a skim. Give them a cover sheet so that the review is targeted rather than a general read.

• Which parts were AI-assisted (structure, wording, challenge prompt) and which were written by the adviser.
• The citation log, completed, with any removals noted.
• The list of "Information needed" and whether the client has now supplied it.
• The confidence level in the short answer and why it is that level.
• Confirmation that identifiers were re-inserted only in the firm's own document system, not in the AI tool.

A short, honest sign-off statement on the file helps later: "Drafted with generative-AI assistance on anonymised facts; all authorities verified against primary sources on [date]; reviewed and approved by [partner]." It is not a legal shield, and it does not replace professional judgement. It does show the process worked.

Four jobs, ranked by risk

Not every use needs the same caution. This ranking helps a team decide where to begin.

Begin at the top, build the habit, and move down only when the checks are routine. Many practices will stop at the third row, and that is a perfectly respectable place to stop.

A realistic week: a three-partner practice trying this on one matter

Say a three-partner practice with two associates is asked by a marketing agency owner whether to move from sole-trader accounting into a company at the start of the next year. The associate takes the file and writes a one-page fact sheet in twenty minutes: turnover band, profit, drawings, number of staff, a planned hire, a spouse who does the bookkeeping, no property in the business. Names and the agency's trading name are not in it.

She pastes the fact sheet and the template into the practice's approved ChatGPT Business project. The first draft takes four minutes. She spends fifty minutes rewriting the analysis section, because the draft compared two tax outcomes using rates it did not state, and she replaces them with a small computation in the practice's own spreadsheet. She has seven citations to check and finds six are fine; one guidance reference is real but covers a different type of income. She fixes it, runs the sceptical-partner prompt, and adds two risks about the spouse's work and the cost of running payroll.

The partner receives a memo with a cover sheet and a completed citation log. Her review takes fifteen minutes because she knows where to look. Before the partner signs, the associate re-inserts the client's name inside the firm's document system. Total elapsed time is about two and a half hours against a typical four, but the saving is not the point. The memo is cleaner on structure and the check trail exists.

Computations stay outside the chat

Do not let a language model do the sums in an advice memo. It can write a formula or describe how to set one up in a spreadsheet, but arithmetic produced in prose is where small, plausible errors hide. Build the comparison in a spreadsheet you control, check it with a second method, and paste the result into the memo. If you want help with the formulas themselves, the ChatGPT and Excel guide shows how to ask for them safely.

The same rule applies to dates, thresholds and rates. Anything that changes with the tax year comes from the official table for that year, copied by a person, not from the model's memory.

When a draft should be thrown away

Some drafts are not worth repairing. Delete and restart from a better fact sheet if any of these show up:

• The draft introduced a fact that was not on the sheet and built an argument on it.
• More than a third of the citations cannot be found.
• The short answer and the analysis disagree with each other.
• The tone promises certainty the facts do not support ("this will qualify").

Each of these tends to mean the fact sheet was thin or the question was two questions. Split it, tighten it and run again.

Writing the policy so it survives a busy season

A one-page policy beats a long one. Cover who may use AI on client work and in which account; what may be pasted (anonymised facts only) and what may not (anything with an identifier, original documents, tax returns); the citation-checking rule and where the log is saved; partner sign-off; and what to do if someone pastes something they should not have (tell the practice's data lead the same day, and note which tool and account it went into). Review it twice a year, because plan terms, settings and the tools change. The readiness checklist for professional firms is a useful starting point, and the engagement-letter tutorial shows how to describe your AI use to clients in plain terms.

If you would rather map this with someone who has done it, that is what my AI implementation consultation covers; you can reach me through the contact page.

Questions tax advisers ask before using AI on memos

Can I paste a client's tax return into an AI chatbot?

Not into a consumer account. A return carries names, identifiers, income and family detail. Use a business plan whose terms exclude training on your content, and even then send only the figures and facts the question needs, with identifiers replaced by labels such as Client A.

Who is responsible if the memo cites a rule that does not exist?

The adviser who signs it, and the firm. AI makes confident errors, so the practice has to treat every citation in a draft as unverified until someone has opened the primary source and checked the wording, the date and that it still applies.

Should clients be told AI helped draft their memo?

Check your engagement letter and your professional body's guidance. A common approach is a short line saying generative tools may assist drafting under partner supervision and that client data is not used to train them. Say what you actually do, no more and no less.

Which memo tasks are safest to give AI first?

Structure and plain-English rewriting. Asking AI to organise your own notes into issue, facts, analysis and recommendation carries little risk because you supply the substance. Asking it to find the law is the riskiest task and needs the heaviest checking.

Further reads

• How Accountants Use AI to Explain Tax to Clients in Plain English — Turn the finished advice into a client-friendly explanation.
• How Small Tax Practices Use AI Through the Busy Season — Fitting AI into a small tax practice's peak weeks.
• Which Legal Tasks Should a Small Firm Never Hand to AI? — A parallel list of tasks professionals should keep human.
• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven ways AI use can damage client trust.
• What to Check in an AI Tool's Privacy Policy and Terms — What to read in a tool's privacy policy before buying.
• How Accountants Use AI to Spot Errors in Client Books — Using AI as a second pair of eyes on client figures.
• AI Tax Research Tools Compared: Citations, Cost and Checks — How AI tax research tools compare on citations and checks.
• AI Software for Tax Professionals: Run a Pilot and Measure Cost per Return — A four-week pilot plan for a small tax practice: one task, pass or fail rules, a consent and WISP check, and a cost-per-return formula.
• Tax Preparation Automation: Source Documents to Software With AI — Map W-2, 1099 and K-1 fields, set accuracy checks and handle scans, so AI extraction feeds your tax software without re-keying errors.
• Tax Services AI for a Small Practice: Intake, Chasing and Review Notes — A three-part workflow for a small tax practice: intake checklist, document-chasing emails and reviewer notes, with a named person owning each step.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: vendor privacy and plan pages for ChatGPT Business, Claude Team and Microsoft 365 Copilot (checked 4 Nov 2026), a professional body's generative-AI guidance for solicitors (30 Jun 2026), and standard memo-writing practice.