Tax Services AI for a Small Practice: Intake, Chasing and Review Notes

A small tax practice can use AI in three places: turning last year's return and a short questionnaire into a client-specific intake checklist, drafting reminder emails for missing documents, and summarising what a preparer did into a note for the reviewer. In each, a named person owns the client data and signs off.

The order matters. Most practices reach for the return itself first, because that feels like the real work, but the cheapest hours to win back sit earlier: the weeks between sending an organiser and having a file complete enough to start. That is where staff re-type the same chasing email forty times, and where AI is least risky because nothing it writes reaches a filing.

Follow me on Instagram@sagnikteaches

Why intake is the right first job for a practice of two to ten people

Intake work has three properties that suit AI. It is repetitive across clients, the output is a message or a list a human reads before it goes anywhere, and mistakes are visible and cheap: a wrong line on a checklist gets noticed when the client says "I don't have that". Compare that with a return, where a wrong figure can sit unnoticed until a notice arrives.

Connect on LinkedInSagnik Bhattacharya

It also targets the real bottleneck. In a typical small practice the preparer is not slow at preparing; the file is slow to become ready. Partners often say they are "waiting on clients", and what they mean is waiting on the three or four items that make the file startable: a prior-year return, a statement of income, a record of payments made. Chasing those well is a communications problem, and drafting is the thing language models are good at.

Subscribe on YouTube@sagnikteaches

What stays with people, in every stage below:

• Client data. A person decides which tool sees which client's information, and records that decision.
• Completeness. AI reports what is on the checklist and what has arrived. It does not decide that a file is complete.
• The return. The preparer is responsible for what is prepared, and the reviewer for what is reviewed. Nothing in this workflow moves that responsibility.

Set the data rules before you open a chatbot

Write a half-page rule and pin it where the team works. It needs four lines: which tools are approved for client information; what may be pasted (descriptions of documents, document types, years, redacted extracts); what may never be pasted (full returns, identity numbers, bank account details, login details); and who to ask when unsure. My starting position for a practice with no policy yet is the cautious one: use AI on descriptions of the file rather than the file.

That sounds restrictive, but it covers most of the value. The intake checklist needs "two earners, two employers last year, a rental property, a new baby" rather than anyone's income figures. The chasing email needs a list of missing items, not their contents. The reviewer note needs the preparer's decisions, which you can write in words. Where your practice software has its own AI features and a data agreement with the vendor, that is a better home for anything touching real documents than a general chatbot. If you are unsure what clients have consented to, read the tutorial on consent for AI tools before going further, and ask your professional body or an adviser where the rules are unclear.

A quick redaction habit

Before pasting anything, replace names with "Client A", strip identity numbers and account numbers, and round amounts or leave them out. The redaction walk-through for tax preparers covers the mechanics. Ten seconds of redaction is cheaper than explaining to a client why their details sat in a chat history.

Stage one: build the intake checklist from last year, not a template

Generic organisers ask for everything, and clients respond by sending everything or nothing. A checklist built from last year's return asks for what this client probably has. Some practice platforms now do part of this themselves: TaxDome's help centre describes AI checklist generation from a prior-year return, for returns prepared in four named desktop packages, and says it is included in the subscription for eligible firms. If your software does that, use it and skip to stage two. If not, do it by hand with a chatbot in about ten minutes per client.

The method: the preparer writes a short profile from last year's file, in words, with no identifying data, then adds this year's known changes from the engagement call. The AI turns the profile into a checklist grouped the way the preparer likes to work.

You are helping a small tax practice prepare a document checklist.
Below is a profile of a returning client (no names, no numbers).
Produce a checklist grouped as: Income, Deductions and credits,
Property and assets, Life changes, Prior-year items to confirm.
For each line give the document name and one plain-English sentence
saying what to look for on it. Flag any line that depends on an
assumption I made. Do not add items the profile does not support.

