Can AI Help Accountants? What It Handles and What You Review

Yes, AI can help an accounting practice, and the reliable jobs have clear inputs and checkable outputs: capturing receipts, suggesting transaction categories, summarising documents and drafting routine emails. Anything involving tax judgement, an opinion or a signature needs an accountant to review and sign. Split the work by those two lists, then name a reviewer for each job.

Brand overviews tell you what each product does. They do not tell you who in your practice checks the output, how long that takes, or what happens when the model is confidently wrong. Those three questions decide whether AI saves a small practice time or quietly adds risk, so the split below sorts a typical service list into three groups, gives each job a reviewer and a sign-off, and ends with a way to test it on your own files before you rely on it.

Follow me on Instagram@sagnikteaches

Three groups: AI-reliable, AI-then-review, accountant-only

Instead of asking whether AI can do accounting, sort each job by two questions. Is the output easy to check against a source? And what happens if it is wrong and nobody notices? The answers give three groups.

Connect on LinkedInSagnik Bhattacharya

Two rules keep the split honest. Nothing moves to a lighter-review group without a test on your own files, and anything that touches a filing or a signature stays in group C no matter how good the tool looks. The sections that follow take each group in turn, with examples.

Subscribe on YouTube@sagnikteaches

Group A: receipt capture and categorisation

This is where most practices should start. A receipt photo or a supplier invoice comes in, the tool reads the supplier, date, total, tax amount and line items, and suggests a category. The inputs are visible, the outputs are fields you can check at a glance, and an error shows up in a reconciliation. For a full setup, the tutorial on AI receipt capture and bank categorisation goes step by step, and Dext vs Hubdoc compares two capture tools.

What can go wrong

• A foreign-currency receipt read as the home currency.
• A date format flipped, putting an expense in the wrong month.
• Split receipts (food and household goods) coded wholly to one category.
• A personal purchase coded to the business because the supplier is one the client uses for both.
• A duplicate, when the same receipt arrives by email and by photo.

Forty clients and 2,400 receipts: the sum

Say a six-person practice looks after 40 small clients, and a typical client sends about 60 receipts a month. That is 2,400 items. Suppose hand-keying takes two minutes each, which is 80 hours a month. With capture and suggested coding, the practice assumes each item needs 20 seconds of glance review, about 13 hours, plus a deeper check on the exceptions the tool flags. The illustrative saving is the difference, minus set-up and the exception work. The practice's own test should replace these assumptions with measured numbers: time ten clients both ways for one month.

The review routine

1. Sample 5% of coded items for each client, weighted towards the largest values.
2. Read every item the tool flagged as low confidence.
3. Check the three biggest categories against the prior month. A category that jumps by a third needs an explanation.
4. Record the sample result: items checked, errors found, error type. Initial it.

If the error rate in the sample is above your threshold, say more than one in twenty, review everything for that client that month and look for the cause: often one supplier or one setting.

Group A: bank categorisation, summaries and email drafts

Bank-feed rules and AI suggestions work well for recurring items: rent, subscriptions, utilities, payroll runs, card fees. They work poorly for one-off items and anything with a mixed purpose. A good set-up lets the tool auto-suggest for items it has seen before and send new payees to a person. Xero's pricing page, for example, lists an Early plan at $27 a month, with auto-reconcile shown as an optional $5-a-month add-on on Early and included on the Growing and Established plans; check current prices and what each plan includes before you plan around them.

Summaries and email drafts are the easiest AI jobs to start and the easiest to get lazy with. Say a cleaning company owner sends a long, rambling email asking about the van she is thinking of buying. The model produces a clean summary of her questions and a draft reply. The accountant checks the facts: is the purchase within the company's means, how would it be financed, what tax treatment is relevant? The draft is faster than writing from scratch. It is not an answer until a person has made the decision it describes. For client emails in particular, keep a rule: nothing is sent that contains a figure or a tax statement which the sender has not checked.

