Year-End Tax Planning Letters With AI: Personalise Without Errors

Personalise year-end tax planning letters by merging each client's figures from your own records into a fixed template, and use AI only to write the sentences around them. Then check every figure against the source before sending. The fastest letters come from keeping facts out of the model's hands.

The risk is plain. A letter that tells a client "you paid $18,400 in estimated tax" when the figure is $14,800 damages trust more than a generic letter ever would. The method below splits the work into data, wording and review, so AI saves time where mistakes are cheap and never touches the numbers.

Follow me on Instagram@sagnikteaches

Merge fields or free-text AI: choose by what can go wrong

There are two ways to personalise at volume. Most practices end up combining them.

Connect on LinkedInSagnik Bhattacharya

The rule that keeps letters safe: numbers, dates and named obligations come from merge fields; AI writes connective tissue and explanation, and only from facts you hand it. If a model writes a figure itself, something is wrong with the process.

Subscribe on YouTube@sagnikteaches

Step 1: build the client data row

One row per client, built from your records, not from memory or last year's letter. Typical columns for a year-end letter, illustrative:

The "planning options relevant" column is the adviser's judgement, not AI's. Which options apply to which client is the value your practice provides, and a model that suggests a strategy the client does not qualify for creates risk. Put the choice in the data and let the letter report it.

Date the row. A letter drafted from numbers as of 31 October and sent after a November bank import can contradict the client's own bank app. Record the "as at" date in a field and print it in the letter.

Step 2: write the template, with decision paragraphs

Write the template once, in your voice, with a small number of alternative paragraphs picked by client type or option. Example, illustrative:

Subject: Your year-end tax planning, [first name]

Dear [first name],

Here is where your tax position stands at [as_at_date], and the decisions worth making before [year_end_date].

Where you are now
Your profit for the year so far is [ytd_profit]. You have paid [tax_paid] towards your tax. Based on our projection, your liability for the year is about [projected_liability], which leaves around [shortfall] to find by [key_date].

{IF option_equipment} Equipment purchase
If you plan to buy equipment for the business, the timing of the purchase can affect which year it counts in. Please speak to us before ordering, so we can confirm the treatment for your purchase.{/IF}

{IF open_items} What we still need from you
[open_items_list]
We cannot finalise the projection without these.{/IF}

These figures are estimates based on the information we hold and may change. This letter is general planning guidance for your circumstances, not a guarantee of any tax result.

Kind regards,
[signatory]

Notice what is in the template: only variables, conditional blocks and fixed, partner-approved sentences about options. The compliance sentence at the end is fixed text, never rewritten by AI.

Step 3: where AI actually helps

AI earns its place in four jobs, none of which involve calculating anything.

Writing the paragraphs in the first place

Ask for several plain-English versions of each fixed paragraph, then a partner picks one. This takes ten minutes and gives the practice a letter library. Prompt, with a sample reply:

Write three versions of a short paragraph for a year-end planning letter to a small business owner, explaining that the timing of an equipment purchase can affect which tax year it counts in. Plain English, no jargon, 50 words each, no figures, no promise of a tax saving, and end by inviting them to talk to us before buying.

Illustrative output: "If you are thinking of buying equipment, when you buy can change which year it counts in. We will confirm how your purchase is treated, so please talk to us before you order." Short, accurate, no promises. Choose, adjust, save in the template.

Personalising the opening line

One sentence per client that shows you know them: "Your new workshop lease will change next year's picture, so it is worth planning for." Supply the fact from your notes, and ask the model only to phrase it. It must not add facts.

Write one friendly opening sentence for a year-end letter. Use only this fact: [client] opened a second site in July. Do not mention any figure or tax rule. Under 25 words.

Explaining a client-specific point

If a client has an unusual situation, for example a one-off sale of an asset, give the model your adviser's own conclusion and ask it to explain it in plain English at a reading age of about 12. It rewords your conclusion; it does not reach one. The plain-English tax explanation tutorial has prompts built for this.

Spotting what is missing

Ask the model to compare the merged letter with the client's data row and list any field that is blank, any sentence that does not trace to the row, and any promise or guarantee wording. It will not catch everything, so treat it as a second reader, not the checker.

Step 4: the fact-check table every letter passes through

This is the heart of the method. For each letter, the reviewer ticks every statement of fact against the source. A one-page checklist:

The reconciliation line catches the commonest slip. In the sample row, profit is $62,300, tax paid is $9,800, projected liability is $14,100. The shortfall is $14,100 minus $9,800, which is $4,300. If the letter says $5,300, the line fails even though every individual number appears in the row. The check also guards against numbers that were right when the letter was drafted and wrong when it went out.

A four-partner practice sends 120 letters in one week

Say a four-partner practice sends year-end planning letters to 120 business clients. They sort the clients into three groups: 70 straightforward sole traders, 35 companies and 15 with special situations (property sales, new partnerships, large asset purchases). The first two groups get merge-field letters with an AI-written opening line; the third group gets a human-written letter with AI help on explanations only.