Profile:
- Married filing jointly, two earners
- One earner changed employer mid-year (two wage statements likely)
- One rental flat, let all year, mortgage interest
- Child started university in September
- Spouse did occasional freelance work, no accounting records last year
- This year: bought a car for the freelance work

A realistic reply (illustrative) gives you around twenty lines: two wage statements for the earner who moved, the mortgage interest statement and rent records for the flat, the university's tuition statement, freelance income summaries and a note that "no accounting records last year" is a risk, and a prompt for the car purchase date, price and business-use mileage. The line worth reading closely is the flagged assumption. Here the AI will usually flag that it has assumed the freelance work continued; the preparer can confirm on the call.

What to fix in the output: it will tend to add items "just in case" despite the instruction, such as retirement account statements nobody mentioned. Delete them. A checklist with twenty relevant lines gets answered; one with thirty-five gets abandoned. Keep a master list of the lines you commonly strip, and add "do not include X" to the prompt until the habit is stable.

Stage two: the intake call notes, turned into a client-ready list

Intake calls produce notes that are messy in the way spoken notes are. The second use of AI is turning those notes into two things: an internal file note, and a short list for the client in plain words. Use your own typed notes, not a call recording, until you have settled consent for recording and for sending it to a tool.

Say a tutoring centre owner who also files as a sole trader calls the practice. The preparer types: "Owner, takes fees by card and cash, pays two part-time tutors, rents rooms from a church hall, bought laptops in March, thinks tutors may be contractors not staff, didn't keep a mileage log." Fed to the chatbot with an instruction to produce a client list and an internal query list, the output splits cleanly: the client list asks for card processor annual summaries, a cash takings record, tutor payment records and the hall rent agreement; the internal list queries tutor status (employee or contractor), cash record quality, and the capital purchase treatment, none of which AI should answer. That split is the point: the AI sorts requests from questions the preparer must resolve.

Stage three: chasing missing documents without nagging

Chasing is where small practices lose the most time and goodwill. The pattern that works is a sequence of short messages, each naming the specific items still missing, each easier to answer than the last, ending with a call from a person. AI drafts the messages; a tracker decides who gets one.

Keep the tracker boring: a spreadsheet or your practice software's task list with columns for client, items outstanding, date requested, number of nudges sent, date of last contact, and the person responsible. If your software already tracks uploads against a checklist, use that as the source and let AI write only the words. If it does not, the spreadsheet is enough. For a practice with sixty returns in progress, updating it takes ten minutes each morning.

A three-message sequence you can copy

Give the AI the list of outstanding items per client and ask it to draft all three messages in your practice's tone. A filled-in example for a yoga studio owner (illustrative):

Message 1 (day 7, friendly):
Hi [first name], thanks for sending most of your paperwork. To start
your return we still need three things: the payment processor's annual
summary, the studio rent statement for the full year, and the
instructor payment records. A photo from your phone is fine for each.
Reply to this email or upload to your portal.

Message 2 (day 14, specific):
Hi [first name], a quick follow-up. We've held a slot for your return
from the week of [date], but we can only start once we have the
processor summary. It is a one-page download from the provider's
dashboard under Reports. The other two can follow.

Message 3 (day 21, offers a call):
Hi [first name], we haven't been able to start yet because the
processor summary is still outstanding. If it's easier, call the office
and [staff name] will talk you through where to find it in about five
minutes.

Notice the structure rather than the wording. Message 1 lists everything and makes the effort small. Message 2 drops to the single blocking item and says how to get it. Message 3 shifts to a phone offer. This is also how the sequence shortens as the client sends things: the AI is re-run on the remaining list, not on a template with stale items.

What to fix in drafts: AI will often write a stern third message ("failure to respond may result in penalties"). Delete threats; a client who feels scolded finds another practice. It will also invent deadlines. Only the practice knows its real cut-off dates, so put them in the prompt as facts, and read every message for a date you did not supply.

Handling the client who sends the wrong thing

The commonest failure is not silence, it is the wrong document: last year's statement, page two of four, a screenshot cropped at the top. Give the AI a description of what arrived and what was expected and ask for a one-paragraph reply that says what was received, what is wrong in plain words, and exactly how to get the right version. A preparer who writes "this is for the previous year, we need the current one, it's under Documents then Annual" in 15 seconds can have a chatbot expand it into a polite message in five. Do not let the AI see the document itself unless the data rules above allow it.