Group B: the accountant reviews every line

Management accounts commentary

Give the model the numbers (not the whole ledger) and ask for a plain-English paragraph on what changed. It will write a tidy narrative. It will also infer causes it cannot know, such as "due to seasonality" or "reflecting higher marketing spend". The reviewer checks each causal claim against what the client told you, and deletes any the practice cannot support. The check takes five minutes and catches the sentence that would otherwise embarrass you.

Reconciliation queries

The model can list unmatched items and propose likely matches. A suggested match is a lead. The reviewer confirms with the bank line and the supporting document before posting. Never accept a bulk "match all" action on a ledger you have not looked at.

Year-end working papers

Drafts of accrual schedules, prepayment calculations and depreciation workings can come from a model. Recalculate each in a spreadsheet. Models make arithmetic errors and sometimes apply the wrong rate or period. If you use AI inside a spreadsheet, the guidance in the ChatGPT Excel guide applies: keep the formulas visible and test against a known case.

Engagement letters and client explanations

Drafts come quickly, and they sound authoritative. The reviewer checks scope, fees, responsibilities and any statement about law or deadlines. The tutorial on AI engagement letters for accountants covers the clauses worth adding, including how you use AI.

Group C: where the accountant decides

These jobs may still use AI in the background for research or to draft notes, but the output is a lead, and the accountant forms the view. Examples: whether a sole trader's van is a business asset for tax purposes, how to treat a refunded deposit, whether a director's loan needs a particular entry, what to put in a filing that carries the practice's name. Research tools can speed up the reading. They can also cite rules that do not exist or have changed. The accountant checks the primary source, records what was relied on, and signs.

A helpful rule: if a regulator, a lender or a court could read the output, a named human has reviewed it and put their name on it.

A categorisation prompt, a sample output and what to fix

If you use a general model on a transaction export (redacted of account numbers), constrain it to your own chart of accounts and make it say when it is unsure. A prompt that works:

Below is a list of bank transactions and my chart of accounts.
For each transaction, return: date, description, amount, suggested
account (from my chart only), confidence (high, medium, low) and a
one-line reason. If the payee could be business or personal, or
could fit two accounts, mark confidence low and say why. Do not
invent accounts. Do not merge transactions.

An illustrative extract of what comes back:

03/04  AMZN MKTP  84.20   Office supplies     medium  Marketplace; could be stock or personal
05/04  SHELL 0412 61.00   Motor expenses      high    Fuel, matches earlier pattern
09/04  TFR [NAME] 420.00  Subcontractors      low     Individual payee, no invoice reference

What you do with it: accept the "high" rows on a sample basis, look up the invoice for the low row, and ask the client what the marketplace purchase was. Then add a rule only for payees whose coding is always the same. A rule written from one guess becomes a repeated error.

One month, end to end: where each group touches the work

Say a six-person practice runs the March close for a small client, a building-services company with about 90 transactions and 55 receipts. This is an illustration of how the three groups interact in a single month, with assumed times.

1. Receipts arrive (Group A). Capture reads 55 receipts in a few minutes. The bookkeeper spends 25 minutes on flagged items and a 5% sample (three receipts, all correct).
2. Bank feed (Group A). Rules code 61 of 90 transactions. The AI suggests the remaining 29. The bookkeeper accepts 18 after a glance, queries 7 with the client, and corrects 4. Time: 35 minutes.
3. Reconciliation (Group B). Two items do not match. The manager finds one is a supplier credit note and the other a payment on the wrong date. 20 minutes.
4. Commentary (Group B). The model drafts four sentences on the month. The manager deletes one that says margins fell "due to material costs", which nobody has confirmed, and replaces it with a question for the client. 10 minutes.
5. Tax-relevant query (Group C). The owner asks whether a new tool purchase should be treated as equipment or a repair. The partner decides, notes the reasoning and replies. AI was used only to pull the relevant paragraph from guidance, which the partner verified.

Roughly ninety minutes of human time across the practice, against a former half day, though your own figures will differ. The important feature is not the saving. It is that every step has a named person who could say what they checked.

How to spot silent errors

The dangerous AI mistakes are the plausible ones. These patterns surface them without re-doing the work.

