Set it up before the season starts: AI-drafted document requests built from each client's prior-year return, chasers, intake checks that flag missing items, triage and draft replies for the inbox, and plain-English summaries of each result. Keep tax positions, calculations and anything filed with the tax authority under preparer review.
Timing is the part most practices get wrong. A new tool introduced in the second week of peak season gets abandoned by the third, because nobody has time to fix its mistakes. The work that pays off in the busy months is done six to eight weeks before them, when there's time to test prompts on last year's files.
Eight weeks out: build the kit
Four pieces of preparation carry the season. Each is a one-off job that saves time on every client afterwards.
- A prior-year summary per client. Half a page: income sources, what documents came in last year, what was late or missing, questions to ask this year, and anything that changed.
- Tailored document checklists, generated from those summaries rather than one generic list for everyone.
- A chaser sequence: three messages of increasing firmness, with the dates they go out.
- Reply templates for the ten questions clients ask most often in peak weeks.
Test every piece on ten of last year's files before the season starts. If the prior-year summary for a client you know well misses something obvious, the prompt needs work, and you'll never have a better time to fix it.
If you take on seasonal admin help, add a fifth piece: a one-page guide for them covering which prompts and templates to use, which tool and account to use them in, the rule that no figure goes to a client without a preparer, and who to ask when an AI output looks wrong. Temporary staff are the people most likely to reach for a free chatbot on their own phone when the practice tool feels slow, so show them the approved tool on their first morning and watch them use it once. AI can draft the guide from your notes in minutes; a preparer should still read it as if they were new. A filled-in extract from an illustrative guide:
SEASONAL HELP - HOW WE USE AI Updated: [date]
Tool: [practice AI tool], signed in with your practice account. Never a
personal account or an app on your own phone.
You may: draft document requests and chasers from the saved prompts; sort
the inbox into Today / This week / Filing; fill reply templates.
You may not: put any amount owed or refundable in a message; paste ID
documents or bank details; send anything to a client a preparer
hasn't seen, except the three approved chasers.
If the AI output looks odd, or a client asks about money: stop, message
[preparer] and leave the draft unsent.The "may not" list is the part to read aloud on the first morning. Everything else they'll pick up by using it.
A filled-in prior-year summary for an illustrative client, a food truck owner trading as a sole trader, drafted by AI from last year's file notes and corrected by the preparer:
PRIOR-YEAR SUMMARY - [client ref] Preparer: [initials]
Income: food truck trading (card payments via one provider, some cash);
occasional private event catering, invoiced.
Last year's documents: card provider annual statement; bank statements
(business account); fuel and ingredient receipts (photos, late);
vehicle finance statement; pitch fees.
Problems last year: receipts arrived in March in one batch; cash takings
not recorded weekly; vehicle finance statement missing until chased twice.
Ask this year: any new vehicle or equipment? any events invoiced through
a platform? has cash recording improved?
Changes noted: client mentioned a second trailer in the autumn.The corrections the preparer made to the AI's first draft are typical. It listed "home office costs" because last year's notes mentioned the client did paperwork at home; the preparer knew it had been discussed and rejected, and deleted it. It didn't spot that the second trailer mentioned in a November email should be a question, because that email wasn't in the file notes it was given. Feed it the right sources, and read what it produces.
Document requests that match each client
Generic document lists are why clients send the wrong things. A request built from their own prior year gets better results, because it names what they actually sent last time. Some practice management systems now do this themselves: TaxDome says its AI builds a document checklist from each client's prior-year return and renames and tags uploads as they arrive. If you use a system like that, set it up and test it now. If not, a prompt does the job in a business-plan AI tool:
Using the prior-year summary below, write a document request for this
client's return. Group items under headings (Income, Expenses, Assets,
Anything new). Name documents the way the client would recognise them
("your card payment provider's annual statement", not "merchant income
summary"). Add a line for each "Ask this year" question.
Keep it under 250 words, friendly and plain. Do not mention tax amounts
or deadlines; we will add the deadline ourselves.For an illustrative e-commerce homeware seller who trades through two marketplaces and their own website, the output's income section might read: "Payout reports from each marketplace for the whole year (download these from your seller dashboard); your website payment provider's annual statement; any refunds you issued outside those platforms." That line alone prevents the most common delay for online sellers: sending one marketplace's figures and forgetting the other.
The weakness of a request built from last year is that it can only ask about last year. A client who started letting a spare room in the summer, or sold shares for the first time, gets a list that says nothing about either, and they may assume that means it doesn't matter. The "Anything new" heading is the guard, but only if it asks specific questions rather than "anything else?". An illustrative version that works: "Since last year, have you started any new work or side income, let or sold any property, sold any investments, or had a big change such as marriage, a new baby or retirement? A yes to any of these means we'll need a few extra documents." Clients answer a list of concrete events far more reliably than an open question.
