How Accountants Use AI to Explain Tax to Clients in Plain English

Give the AI the finished numbers and your technical conclusion, then ask it to rewrite the explanation for a named reader at a set reading level, with one worked figure and no new facts. The AI handles the wording, not the tax: every rule, rate and date comes from you, and you check the draft against your working papers before it goes.

That split is the whole method, and it exists because of how chat assistants fail on tax. Ask one "why might my client's tax bill be higher this year?" and it will list reasons drawn from whichever tax systems dominated its training data, including allowances and thresholds that may not exist for your client, stated with complete confidence. The explanation must never be the place where the tax is worked out. You've already done the computation; the AI's job is to make it readable.

Follow me on Instagram@sagnikteaches

Who does what in a plain-English explanation

Write the facts block first

A facts block is a short, structured summary of the computation, written for the AI rather than the client. It takes three or four minutes once you've done the work, and it is what keeps the model from inventing. Here's a filled-in one for an illustrative client, a self-employed financial planner:

Connect on LinkedInSagnik Bhattacharya

FACTS BLOCK (use only these facts)
Client first name: [first name]
Reader: self-employed, comfortable with numbers, not with tax terms
Tax due for the year: $21,400 (last year $14,900)
Main reason 1: profit rose from $71,000 to $92,000
Main reason 2: last year's bill was reduced by an $8,000 loss
  brought forward from the year before; there is none this year
Not a reason (client may assume it is): no change to the rates
  that apply to the client
Payment: $21,400 due by [date from computation]
What the client needs to do: set aside the balance; reply to confirm
  whether to set up a payment plan
Tone: reassuring, direct; the client is anxious about the increase

Two lines in there do most of the protective work. "Use only these facts" sets the rule. "Not a reason" heads off the model's instinct to explain an increase with a rate change, which is the most common invented cause, and it pre-empts the question the client was going to ask anyway.

Subscribe on YouTube@codingliquids

The rewrite prompt, with the reading level set

Using only the facts block below, write an email to the client
explaining their tax bill.
- Reading level: a Flesch Reading Ease score of 60-70.
  Sentences under 20 words. No tax jargon; if a term is
  unavoidable, explain it in the same sentence.
- Structure: what the bill is; why it is higher (the two reasons,
  biggest first, with one worked figure); what is NOT the reason;
  what to do next and by when.
- Under 220 words.
- Do not add any rule, rate, threshold, date or figure that is not
  in the facts block. If something is missing, write [CHECK].
- Do not give advice or suggest ways to reduce tax.
{facts block}

Illustrative output from that prompt, lightly trimmed:

"Hi [first name], your tax bill for the year is $21,400, up from $14,900 last year. I know that's a jump, so here's why. The main reason is that your business did better: your profit rose from $71,000 to $92,000, and tax follows profit. The second reason is a one-off from last year. Your bill then was reduced by an $8,000 loss carried over from the year before. That loss has now been fully used, so there's nothing to reduce this year's bill. You'll have that benefit again only if the business makes a loss in future, which is not something to plan for! The rates that apply to you haven't changed. The $21,400 is due by [CHECK]. Could you reply to let me know whether you'd like to set up a payment plan? Best, [your name]"

Before sending, three edits. The line about having the benefit "again only if the business makes a loss" is invented reassurance in an odd tone; it's also a statement about future tax treatment you didn't supply. Delete it. The exclamation mark goes with it. The [CHECK] on the date is correct behaviour, because the placeholder in the facts block was never filled; fill it. Everything else matches the computation. That's the typical pattern: the draft is 90% right, and the 10% that's wrong is an extra sentence the model added to be helpful.

The model also takes its tone from what it assumes about the news, and a lower bill is where that goes wrong. Run the same prompt for an illustrative bakery owner whose bill fell from $11,200 to $6,800 and the first draft opens "Good news! Your tax bill has dropped by $4,400 this year." The bill fell because profit fell after the bakery lost its biggest wholesale customer, which the owner knows only too well. A cheerful opening reads as if the accountant hasn't noticed. The fix is one line in the facts block, "Context: profit fell after a major customer left; don't present the lower bill as good news", and the second draft opens with the figure and the reason, plainly, then moves to what's due and when.

Measuring readability instead of guessing it

"Plain English" is easy to claim and hard to judge on your own writing. Word will score a draft for you. In Word for Microsoft 365, go to the Home tab, choose Editor, then Document stats. Alternatively, tick "Show readability statistics" under File, Options, Proofing, and the scores appear after a spelling and grammar check. You'll see a Flesch Reading Ease score (higher is easier) and a grade level. Microsoft's own guidance suggests 60 to 70 for most standard documents. Explanations of tax bills written by accountants for accountants usually score in the 30s or 40s.

Use the score as a check, not a target to game. A draft can score 70 and still bury the one sentence the client needs. Read it once as the client would: could they tell someone else, in one sentence, why the bill went up and what to do next? If not, it isn't plain yet, whatever the number says. Checking AI content for plain English and readability has more on reading-level tools.

