Free ChatGPT, Claude and Gemini plans, plus the Copilot Chat included with eligible Microsoft 365 business licences, handle drafting, summarising and explaining work that uses anonymised or public text. Anything with identifiable client records, arithmetic or a filing consequence needs a paid plan with no-training terms, a human check, or both.
That is a rule of thumb, not a verdict, and it comes with two warnings. Free plans publish few numbers: most vendors say "limited usage" and leave you to find the cap. And an accountant's duty of confidentiality doesn't change because a tool is free. The sections below sort fourteen common practice jobs, set out what each free plan actually states about its limits, and give a client-data rule you can adopt. Bookkeeping capture tools, which read receipts and bank feeds, are a different category and are covered elsewhere.
Fourteen practice jobs, sorted: free, paid or human
The question for each job is what goes into the tool and what could go wrong in what comes out. Jobs that need no client identifiers suit free plans. Jobs that do need them belong on a paid plan or with a person. A third group, calculations, belongs in a spreadsheet whatever the plan.
| Job | Free plan? | What breaks first | The human check |
|---|---|---|---|
| Chaser letter for missing records | Yes, with placeholders for names | A consequence or deadline stated wrongly | Read it against the engagement terms before sending |
| Plain-English explanation of a concept | Yes | A local rule presented as universal | Check against current published guidance |
| First draft of engagement letter clauses | Yes, from your own template | A clause missing or invented | A partner reviews every clause |
| Formula help, using dummy data | Yes | Works on the example, fails on edge cases | Test on a dozen real-shaped rows |
| Explaining an error in a client's spreadsheet | Only an anonymised extract | Hidden rows and links that the extract loses | Re-perform the check in the sheet |
| Meeting summary and action list | Fathom Free; others hit caps quickly | Consent, storage, retention | Read the notes; delete audio under your policy |
| Summary of a long email thread | Yes, once anonymised | Wrong order of events | Check the chronology against the thread |
| Reading a client PDF such as a statement or contract | No, not with identifiable data on a consumer account | Data risk, page limits, scanned pages | Paid no-training plan, then human reconciliation |
| Variance commentary from anonymised monthly figures | Yes, with the numbers rounded and the name removed | Invented causes | Recompute every figure and strike unsupported reasons |
| Technical research on how an item is treated | A starting point only, with web search on | Wrong or outdated citations | Open the source and confirm it is current |
| Calculations such as payroll or depreciation | No: use a spreadsheet | Arithmetic slips presented confidently | Recalculate and keep the working |
| Categorising bank transactions at scale | Not a free-chat job | Volume and data protection | Use a capture tool and review exceptions |
| Client portal chat answering questions | No | Liability for wrong answers | Paid tool with review rules, if at all |
| Working-paper or file review | No | Confidentiality and an audit trail | A person, with a paid tool at most as an aid |
Count the rows and a pattern appears. Six jobs suit a free plan outright: chaser letters, explanations, engagement clause drafts, formula help, email threads once anonymised, and variance commentary. Three more are conditional: spreadsheet extracts, meeting summaries and research as a starting point. The remaining five need a paid plan, a spreadsheet or a person. If your practice only does ChatGPT-based work, the free ChatGPT tutorial for accountants goes deeper on that single plan; this one compares across tools.
What each free plan states about its limits
I read each vendor's own pages in early November 2026. Where a vendor publishes no number, the table says so, because "unlimited" and "unstated" are different things when month-end lands.
