Choose an accountant who can show you how AI fits their process: which tools touch your data, who checks every AI output before it reaches you or a tax return, and what the engagement letter says about it. Specific answers and a named reviewer are good signs; vague enthusiasm, or a refusal to discuss AI at all, are both warnings.
Keep the usual criteria first: a qualified accountant who knows businesses like yours, replies within a reasonable time and quotes a clear fee. AI competence is the tiebreaker and the risk check. A practice that uses AI well should be quicker and more accurate on routine work and have more time left for advice. One that uses it badly can put your payroll in a consumer chatbot, or send you a confident explanation of your accounts that nobody actually read.
The example running through this tutorial is a café with two owners and six part-time staff, about 1,100 card transactions and 60 supplier bills a month, and a weekly payroll. Its owners wanted a new accountant after their old one retired, and they wanted one who would make month-end less painful.
What good AI use looks like inside a small practice
Before you interview anyone, it helps to know what sensible AI use in a practice actually is, so you can recognise it. Most of it is unglamorous: software that suggests matches and categories, and a person who approves them. The table sets out the common tasks and where the human must stay involved. If you'd like to see these tasks described from the practice side, the tutorial on how small accounting firms use AI day to day gives real examples.
| Task | What the AI does | What a person must still do |
|---|---|---|
| Bank reconciliation | Suggests matches between bank lines and invoices; in Xero, JAX auto-reconciliation is available on the Growing plan and above | Review exceptions and anything unusual before month-end is closed |
| Receipts and bills | Reads supplier, date, total and tax from photos or PDFs; Xero's built-in Smart Document Capture, announced in July 2026, is free on all plans | Spot-check totals and tax amounts, especially on handwritten or faded receipts |
| Error spotting | Flags duplicates, odd amounts and sudden changes in a category | Decide which flags matter and ask you about them |
| Client emails and explanations | Drafts plain-English summaries of your figures | Check every number and claim against your actual accounts |
| Tax and planning | Can help research rules and draft options | A qualified person makes the judgement and signs it off |
| Payroll | May check for anomalies inside payroll software | Keep payroll data out of general chatbots unless on a business plan with a clear policy |
The last two rows matter most. AI tools can draft a tax explanation that reads well and is wrong. OpenAI's own usage policies, updated in October 2025, rule out using its services to give tailored advice that needs a licence without a licensed professional involved. An accountant who uses AI well treats it as a fast junior, never as the person signing.
Eight questions for the first meeting, with answers to listen for
Ask these in the first meeting, and note how quickly and specifically each is answered. A practice that has thought about AI will answer most of them in a sentence or two. The answers below are illustrative.
- "Which AI tools will touch our data?" Good: "Xero's reconciliation suggestions, our receipt capture app, and Microsoft 365 Copilot for drafting emails, all on business accounts." Weak: "We use all the latest AI."
- "Who reviews AI output before it reaches us or a return?" Good: a named person and a described check, such as "your client manager reviews every reconciliation exception and reads any AI-drafted email before it's sent". Weak: "The software's very accurate."
- "Are those tools on business plans, and is our data used to train models?" Good: "Business plans only; none of them train on our client data by default, and staff can't use personal accounts for client work." Weak: silence, or "I'd have to check with our IT person" with no follow-up.
- "Do you have a written AI policy for staff?" Good: "Yes, a one-page policy; I can send it." Many small practices won't have one yet, and that's forgivable if the other answers are strong.
- "What does AI change about your fee to us?" Good: an honest answer about where time is saved and what they now include, such as a monthly check-in. Weak: "Nothing, it just helps us."
- "What's a mistake your software made, and how did you catch it?" Good: a specific story. A practice that claims its tools never err isn't checking them.
- "What will you ask us to change on our side?" Good: "Photograph receipts on the day, use one card for business spending, and connect your till payouts." These answers show they understand where AI accuracy really comes from.
- "If we leave, what do we take with us?" Good: "Everything is in your own Xero file; you're the subscriber and we're invited as advisers." Weak: "It all lives in our portal."
The café's owners asked all eight questions of three practices. Two gave specific answers to at least six; the third gave warm but general answers to every one, which told them what they needed to know.