1. Monday: a bookkeeper builds the data rows from the practice system and reconciles them (about 3 minutes per row, 6 hours for 120).
2. Tuesday: a partner approves the template and conditional paragraphs, with AI offering alternative wording (1 hour).
3. Tuesday afternoon: the merge produces 105 letters automatically; AI writes 105 opening lines from one-fact prompts, reviewed in a batch (1.5 hours).
4. Wednesday to Thursday: two reviewers each check about 50 letters at six minutes each (10 hours across two people). The 15 special-situation letters take a partner about 25 minutes each (6 hours).
5. Friday: send, log, and note which fields caused errors.

Total, illustratively, about 30 hours across the team, compared with the 50 or so such a practice would spend writing each letter from scratch, and the letters are more accurate because figures are checked rather than retyped. The saving does not come from a faster check; it comes from not writing 105 letters from scratch.

For busy-season context, see how small tax practices use AI through the busy season, which covers where this sits in the calendar.

Five mistakes, each seen in a different form

The transposed figure

A model, asked to "summarise this client's position", returns $41,200 for a figure that was $42,100. Nothing in the text looks wrong. Defence: figures are never produced by AI, and the checklist compares each to the row.

The option that does not apply

A letter tells a client they could make a pension contribution when their situation, known to the adviser, makes that unsuitable. Defence: options come from the adviser's field, not from the model's general knowledge.

The stale date

The key date in the template is last year's, because the template was copied. Defence: the date field is entered each season from the tax authority's own page and the reviewer opens that page during the check.

The guarantee in friendly clothing

"This will save you about $3,000." The AI added it to sound helpful. Defence: a banned-phrases list for reviewers (will save, guaranteed, definitely, you will not pay) and the fixed compliance sentence.

The mail-merge crossover

One client's name with another's figures, after a sort order changed in the spreadsheet. Defence: lock the row ID in the letter footer in small type for internal copies, and have the reviewer match name and ID first.

A free-text draft that went wrong, and the fix

To see why the split matters, imagine a practice that skips the merge fields and pastes a client summary into a chatbot: "Write a year-end planning letter for [client], a landlord with two rental properties, rental profit about $28,000, paid $5,000 so far." The model returns a warm, fluent letter containing this paragraph (illustrative):

With your rental profit of $28,000 and payments of $5,000, you could reduce your bill by around $2,500 by bringing forward repairs to your second property before the year end. We estimate your remaining liability at roughly $3,900.

Four problems, none obvious on a quick read. The "$2,500 saving" is invented. The suggestion to bring forward repairs may not be appropriate for this client, and the adviser never chose it. "$3,900" was calculated by nobody. And the letter speaks as "we" about a projection the practice has not made. The fix is not a better prompt; it is a different design. Put the figures and the adviser-approved option in the data row, use the template paragraph about repairs and timing without any figure of benefit, and let AI do only the opening sentence. The reviewed version reads:

If you are planning repairs to either property, the timing can affect which year the cost falls in, and whether it counts as a repair at all. Please speak to us before booking the work so we can confirm the treatment.

It is less exciting and far safer. A client who wants a number will ask, and the adviser can then give a worked figure with the facts to hand.

Adapting the letter by client type

One template with conditional blocks covers most of the list. Differences that matter:

• Sole traders and partners: emphasis on estimated payments, record gaps and equipment timing. Plain language; most are not tax specialists.
• Company owners: the position inside the company and the position personally are separate. Make sure the letter does not mix them. A "profit" figure should say which entity it belongs to.
• Landlords: repair or improvement timing, records of expenses, any change in rents or ownership.
• Employees with side income: a short letter focused on the second income and any missing records.
• Clients with a change of circumstances: a human-written letter. Marriage, a sale, a death in the family, retirement or a move all deserve a call, and AI should only tidy language afterwards.

Keep the number of variants small. Five template versions that people maintain beat fifteen that drift. Review the variants each year, with a partner reading every fixed paragraph again.

Handling client replies after the letter goes out

A good planning letter produces replies: questions, documents, decisions. Plan for them before sending. Include a clear action list and one contact route. AI can help here too, with replies that you review before sending.

A client replied to our year-end planning letter. Their message is below. List what they are asking, what they are telling us that changes the file, and any document they attached. Then draft a short reply that answers only what our file notes below support, and asks for anything missing. Do not give a tax figure; say a partner will confirm.

File notes: [notes]
Client message: [message]

The "changes the file" list is the useful part. A reply saying "we also sold a van in October" changes the projection and may need a fresh data row and a new letter. If nobody updates the row, the next letter repeats the mistake.

A second quick sum: what the split saves and costs

Take 100 clients and a reviewer who is paid $45 an hour in loaded cost, as a round number. Writing each letter from scratch at 25 minutes is about 42 hours, or $1,875. With merge fields and AI wording, set-up is roughly 8 hours in the first year (template, prompts, data row build), then 8 minutes per letter for drafting-and-review, which comes to about 13 hours, so 21 hours in total or $945. The second year drops the set-up and costs about 13 hours of review. The table is illustrative, and your numbers will differ, but one point holds: the saving is real only when the review step stays in. If you cut the review to save another six minutes a letter, you have removed the control that makes the saving safe.