| Tool | What the vendor states | What matters in a practice |
|---|---|---|
| ChatGPT Free | $0; no message count published; 3 file uploads a day; Projects hold 5 files; Temporary Chat isn't used for training but may be kept up to 30 days | The daily upload cap rules out batch document work |
| Claude Free | Limited usage every five hours and every week, no numbers; up to 5 projects; up to 20 files per chat; training listed as opt-out | Long documents drain usage quickly; set the training choice deliberately |
| Gemini Free | No numeric limits on the plan page; up to 10 files per prompt; temporary chats kept up to 72 hours | Fine for text jobs; measure the cap on a busy day |
| Gemini Notebook (formerly NotebookLM) | Standard (free): up to 100 notebooks and 50 sources per notebook, limits subject to change; Google says your data isn't used to train it unless you send feedback | Good for questions across public standards or your own manuals, not for client files |
| Copilot Chat | Included with eligible Microsoft 365 licences; covered by enterprise data protection when staff sign in with their work account | Make sure staff use the work sign-in, not a personal account; it mainly uses web data unless you add a file to the prompt |
| Fathom Free | Unlimited recordings and transcriptions, instant AI summaries | The natural free choice for meeting notes, with consent |
| Fireflies Free | Unlimited transcription, limited AI summaries, 400 minutes of storage per team | About 13 half-hour meetings, then you must delete or pay |
| Otter Basic | 300 minutes per user a month, 30 minutes per conversation | Ten half-hour meetings, and anything longer is cut off |
The trouble with unstated caps is that you find them at the worst moment. A fortnight before a filing deadline, an assistant that has been fine all quarter tells you to come back in five hours. Plan for that in two ways: don't make a deadline-critical task depend on a free plan, and keep a second tool open for the same kind of work so one cap doesn't stop the day.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowClient-data rules: four tiers a whole practice can follow
Staff won't read a policy, but they will follow a four-line rule. Sort material into tiers, and say which plan each tier may touch.
| Tier | Examples | Free plan allowed? | Condition |
|---|---|---|---|
| 0: Public or yours | Published guidance, your templates, your own marketing | Yes | None |
| 1: Anonymised | Rounded figures with the client name, numbers and dates of birth removed | Yes, with a check | A second person can't tell who the client is |
| 2: Identifiable client records | Statements, ledgers, correspondence, signed documents | No | A paid plan with no-training terms, or the practice software's own AI, and client terms that allow it |
| 3: Credentials and identifiers | Tax IDs, bank logins, payroll with personal data | Never | Not in any general AI tool |
Here is the anonymising step on a real-shaped query, so the line between tiers is concrete. The "before" is what is in the file, and the "after" is what may go into a free plan.
| Before (tier 2: don't paste) | After (tier 1: may paste) |
|---|---|
| Client name, a named bakery business, sales of 84,213.40 in March, supplier invoice 7731 from a named wholesaler disputed, owner's tax reference and bank account details in the footer | A small food retailer, sales of about 84,000 in the month, one supplier invoice disputed, no references or account details |
If staff are already using free tools without a rule, the first job is finding out what has been pasted. The shadow AI tutorial covers how to do that without a witch-hunt, and the anonymising walkthrough has more patterns than the one above.
A six-person practice's month on free plans
The example practice has six people: two partners, two seniors and two juniors. The table lists a month of the jobs that suit free plans, with before and after times that are my assumptions for illustration. They are not measured results, so time your own team for two weeks before you rely on them.
| Job | Per month | Minutes each, before and after | Minutes saved |
|---|---|---|---|
| Chaser letters | 40 | 8, then 3 | 200 |
| Meeting write-ups (Fathom Free) | 12 | 15, then 5 | 120 |
| Plain-English client explanations | 20 | 10, then 4 | 120 |
| Spreadsheet formula problems | 15 | 12, then 5 | 105 |
| Variance commentaries from anonymised figures | 8 | 25, then 15 | 80 |
| Total | 625 (about 10.4 hours) |
The variance commentaries save the least, because anonymising and checking the numbers eats most of the gain, which is a useful warning that not every job speeds up by the same fraction. Now look at the caps. Twelve half-hour meetings is 360 minutes: more than Otter Basic's 300 minutes a month, inside Fireflies' 400 minutes of storage (until the next month's calls pile up), and no problem for Fathom Free. The chat assistants' caps are the unknown, and in a busy month the most likely casualty is the afternoon the whole practice tries to finish commentaries at once.
At an assumed blended $60 an hour, 625 minutes is about $625 of time a month. Six business-plan seats cost $120 on annual billing or $150 monthly, which is 19 to 24 per cent of that. The real argument for paying, though, is the work free plans shouldn't do. Suppose the practice reads 30 client PDFs a month, such as statements and contracts, and a business plan with a human reconciliation cuts 12 minutes from each: 360 minutes, or $360 of time, against $120 to $150 of seats. That job is the one the free tier locks you out of, and it is worth more than the whole free list combined.
Three prompts that work on free plans, and what to fix
A chaser letter
Write a polite, firm email of about 120 words to a client who hasn't sent the records we need to finish their year-end accounts.