The reading level also has to suit the reader rather than a house default. An illustrative case: a retired engineer with rental income and a small pension asked why his advance payments had changed. The draft written to the 60-70 target came back in short, simple sentences with no table, and he replied asking for "the actual numbers, laid out properly". The fix was in the facts block's Reader line, changed from "comfortable with numbers, not with tax terms" to "detail-minded; wants every figure and how it was reached; still avoid jargon", with the prompt allowing a small table. The second draft scored lower on reading ease, in the low 50s, and was the better letter for him. The score tells you whether the sentences are easy; only the reader tells you whether the letter is right.

Before and after: one paragraph a client actually has to read

Here's the kind of paragraph that often ends up in year-end letters, followed by what the method produces from the same facts for an illustrative client who owns a small recruitment agency and takes most of their income as dividends.

Before: "Your liability has increased principally due to the higher level of distributable profits extracted by way of dividend, which, when aggregated with your other income, has resulted in a greater proportion of that income falling within the higher band, together with the restriction of the relief previously available."

After: "You paid yourself more in dividends this year: $64,000, up from $40,000. Added to your salary, that pushed more of your income into the band that's taxed at a higher rate. There was also a relief you had last year that you no longer qualify for. Together, those two changes explain the higher bill."

The "after" version needed the facts block to say which relief and why it stopped, so the reviewer can check it's accurate. The model can simplify the sentence, but it can't know that the relief was restricted rather than removed, and the difference matters if the client asks.

Five explanations worth building once and reusing

Most of the plain-English work in a year comes from the same handful of questions. Build a facts-block template for each and you'll rarely start from nothing.

1. "Why is my bill higher than last year?" The facts block needs both years' figures, the two or three real causes ranked by size, and at least one "not the reason" line.
2. "Why do I owe tax on money I haven't taken out?" For owners who confuse profit with drawings. Needs profit, drawings and the cash left in the business. An analogy that works: "tax follows what the business earned, not what you moved to your personal account."
3. "Why wasn't this expense allowed?" Needs the item, the amount and your reason in one terse line. Keep the rule itself out of the AI's hands; say "the rules for this type of cost" and let your reason do the work. For an illustrative personal trainer who claimed $2,400 of home gym equipment used mostly for family workouts, the facts-block line was "Home gym $2,400: not claimed in full; used mainly privately; we claimed the 20% business share, $480." The draft explained it in three sentences: what was claimed, why only part of it, and that keeping a simple log of client sessions at home would support a larger share in future. That last sentence was the model's addition, and it's advice, so it came out and went into the accountant's own follow-up note instead.
4. "What are these advance payments?" Only relevant where the client's tax system collects tax in instalments. Needs each amount and date from the computation, and a one-line reason instalments exist. This is where invented dates most often appear, so check every one.
5. "What happens if I pay late?" Needs your firm's approved wording on consequences. Don't let the model estimate interest or penalties; give it a sentence to use verbatim.

For a practice whose clients are mostly landlords, explanation 2 becomes "why do I owe tax when my rent all went on the mortgage?", and the facts block has to separate interest from capital repayments, because the treatment of each depends on the rules that apply to the client. That separation is yours to supply; the AI only phrases it.

A season of year-end letters, before and after

Consider an illustrative sole-practitioner accountant who sends about 60 year-end explanation letters. Written from scratch, each takes around 25 minutes to get from computation to a letter the client understands: roughly 25 hours across the season.

With the method: four minutes on the facts block, a minute for the draft, four or five minutes reviewing and editing. Call it ten minutes a letter, or about ten hours in total, plus two hours in the first week building the five templates and testing the prompt on last year's letters.

The review is where the time is well spent. In one realistic slip during a season like this, a draft for a client with instalment payments converted "second advance payment of $3,200" into "your $3,200 refund", because the facts block had the payment listed under a heading called "balance". The client would have expected money back. The reviewer caught it because the checking routine below includes reading every figure in the draft against the computation, not just skimming for tone.

Time saved is only half the test; the other half is whether clients understood. The simplest measure is the follow-up email. Count the replies that ask what a letter means ("so do I owe this or not?", "what's a loss brought forward?") across one season, then count them again the next. Illustratively, if 60 letters drew 21 clarifying replies last year and 8 this year, that's 13 fewer back-and-forth threads, each of which typically costs another 10 to 15 minutes to answer. Keep the questions themselves too: each one points to a line in a facts-block template that needs rewording.

The four checks before an explanation leaves the office

1. Every figure matches the computation. Tick each number in the draft against the working papers. Amounts, years, dates, and whether each is money owed or money coming back.
2. No new facts. Anything that isn't in the facts block gets deleted, however plausible. Watch for rates, thresholds, "usually", "most people" and statements about next year.
3. No advice drift. Delete "you should", "you could consider" and "a good idea would be" unless you intend to give that advice and have documented it.
4. Tone fits the client. A worried client needs the answer in the first sentence; a detail-minded one wants the figures in a small table.