Add the tool's subscription cost from the vendor's current pricing page; business-plan seats for ChatGPT or Claude are typically a modest monthly figure per user, small beside the labour.

Keeping a letter library that improves each year

After the season, hold a one-hour review. Which paragraphs prompted client questions? Which were never used? Which caused review failures? Move the best paragraphs into a library, retire the weak ones, and note the year they were last approved. Next season, start from the library and change only what the law, the dates or the practice's advice has changed. That also gives a partner a defined list of paragraphs to re-approve, which is quicker and safer than rereading every letter in full. Store old letters and their data rows, so that if a client asks in two years what you told them, you can show it.

Compliance wording and confidentiality

Two sets of rules sit behind every letter. The first is your professional body's standards for communications and planning advice, which may require specific disclaimers or engagement terms. The second is confidentiality: client data is confidential, and sending it to a third-party tool needs a lawful basis and may need client consent written into your engagement terms. Check both with your professional body or counsel before you start. The engagement letters tutorial shows where AI-use clauses go.

A practical set of habits for the data: use only business plans that do not train on your content; send the model only the fields it needs for the task (for the opening line, that is one fact and no name); keep the merged letters in your own systems; and keep a record of which tool drafted which part. Fixed sentences such as the disclaimer are partner-approved and not editable by the drafting tool.

The send-ready QA step

Treat this as a gate with a name on it. A letter is send-ready only when a second person, not the drafter, has completed the checklist and initialled the data row. Run these last three checks on the final PDF, not the draft:

1. Open the PDF and read the first and last paragraphs aloud. Merge errors hide in the joins.
2. Search the file for square brackets, curly braces and the words "TODO", "CONFIRM" and "[".
3. Check the recipient name and address against the engagement record.

Sample ten letters at random after sending and compare each against its data row again. If more than one has an error, pause the next batch and find the cause.

What to measure so you know it works

Keep a small log for the season: letters sent, errors caught at review (by type), errors found after sending (aim for none), review time per letter, and client replies that mention confusion. If errors caught at review cluster in one field, fix the data source. If client replies show confusion about a paragraph, rewrite that paragraph in the template once, and every letter improves. Over two seasons the library of approved paragraphs becomes the most valuable asset, because each one has been written, checked and tested on real clients.

If you would like help setting up the data row and the review gate for your own client mix, that is something my AI implementation consultation covers, using the tools your practice already has.

Year-end letter questions

Should AI write the figures in a client letter?

No. Figures come from your records through merge fields or a table you control, and the AI writes around them. A model asked to produce numbers can transpose, round or invent them. If a figure must be calculated, calculate it in your own software and paste the result into the data row.

Can I paste client tax data into ChatGPT or Claude?

Only under your practice's data policy and professional confidentiality duties. Use a business plan that does not train on your content, remove names and identifiers where the task allows, and keep the real data in your own systems. Check the consent your engagement terms give for third-party tools.

How many letters should one person check per hour?

As a planning figure, allow five to eight minutes for a careful check of a personalised letter against its data row. At that pace one reviewer clears seven to twelve letters an hour. Plan the calendar around that, not around how fast the drafts appear. Faster drafting only moves the queue to the checker.

Do I need to say AI helped write the letter?

That depends on your professional body's rules and your engagement terms, so check both. Whatever the rule, the signatory owns the content. Many practices describe AI use in their engagement terms and do not label individual letters.

Further reads

• Client Tax Organiser: Build One With AI and Send It by Portal — Get the year-end data from clients in a usable form.
• Can AI Draft a Suitability Report That Passes Compliance Review? — How compliance reviews AI-drafted client documents.
• Using AI in Financial Planning: Tax Return Data Into a Client Plan — Turning tax return data into a client plan.
• Rolling Out AI in a Bookkeeping Practice Without Losing Control — Rolling out AI in a practice without losing control.
• How to Automate Client Onboarding in a Small Accounting Practice — Capture the facts at onboarding so letters stay accurate.
• FP Alpha vs Holistiplan: Tax Report Tools for Planners Compared — Per-household against per-credit pricing, a feature table from both vendors' pages, and a test you can run on one synthetic return before paying.
• Holistiplan Pricing: Basic vs Premium Tax and Who Needs Which — Holistiplan Basic Tax vs Premium Tax priced by household tier from 30 to 750, with a Premium break-even, the billing rules and a 7-day trial checklist.
• AI in Bookkeeping and Tax: The Year-End Handoff to Your Preparer — A year-end handoff pack, a queries list, an adjustments log and a missing-document chaser, with AI drafting and both sides checking.
• Can AI Draft an IRS Notice Response? What to Check Before Sending — A practitioner's routine for AI-drafted IRS notice responses: triage by type and deadline, redact first, verify every code, amount and date, then sign.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI and Anthropic pages on business-plan data use (no training on business content by default); practice workflow guidance. No tax rates, limits or deadlines are stated here because they change by year and place of practice. All figures in the examples are invented for illustration. Confirm current rules from your tax authority's own pages before putting them in a letter.