Stage four: reviewer notes that a reviewer will actually read

A review note exists so a second person can see what the preparer decided and why, without redoing the work. Preparers hate writing them, so they are short and uneven. The use here is narrow: the preparer dictates or types rough bullets about what they did and the AI organises them into a standard layout. The AI adds no content of its own.

Reformat my rough notes into a reviewer note with these headings:
Documents received, Judgement calls, Items not resolved, Questions for
reviewer. Use only what is in my notes. If a heading has nothing under
it, write "None noted". Do not add advice or figures.

My notes:
- all income docs in, employer 2 wage statement was a reissue, used it
- rental: treated roof repair as repair not improvement, invoice says
  "replace damaged tiles" only, no photos
- no mileage log for freelance, used client's estimate of 3,000 miles
- not sure the tuition credit applies, child may be part time

The output keeps the same four judgement items and pulls them under the correct headings: the reissued wage statement under Documents received, the repair-versus-improvement call under Judgement calls, the mileage estimate under Items not resolved with a note that no log exists, and the tuition question under Questions for reviewer. A reviewer scanning that in two minutes knows where to look. Without it the same reviewer reads the whole file to find the roof invoice.

The key check is that nothing was added. Models are tempted to tidy "not sure" into a confident sentence ("the tuition credit applies, as the child is a student"). Compare the note to the original bullets and strike any sentence with no source. Do this for the first twenty notes; after that, spot-check one in five.

A worked example: one season in a six-person practice

An illustration with invented figures, to show where the hours move. Picture a six-person practice with 220 returns, two preparers, one reviewer, an administrator and two partners. Before the change, the administrator spends about 12 hours a week in the busiest six weeks chasing documents by hand, partners spend about 15 minutes per return writing checklists on calls, and review notes are written for roughly a third of files.

Do the sum. Checklists: 220 clients times 10 minutes saved is about 37 hours. Chasing: 7 hours saved for six weeks is 42 hours. Reviewer notes cost 3 minutes times the roughly 150 files that previously had none, about 7.5 hours added, which is deliberate, because reviewers benefit. Net around 70 hours across the season, minus a week or so of setup and the hour-a-week of reading drafts. These are numbers to test on your own season, not a forecast: if your pilot shows half of this, weigh it against the setup week, and if it shows nothing, you have learnt that cheaply.

How to run a four-week pilot and know it worked

1. Week 1: write the data rules, pick ten returning clients, build their checklists from profiles, and time each one.
2. Week 2: send the checklists, set up the tracker, and draft the first chasing messages for anyone with items outstanding after seven days.
3. Week 3: run the three-message sequence on the stragglers, with the third message always followed by a call from a person. Count replies after each message.
4. Week 4: use the reviewer-note layout on every file that reaches review. Ask the reviewer one question: did the note save you looking anywhere?

Measures worth keeping: median days from organiser sent to file startable; checklist lines the client said did not apply (this is your error rate); chasing messages sent per client; and reviewer yes or no on the notes. If days-to-startable has not moved, the problem is probably not wording, and a software reminder or a phone call at day five will beat a better email.

Where this goes wrong in a small practice

• The checklist silently omits a line. A model working from a profile cannot list what you forgot to mention. The preparer reads last year's return against the checklist once, for the first season.
• Tone drift. Over a long chat, messages get longer and warmer or colder than your house style. Start a new chat per client and paste your style rules in each time, or save them in a Project or custom instruction on the plan you use.
• Chasing someone who already sent it. The tracker was stale. Always check the portal or inbox before a message goes out. Nothing annoys a client faster.
• Treating a client's silence as a refusal. Some clients want a call, some are on holiday, some are ill. The third message offers a person, not a deadline.
• Data in the wrong place. One staff member pastes a full return into a free chatbot "just this once". Your rule should say why, not only that, and the response to a breach of the rule should be a conversation, not silence. The tutorial on free AI tiers and client data explains what the difference actually is.