Use [client first name], [your name], [list of missing items] and [deadline date] as placeholders.
Say neutrally what happens if we don't receive them by the deadline, offer a 15-minute call, and don't threaten.An illustrative result: a short email that opens with the year-end, lists the missing items, names the deadline, says that without them "we may have to prepare the accounts on the information we hold and it may be less accurate", and offers a call. What to fix: the consequence sentence is where models go wrong, because they pick a plausible penalty that isn't in your engagement terms. Replace it with what your letter actually says, check the deadline, and remove any wording that sounds like a promise about outcomes. The engagement letters tutorial shows how to keep clauses and chasers consistent.
A duplicate-invoice formula
In Excel, column B holds supplier names, column C invoice numbers and column D amounts for rows 2 to 500.
Write a formula for column E that returns "CHECK" when the same supplier and invoice number appear more than once, and "" otherwise.
Explain it in two sentences.A typical answer is this formula, filled down from row 2:
=IF(COUNTIFS($B$2:$B$500,B2,$C$2:$C$500,C2)>1,"CHECK","")What to fix is not the formula but the test. Build a dummy sheet with three planted duplicates, one with a trailing space after the supplier name and one with a different spelling. The formula will catch the exact duplicates and miss the other two, because the matching is literal. Tell the model what you saw and ask for a version that trims spaces, and extend the range beyond row 500 if the real data can grow.
Variance commentary from anonymised figures
Anonymised monthly figures for a small retail client, this month against the same month last year: revenue 84,000 against 91,000; cost of sales 52,000 against 51,000; wages 21,000 against 19,500; other costs 6,500 against 6,000.
Write three sentences of commentary a client would understand. State only what the figures show. Do not suggest causes. List two questions to ask the client.The correct figures, which you should compute yourself before reading the model's answer: revenue down 7,000, or 7.7 per cent; gross profit down from 40,000 to 32,000, a margin of 38.1 per cent against 44.0 per cent; profit after wages and other costs down from 14,500 to 4,500. A good answer states those and asks whether prices or discounts changed and whether anything unusual sits in cost of sales. The failure to watch for is an answer that adds "due to higher supplier prices". Nothing in the figures says that. A free model will often produce such a sentence unprompted, and a client may repeat it as fact. Strike every cause that the figures don't show, and recompute every number the model quotes.
Recording client meetings: the consent line and the retention rule
Fathom's free plan makes meeting notes cheap, which is exactly why a practice needs a rule before the first recording. Two decisions matter: how you ask, and how long anything is kept. Here is wording for the start of a call, and a retention rule to pair it with.
Before we start, I use an AI notetaker that records and summarises our calls so I can send you accurate notes. Only our team sees the recording. Is that all right with you? If you'd rather I took notes by hand, that is no problem.
For retention, a workable rule is to keep the approved notes on the client file, check the summary against the recording once, and delete the audio and raw transcript when the notes are approved. That is my suggestion, not a regulation; your professional body, insurer and engagement terms may require something different, so check them. If a client says no, believe them, and make a note on their file so the next person doesn't switch the notetaker on by habit. For a comparison of notetakers on cost and retention settings, the best AI note taker for accountants goes through the options.
One line for your engagement terms
If staff are going to use any of these tools, tell clients in the document they sign, not in a footnote nobody reads. A short paragraph is enough, and it should say what you do and what you don't. Here is a draft to adapt; have your adviser look at it before it goes into a live engagement letter.
We may use AI tools to help draft routine correspondence and to summarise meetings. We do not put your identifiable records into tools that use content to train their models. A member of our team reviews everything produced with these tools before it reaches you. If you would prefer that we did not use AI tools on your work, tell us and we will note this on your file.
The last sentence matters. It converts a policy into a choice the client can make, and it gives staff a flag to look for before they open a tool. Write the paragraph once, add it to the template, and the rule about tiers has something to point at when a client asks.
Checks a person still does, whatever the plan
- Recompute every figure. Models state sums with confidence and sometimes wrongly. Rebuild the number in the spreadsheet before it reaches a client.
- Open every source. If an answer cites guidance, open the page and confirm it says what the model claims and is still current.
- Read for invented causes and promises. The commonest slips are a reason the data doesn't support and a commitment the practice didn't make.