A short list of words that should make you slow down in any AI draft: always, never, guaranteed, refund, won't have to, next year, allowance, threshold. None is wrong in itself, but each is a place where the model tends to add something you didn't give it. For the wider pattern of confident errors, AI hallucinations explained for business owners covers why this happens.

Formats that land better than a long email

• The three-number summary. "This year: $21,400. Last year: $14,900. Due by: [date]." Then two sentences of why. Works for clients who skim.
• A one-page letter with a "what you need to do" box. For bigger news, such as a first large bill. The box holds the action and the date, nothing else.
• A 90-second call script. Ask the AI to turn the approved email into talking points for a phone call. Some clients will only really take it in when they hear it.
• A small table for instalments. Date, amount, what it's for. Easier to stick on a fridge than a paragraph.

The call script is the format most worth checking, because it's the one furthest from the approved wording. Asked to turn the email to the financial planner into talking points, an assistant might return, illustratively:

1. "Don't worry, this is completely normal for a growing business."
2. "Your bill is $21,400, up from $14,900."
3. "Profit went up from $71,000 to $92,000, and tax follows profit."
4. "Last year an $8,000 loss from the year before reduced the bill; that's used up now."
5. "Nothing about your rates has changed."
6. "Payment is due by [date]. Shall we set up a payment plan?"

Two fixes. Point 1 is reassurance nobody approved, and "completely normal" is a claim about other people's tax, so it goes; the call should open with point 2, the number the client rang about. And the script ends with a yes-or-no question on a payment plan, which invites an answer before the client has thought about cash flow. Better to close with "have a think about whether a plan would help, and let me know by [date]", then note the answer on the file.

A filled-in instalment table for the same client might have three rows: the $21,400 balance with its due date, then the two advance payments towards next year with their amounts and dates, each labelled "towards next year's bill" in plain words. Build it from the computation, never from the AI draft, and paste it in after the prose has been approved, so no figure passes through the model twice.

Where plain English has to stop and advice begins

Explaining what happened is different from advising on what to do next. The moment a client asks "so should I pay myself differently?", the conversation has moved from explanation to advice, and that belongs in a meeting or a properly scoped piece of work, not an AI-drafted email. What AI can and can't safely prepare on business taxes draws the same line from the client's side.

Keep your approved explanations. After a season, you'll have a library of reviewed paragraphs for the common situations, and your house style will be visible in them. Put the best ones and your style rules somewhere the AI and the team both follow; writing a house style guide your team and AI both follow shows how. For the busiest weeks, how small tax practices use AI through the busy season fits these letters into the wider workload.

Plain-English tax explanations: follow-up questions

Should I tell clients that AI helped write the explanation?

There's no universal rule for a reviewed letter, but many firms mention AI use in their engagement terms or privacy notice, which covers it without a note on every email. What matters more is that a qualified person has checked every figure and statement, because the explanation goes out under your firm's name and you're responsible for it either way.

Can the same method help clients whose first language isn't English?

Yes, with care. Write and approve the English version first, then ask the AI to translate it and to keep all figures and dates unchanged. If nobody in the firm reads the target language, keep the translation short, send the English alongside it, and say which version governs. Don't translate advice or anything contractual without a qualified translator.

Is it safe to paste a client's tax figures into a chat assistant?

Use a business plan that doesn't train on your content by default, such as ChatGPT Business or Claude Team, and paste only what the explanation needs. The facts block in this method uses a first name and figures, not tax reference numbers, addresses or dates of birth, which keeps the exposure small if anything goes wrong.

Further reads

• Turning Hours Saved by AI Into Advisory Work Clients Pay For — Turn the hours saved on letters into advisory work clients pay for.
• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — The AI mistakes that damage client trust fastest.
• Is It Safe for an Accountant to Use ChatGPT With Client Data? — Settings and plans for using client figures safely.
• How to Anonymise Client Data Before You Paste It Into AI — A quick routine for stripping identifiers before you paste.
• ChatGPT Prompts for Bookkeepers: Client Queries, Chasers, Notes — More prompts for client queries, chasers and notes.
• AI Engagement Letters for Accountants: Faster Drafts, Clear Scope — Which parts of an engagement letter AI should draft, five prompts with sample outputs, and the partner checks that keep scope clear and terms untouched.
• How Small Accounting Firms Use AI Day to Day: Real Examples — A five-person practice's working week with AI, day by day: the tools, the prompts, what came back, and what the team corrected.
• Explaining Dental Treatment Plans to Patients With AI — A dentist-checked workflow for turning clinical treatment plans into plain-English explanations patients can reread and compare at home.
• Explaining Lens Options and Eye Test Results With AI Help — Prompts and checked templates for explaining prescriptions, lens choices and eye test findings to patients in plain English, with the clinician in charge.
• How to Choose an Accountant Who Uses AI Well — What good AI use inside an accountancy practice looks like, eight questions for the first meeting, and how a café compared three firms.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Support, Get your document's readability and level statistics; OpenAI and Anthropic business plan data-use pages.