Intake: spot what's missing before a preparer opens the file
When documents arrive, the most expensive moment is a preparer opening a file, starting work and discovering something's missing. AI can do the comparison at intake. Give it the tailored checklist and the list of documents received (file names and a line on each), and ask for three lists: received and complete, received but possibly incomplete (a bank statement covering eleven months), and not received. Then have it draft the chaser.
An illustrative intake check for the food truck owner above:
RECEIVED AND COMPLETE: card provider annual statement; pitch fee invoices
(4); ingredient receipts (photos, 212 files).
POSSIBLY INCOMPLETE: "business bank statements full year.pdf" - file name
says full year; contents not checked.
NOT RECEIVED: vehicle finance statement; details of the second trailer;
private event invoices.The middle line is the honest one, and it's there because the prompt asked the AI to judge contents rather than file names. When the administrator opened the PDF it held eleven monthly statements: March was missing, which is where the "after" chaser below comes from. An earlier version of the prompt had marked the same file complete on its name alone. If your intake check only ever reads file names, test it by uploading a deliberately short statement and seeing whether it notices.
The difference a specific chaser makes:
Before: "Dear client, we are still waiting for some documents. Please send
them as soon as possible."
After: "Thanks for sending your documents last week. We have everything
except two items: your vehicle finance statement for the year, and
the bank statement for March, which was missing from the batch.
Once we have those, your return goes straight to review. You can
upload them to the portal or reply to this email."The specific version gets documents back days sooner, because the client knows exactly what to look for and that it's the last thing standing between them and a finished return. The full chasing method, including deadline-driven sequences, is in how to chase missing client records with AI.
The inbox in peak weeks
In the busiest weeks, client email can take more preparer time than preparation. Three things help:
- Triage that flags the urgent few. Notices from the tax authority, "I've just realised I forgot...", and replies to your own queries need a preparer the same day. The setup for a professional firm, including Copilot's Prioritize my inbox and practice management triage such as Karbon's, is covered in AI email triage for professional firms.
- Reply templates for the repeat questions. "Have you got everything?", "When will it be done?", "Can you confirm my appointment?", "What's my login for the portal?". AI personalises the template from the file status; a person sends. Filled from a file marked "documents complete, in queue for preparation", the illustrative reply to "have you got everything?" reads: "Yes, we have everything we need, thank you. Your return is in the queue and a preparer will be in touch if anything needs explaining." It says nothing about when, because the file status doesn't say.
- A second look at attachments. In an illustrative peak week, triage filed a client email headed "quick question about my photos" under "this week". The attachment was a photograph of a letter from the tax authority with a reply date inside it. Triage that reads the email's words but not what's in the attachment will miss this, so any client email with an image or PDF attached goes to a person the same day, whatever the subject line says.
- One rule for money questions. "How much will I owe?" never gets an AI-drafted figure. The template says the preparer will confirm once the return is reviewed, and when.
Karbon, for example, says its AI summarises threads and drafts replies in the firm's voice and is included for its customers at no extra cost, so if you already use a practice management system, check what it offers before adding anything.
Review support without letting AI do the tax
The preparer's review is the part no tool should shortcut, but AI can make it faster by turning numbers into questions. Calculate the year-on-year comparison in your tax software or a spreadsheet first, by line. Then give AI only the lines that moved more than a threshold (20% is a common starting point) and ask for client questions. Add a money floor as well, or the percentage rule misfires both ways: a line that goes from $40 to $90 is up 125% and not worth a question, while a $30,000 line up 15% has moved $4,500 and is. A rule such as "more than 20% or more than $2,000, ignoring lines under $500" suits many small-business clients; adjust it after the first twenty files.
Below are the lines in this client's figures that changed by more than 20%
from last year, with both years' amounts. For each, write one short,
neutral question we could ask the client to explain the change. Do not
suggest tax treatments or reasons.An illustrative output for a furniture maker:
1. Materials: up 46%. "Your timber and materials costs rose a lot this
year. Was that more orders, higher prices, or stock you're still holding?"
2. Vehicle costs: down 38%. "Vehicle costs fell compared with last year.
Did you change vans, or use the van less for deliveries?"
3. Subcontractors: new this year. "We can see payments to other makers or
finishers this year. Could you send a list of who you paid and for what?"Before sending: question 1 is good, and the stock point matters to the return. Question 3 assumes who the payees were; reword it to "payments to other businesses or individuals" until the client confirms. These questions go into one email per client, rather than three separate calls from a preparer mid-review.
The no-figures rule exists because of errors like this illustrative one: a practice let its AI tool draft "your return is ready" emails. For one client the draft helpfully added "you should expect to owe around" a figure, estimated from a threshold it had assumed. The threshold was a year out of date, the figure wrong. It was caught at review only because the preparer read the whole email. The fix is in the prompt ("never state or estimate an amount") and in the review step: every number in a client email comes from the reviewed return.