- Keep a log. Note which tool touched which piece of work, at what tier of data, and who reviewed it. If a client or insurer asks, the log is your answer.
- Tell clients where it matters. If engagement terms say how you handle client information, add a line on AI tools rather than hoping nobody asks.
What this tutorial leaves out
Two groups of tools sit outside this tutorial on purpose. The first is bookkeeping capture: software that reads receipts and bank feeds and proposes entries. It is priced and judged differently, and the receipt capture and categorisation tutorial covers it. The second is the "AI solver" websites that top the search results for this topic. Many are built for students, answer in a form you can't audit, and invite you to paste your problem into a site whose terms you haven't read. For client work they sit at tier 3 of the rule above: not in any general AI tool until you have read what happens to the text.
When to pay, and for what
The triggers are about the work, not the vendor. Pay when any of these is true: staff are already putting identifiable client records into free accounts; a job in the table above that needs tier 2 data would save you more than the seats cost; or a free cap stops work in the two weeks before a deadline.
| The need | Paid step | List price |
|---|---|---|
| Client-data jobs with no-training terms and one shared workspace | ChatGPT Business or Claude Team | $20 a seat a month annual, $25 monthly; two seats minimum |
| Meeting notes past the free caps | Fireflies Pro or Otter Pro | $10 or $8.33 a seat a month annual; $18 or $16.99 monthly |
| AI inside the Microsoft tools the practice already runs | A Copilot licence on top of the business plan | Per-seat price on Microsoft's page; confirm the billing period |
| AI inside practice or accounting software | The vendor's own add-on | Priced per vendor; ask each for a written quote |
My order for a small practice: write the four-tier rule this week, move the two or three people who handle tier 2 material onto a business plan, and leave everyone else on free plans for tier 0 and tier 1 work, with the checks above. Time the jobs for a fortnight, compare with the table, and revisit the paid question when a cap stops work or a client asks how you handle their data.
Further reads
- Best AI Tools for Small Accounting Firms, Sorted by Task (2026) — A paid-tool shortlist sorted by task for small firms.
- Is Copilot Worth It for a Small Accounting Firm? Tasks and Seat Cost — Whether a Copilot seat earns its price in a practice.
- Bookkeeping AI Reviews: How to Test a Tool on Your Own Client Files — How to test a capture tool on your own client files.
- Is It Safe for an Accountant to Use ChatGPT With Client Data? — The privacy question for one tool, worked through.
- How Accountants Use AI to Spot Errors in Client Books — More checks like the duplicate-invoice formula.
- How Much Does AI Cost a Small Accounting Firm? — What AI costs a small firm across a year.
- Claude vs ChatGPT for Accounting Firms: Handling Client PDFs — Handling client PDFs: Claude and ChatGPT compared.
- How to Chase Missing Client Records With AI Before Deadlines — Chasing missing records before deadlines, in more depth.
- AI Powered Bookkeeping Platform: How a Practice Chooses One — A test-file method for picking an AI bookkeeping platform: three real client months each, scored on accuracy, review time and cost per client.
- Karbon Pricing: Team vs Business Plans and Where the AI Sits — Karbon's Team and Business plans priced for 3, 6 and 10 users, with where Kai and the AI agents sit and which add-on fees to check.
- AI for Accountants in Australia: Tax Agent Duties and Client Data — The Code of Professional Conduct, client consent and Privacy Act steps for tax and BAS agents using AI, plus three safe first jobs.
- AI for Accountants in Canada: Privacy Rules and First Uses — Which privacy laws and CPA conduct rules apply when a Canadian firm uses AI, how to vet a vendor, and where to start safely.
- Free AI for Tax Professionals: Where the Free Tier Stops at Client Data — Which tax-practice jobs suit a free AI tier, which never should touch return information, and what a paid business plan changes and does not change.
- Ten AI Use Cases in a Small Accounting Practice, Ranked by Set-Up Effort — Ten AI jobs for a small accounting practice, ranked by how much set-up and risk each carries, with an example and a first step for every one.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: pricing pages for ChatGPT, Claude, Fathom, Fireflies and Otter; Google's Gemini and Gemini Notebook help pages; Microsoft's Copilot Chat documentation; OpenAI and Anthropic data-controls information.