The engagement letter test
The engagement letter is the contract that sets out what the accountant will do. More practices now add wording about AI to it, and reading that wording is a quick way to judge how seriously they take the subject. If the letter says nothing at all, ask them to add a short clause; how they react is informative. Accountants who want a template for their side can find one in the tutorial on AI engagement letters for accountants.
Here's the difference in practice. Before, a typical letter from one of the café's candidates:
"We may use third-party software and service providers in delivering our services."
After the café asked for more, the same practice sent this:
"We use AI features within Xero and our document-capture software to suggest transaction matches and read receipts, and Microsoft 365 Copilot on our business licence to draft correspondence. These tools do not use your data to train AI models. A qualified member of staff reviews all AI-suggested reconciliations before month-end is closed, and all figures in correspondence are checked against your records before sending. We do not use AI tools to make tax judgements. We will tell you before we introduce any new AI tool that processes your data."
The second version names the tools, says who checks, draws a line around tax judgement and promises notice of change. You don't need identical wording, but those four elements are worth asking for.
Your data: tools, plans and settings to ask about
Most of the real risk with AI in a practice is ordinary data handling. A staff member pasting your payroll export into a personal chatbot account to tidy the formatting is a more likely problem than any exotic AI failure. The question-by-question version of this, written for accountants, is in is it safe for an accountant to use ChatGPT with client data; from your side, three checks cover most of it.
- Business or consumer accounts. Business plans such as ChatGPT Business, Claude Team, Gemini in Google Workspace and Microsoft 365 Copilot don't train on business content by default. Consumer plans let each user switch model training off in their privacy settings, which depends on every individual remembering to do it. Ask which the practice uses.
- AI inside the ledger. Some AI features can't be avoided. Intuit says QuickBooks Online's AI features can't currently be switched off individually, so if your books are in QuickBooks, the question becomes how the practice reviews what those features do, not whether they're on.
- Where the files live. Ask where source documents go after capture, who at the practice can see them, and how long they're kept. You want a short, confident answer.
If your business handles especially sensitive data, such as health information about customers, ask your accountant how their tools fit data-protection law such as the GDPR, and involve your own adviser if the answers leave you unsure.
Will AI make the fees lower? Comparing quotes properly
Sometimes, a little. More often the saving shows up as more included in the same fee, a faster month-end, or time for a quarterly conversation about your numbers. Be wary of any quote that is much cheaper because "AI does it all", and just as wary of a quote that has risen because of new "technology fees" without any change in what you get.
The way to compare is by scope, not headline price. Put each quote into the same grid. These are the three quotes the café received, with illustrative figures:
| What's included | Practice A (online, AI-first) | Practice B (two partners, little AI) | Practice C (small practice, AI with review) |
|---|---|---|---|
| Monthly fee | $240 | $420 | $360 |
| Bookkeeping and reconciliation | Yes, by software, reviewed quarterly | Yes, by hand, monthly | Yes, AI-suggested, reviewed monthly |
| Payroll (weekly) | Extra $60 | Included | Included |
| Year-end accounts and tax return | Included | Included | Included |
| Named contact | Rotating team | Partner | Client manager |
| Month-end deadline | Not stated | Day 20 | Day 10 |
| Management check-in | None | Annual | Quarterly, 30 minutes |
| Books kept in | Firm's own platform | Client's Xero file | Client's Xero file |
Once payroll is added, Practice A costs $300 a month, only $60 less than Practice C, and it offers a rotating team, quarterly review of the AI's work and books held on its own platform. Practice B is the most expensive and the slowest at month-end. On price alone A looks best; on what the café actually gets, C is ahead.
Worked example: how the café chose, and the sums behind it
The café scored the three practices on the eight meeting questions (one point for each specific, credible answer) and on the quote grid. Practice A scored 3 of 8: good on tools, vague on review, and unable to say what happens to the books if the café leaves. Practice B scored 4: honest that it barely used AI, strong on review, weak on turnaround. Practice C scored 7, missing only a written AI policy, which it promised to finish within a month.
Practice C then made a recommendation that shows what "using AI well" means in money. The café was on Xero's Early plan. Practice C suggested moving to Growing so that JAX auto-reconciliation would handle the 1,100 monthly card payouts and bills. At list prices that's $55 a month instead of $25, so an extra $30 a month or $360 a year. Practice C estimated that reconciliation would drop from about six hours a month to two hours of reviewing exceptions, and agreed to hold its fee for two years in return. At the owners' own valuation of their time, even two saved hours a month on their side (fewer queries, fewer shoebox receipts) was worth more than the $30.