What your practice software may already do

Before you build any of this by hand, read your own platform's AI page. TaxDome's help centre, for example, lists AI checklist generation and checklist matching as part of the subscription for eligible firms, and document renaming, tagging and return-delivery help as part of a separate AI add-on ($200 a month per firm on its help page in November 2026); features vary by region and change often, so confirm against the live page. If your platform matches uploads to checklist lines, you can drop the manual tracker for those clients and keep the AI-drafted messages. The comparison of tax preparation software with AI features goes through what the main platforms offer, and the tutorial on chasing client documents goes deeper on the reminder schedule.

The test for any software feature is the same as for a chatbot: who sees the data, where does it go, can you see what the AI did, and can a person override it? If the vendor's answer to any of these is vague, treat the feature as unready for client files.

Checks that keep responsibility with a person

Four lines to put in your procedures manual, in your own words:

1. Every AI-drafted client message is read by a staff member before sending.
2. Every checklist is confirmed against the prior-year file by the preparer in the first season, then spot-checked.
3. "File ready" is a recorded decision by a named preparer, never an inference from the tracker.
4. Reviewer notes are drafts of the preparer's own decisions; the preparer initials them.

If you want to see how the next stage works once documents are in, the tutorial on extracting, reviewing and signing off a return picks up where this one ends.

Questions small tax practices ask before switching this on

Should client documents go into a general chatbot?

Only with a plan you have checked: whether the provider trains on inputs, how long it keeps them, and what your engagement letter and consent forms allow. Safer habit: paste descriptions and redacted extracts, never full returns or identity numbers. Practice software with built-in AI and a data agreement is usually the cleaner route.

How many chasing emails is too many?

Three automated nudges over about three weeks is a sensible ceiling, with a phone call from a person after the third. More than that stops reading as helpful and starts reading as noise, and the clients who ignore three emails rarely answer a fourth.

Can AI decide that a return is ready for review?

No. It can tell you which checklist lines have a matching upload. Whether the documents are complete, correct and the right year is the preparer's judgement, and the readiness decision should be recorded by a named person.

What does this cost to try?

A pilot can run on a paid chatbot plan and a spreadsheet for a few weeks. Practice-management software with AI features changes its pricing often, so read the vendor's current page and ask what is included in the base subscription and what is an add-on.

Further reads

• AI Tax Organiser: Build One That Chases Missing Documents — Build the organiser itself, with chasing built in.
• ChatGPT Prompts for Tax Preparers: Client Emails and Checklists — More prompts for client emails and checklists.
• AI Software for Tax Professionals: Run a Pilot and Measure Cost per Return — Run a pilot and measure cost per return.
• AI Tax Return Review vs Software Diagnostics: What Each One Catches — What AI review catches compared with software checks.
• AI for Tax Advisory: Drafting Client Advice Memos Safely — Drafting advice memos safely.
• Tax Workpaper Prep With AI: Tie the Return Back to Source Documents — The next stage: tie the return back to sources.
• How Small Tax Practices Use AI Through the Busy Season — How intake, chasing and review fit across a whole season.
• MTD for Income Tax: Triage Your Client List and Send Reminders With AI — A practice-side list, exception queue and three reminder drafts for clients approaching digital record duties.
• Making Tax Digital With AI: Chase Quarterly Records From Clients — A small-practice quarterly chase plan with evidence requests, a response tracker and human approval gates.
• TaxDome Automations: What to Automate First in a Small Practice — Which TaxDome automations to set up first in a small practice: onboarding, document requests and reminders, in order, with what to leave manual.
• Client Tax Organiser: Build One With AI and Send It by Portal — A conditional tax organiser example with client questions, portal instructions, staged reminders and a completeness review.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: TaxDome Help Center page on TaxDome AI (checked 4 November 2026); AI provider data-use and retention pages; the author's own workflow design. Examples are illustrative.