Explaining the result in plain English
Once the return is reviewed, AI is good at turning it into a short explanation the client will read: what went in, what changed from last year and why, what they owe or are due, when, and what to do differently next year. Give it the reviewed figures and your notes, not the raw documents, and check the finished text against the return line by line. An illustrative draft for the furniture maker, with the preparer's edits noted:
Your return is finished and you're due a refund of [amount from return].
Your profit was lower than last year mainly because materials cost more,
and some of that timber is still in your workshop as stock.
[EDIT: "some of that timber is still in stock" - confirm the stock figure
the client gave us is the one we used.]
Your payments during the year were higher than the final bill, which is
why you're getting money back. Next year's payments will be lower too.
[DELETE: "next year's payments will be lower" - not something we've
worked out or agreed.]
Next year: please send the list of other makers you paid by [month].The pattern of edits is typical. Explanations of what happened are usually accurate, because they come from the figures you supplied. Predictions about next year creep in because they sound helpful, and they're the sentences a client quotes back twelve months later. The method, with examples, is in how accountants use AI to explain tax in plain English. Clients who understand their result ask fewer questions and send better records next year.
Rules for tax data in AI tools
- Business plans only. ChatGPT Business, Claude Team, Microsoft 365 Copilot and Gemini in Google Workspace leave your content out of model training by default. Free consumer tools are not for client returns.
- Check the rules that apply to preparers. Some tax regimes put specific limits on how preparers use or disclose return information, including to software and service providers, and some require client consent for certain uses. Confirm what applies to you with your professional body or adviser before return data goes into any AI tool.
- Keep identity documents and bank details out. No task in this plan needs them in an AI chat.
- Say it in your engagement letter. A sentence on the tools you use and how client data is handled. The wording is covered in AI engagement letters for accountants.
What it adds up to for a three-preparer practice
For an illustrative practice with three preparers and an administrator handling around 600 returns, the peak-season savings look roughly like this:
| Task | Before (per return) | After (per return) | Saved across 600 returns |
|---|---|---|---|
| Document requests and chasers | 20 min | 8 min | about 120 hours |
| Intake check for missing items | 10 min | 4 min | about 60 hours |
| Client email in peak weeks | 25 min | 15 min | about 100 hours |
| Result explanation to client | 15 min | 6 min | about 90 hours |
| Preparation and review | Unchanged: this is where the preparer's time should go | ||
Around 370 hours over a season is roughly two working months of one person's time, moved from admin into review and advice. Your figures will differ; time a sample of returns this season so you have your own baseline. The broader question of what AI can safely prepare in tax work is covered in can AI do my business taxes.
The week after the deadline
Hold a 30-minute debrief while memories are fresh. Three questions: which prompts or templates did people stop using, and why? Which AI outputs had to be corrected most often? Which clients were late despite tailored requests? Update the prior-year summaries with this year's lessons the same week. That's the first step of next season's eight-week build, done while it's easy.
Write the answers down as changes, not impressions. Illustrative notes from a debrief:
- Dropped: the reply template for "when will it be done?", because preparers found it quicker to type one line. Delete it rather than keep a template nobody uses.
- Corrected most: intake checks on marketplace sellers, which kept marking one platform's payout report as the whole year. Add "list each platform separately" to the prompt.
- Late despite tailored requests: nine clients, seven of them the same people as last year. Move their first request two weeks earlier and send it by text as well as email.
Three concrete changes like these, made in the quiet month, do more for next season than any new tool.
Questions tax preparers ask about AI in peak season
Should we use AI for tax research questions?
Use it to find your way into a topic and to draft questions, not for the answer you rely on. Rates, thresholds and rules change every year and AI tools may give last year's figure or another jurisdiction's with complete confidence. Check every point against the tax authority's own guidance or your usual research service, and note the source on the file.
Can AI read receipts, slips and statements for us?
Extraction tools and many practice management systems now read uploaded documents and pull out names, dates and amounts, and general assistants can read clear scans too. Accuracy drops with photos of crumpled receipts and multi-page statements. Use extraction to save typing, then reconcile totals against the source before anything goes into a return.
Is peak season too late to start?
For new tools, usually yes. Adding reply templates or a chaser prompt mid-season is fine, because they're small and easy to drop. Anything that changes how documents arrive or how work is tracked is better started the month after the deadline, when you can test it on the files you've just finished.
Further reads
- ChatGPT Prompts for Bookkeepers: Client Queries, Chasers, Notes — Tested prompts for client queries, chasers and notes.
- AI Receipt Capture and Bank Categorisation: A Bookkeeper's Setup — Setting up receipt capture so records arrive cleaner.
- Best AI Tools for Small Accounting Firms, Sorted by Task (2026) — Tools sorted by task for small practices.
- Is It Safe for an Accountant to Use ChatGPT With Client Data? — The data-safety question for client financial records.
- How Much Does AI Cost a Small Accounting Firm? — What the tools in this plan cost a small firm.
- 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — The AI mistakes most likely to cost you clients.
- 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.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: TaxDome AI product page; Karbon AI feature page; Microsoft Support, Prioritize my inbox; OpenAI and Anthropic business plan data terms.