The café also asked Practice C to show an example of an AI-drafted client email, and the example is worth reproducing because it shows the review step working. The draft, produced by the practice's assistant from the café's quarterly figures, read:
"Your gross margin fell from 68% to 63% this quarter, mainly because of the rent increase in April. Consider negotiating with your landlord."
The client manager caught the error before sending. Rent doesn't sit in cost of sales, so it can't move gross margin; the real cause was a jump in dairy and coffee bean prices across three suppliers. The version the café received read:
"Your gross margin fell from 68% to 63% this quarter. The main cause is ingredient costs: milk and coffee beans rose across three suppliers, adding about $1,900 over the quarter. Rent rose too, but that affects your net profit, not gross margin. Worth a look: whether your two best-selling drinks need a small price change."
That is the whole case for a human reviewer in one paragraph. The AI draft saved time and was plausible; only someone who knew where rent sits in a profit and loss account could see it was wrong.
Warning signs that show up after you've signed
Even a careful choice can go wrong, and the early signs are usually small. Three illustrative cases show what to watch for; the tutorial on AI mistakes that damage an accounting firm's client trust covers more.
- A category that is quietly wrong for months. A bank rule, set up by AI suggestion, filed a café's monthly coffee-machine lease payment as "repairs and maintenance" for eight months. It showed up only when the owners asked why repairs looked so high. Ask your accountant to show you the bank rules they've created, once, in the first quarter.
- Generic answers to specific questions. An owner asked why cash in the bank didn't match profit and received three paragraphs of textbook explanation that never mentioned their actual loan repayments. That is often an unreviewed AI draft. Reply asking for the answer in terms of your own numbers.
- A request to send data somewhere new. If you're suddenly asked to upload bank statements to an unfamiliar portal or link, ask what the tool is and whether it's in the engagement letter. The practice promised notice of new tools; hold it to that.
None of these is a reason to leave on its own. How the practice responds when you raise them is the real test: a good one fixes the rule, rewrites the answer, or explains the tool within a day or two.
The 90-day review that confirms your choice
Put a date in the diary three months after you start and check these points. The café's owners did it in twenty minutes over coffee, with their first quarter's reports open.
- Was each month closed by the promised day? Count the misses.
- Pick ten transactions at random from the bank feed. Are they in the categories you'd expect? One odd one is normal; four is a pattern.
- Did every AI-drafted email or report you received make sense against your own figures?
- Did the practice ask you sensible questions, such as about unusual payments, rather than guessing?
- Has the practice introduced any new tool touching your data? Were you told first?
- Is your ledger still in your own name and subscription, with you as the owner?
If your bookkeeping is done by a separate service rather than the accountant, the questions are slightly different, because you're judging a monthly processing operation rather than an adviser. The tutorial on choosing an outsourced bookkeeping service that uses AI covers that side, and how accountants use AI to spot errors in client books shows the checks a good practice should be running on your behalf.
Further reads
- Can AI Do My Bookkeeping? What Still Needs an Accountant — Which jobs software can take over and which still need your accountant.
- Xero vs QuickBooks AI: Which Saves More Bookkeeping Time? — Compare the AI inside the two ledgers most accountants will suggest.
- How Accountants Use AI to Explain Tax to Clients in Plain English — What good AI-assisted client explanations look like from the practice's side.
- AI Receipt Capture and Bank Categorisation: A Bookkeeper's Setup — How receipt capture and bank rules are set up properly.
- How Much Does AI Bookkeeping Software Cost per Month? — Price the software side before you compare accountants' fees.
- Turning Hours Saved by AI Into Advisory Work Clients Pay For — What a practice should do with the hours AI frees up.
- Can AI Do My Business Taxes? What It Can Safely Prepare — What AI can safely prepare for your business tax return, where its tax answers go wrong, and a year-end pack you can build before the accountant sees it.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Xero pricing and product pages (plans, JAX reconciliation, Smart Document Capture) and Intuit's QuickBooks Online information as summarised in the series fact sheet; OpenAI usage policies (29 October 2025).