• Round-number check. A total that comes out suspiciously neat, or an item that matches the budget exactly, often means the model smoothed something.
• Control totals. Compare the sum of captured receipts with the total of the batch you sent. A gap means missing or duplicated items.
• Prior-period comparison. A category that doubles or disappears needs an explanation before it is accepted.
• Reverse lookup. Pick five coded items at random and trace each back to its document. If you cannot find the source quickly, the audit trail is weak.
• Ask for the reason. When a model must give a one-line reason for each suggestion, vague reasons ("typical expense") are a sign to check that item.

Say a cleaning company's March fuel spend appears at double the usual amount. The sample check finds two photos of the same receipt, captured once from the phone and once from email. Control totals caught it. The fix is a duplicate-detection setting, plus a line in the client onboarding note asking them to use one route only.

Name the reviewer for every job

A split on paper is not enough. Write one line per job: who reviews, how much, and where the record lives. An illustrative extract for a six-person practice:

When a staff member leaves or a client changes, the table tells the new person what to do. It also gives you an answer when a client or an insurer asks how you supervise AI output.

Three cases that test the split

A plumbing firm with a messy shoebox

A plumbing firm sends three months of receipts in one go, many creased, some for fuel and tools, some personal. Capture handles 80% cleanly. The flagged 20% need a person, and the personal items need a conversation with the owner. The lesson: AI sorts the pile faster, but only the owner knows which items were personal, and no tool should guess.

An electrical contractor paid net of deductions

A contractor's income arrives net of a scheme deduction that the tool reads as a short payment. The extraction is accurate; the interpretation is wrong. Group B review catches it by comparing to the statement. Where industry-specific rules apply, the tool's suggestion needs a reviewer who knows them.

A roofing company's year-end

The model drafts a prepayment schedule for an insurance policy paid in October. It spreads the cost over twelve months from the payment date, not from the policy start date. The reviewer catches it because the policy dates are on the invoice. A check against the document, not the model's confidence, found the error.

Moving a job up or down the groups

The split is not permanent. Set a rule for promotion and demotion so the decision does not rest on how enthusiastic someone feels that week.

• Promote from B to A only after three consecutive months where the reviewer found no material errors in that job across at least five clients, and the sampling record shows it. Promotion lowers the review level to a sample, never to none.
• Demote at once when a material error reaches a client, a filing or a bank. Return the job to full review, find the cause, and write down what changed.
• Re-test after any change of tool, model, plan or client mix. A software update can alter behaviour without announcing it, so treat each as a new test.
• Never promote into C. Judgement, advice and signatures stay with a named person whatever the tool can do.

An illustrative case: a practice promotes bank categorisation to a sampled check for its twelve simplest clients. In month five a bank changes how it labels card payments, the tool starts coding a regular supplier to the wrong account, and the sample catches it at the first review. The practice demotes the job for two weeks, fixes the rule, and promotes it again. Nobody outside the practice ever knew, which is how a well-run review step should feel.

Test the split on your own files before you commit

1. Pick three clients of different types: a simple sole trader, a company with stock, and one with messy records.
2. Run one month of receipts or transactions through the tool, then do the same month by hand, or compare to the previous month's finished work.
3. Count differences and classify them: wrong amount, wrong category, wrong date, duplicate, missed item.
4. Time the review, not just the tool. A tool that saves ten minutes and costs ten in checking has saved nothing.
5. Decide per job whether it stays in its group, moves to lighter or fuller review, or is dropped.
6. Repeat after a quarter, because tools, models and your clients change.

The tutorial on using AI to spot errors in client books shows another way to use the same tools, on the review side. For software that bundles agents into the ledger itself, see the accounting agents from QuickBooks, Xero and Sage, and apply this same split to what they do.

Keep client data and trust in view

A practice handles sensitive data, and clients assume it stays within the practice's controls. Before any client information goes into an AI tool, check the plan's data terms, whether inputs are used for training, where data is stored and who can see it. Put your use of AI into the engagement letter, in plain words, and expect clients to ask how you supervise it: the reviewer table above is the honest answer, and it is easier to show than to describe.

If you would like help writing your own split, naming reviewers and choosing a first pilot, that is what my AI implementation consultation covers. Start with the receipts, write down who checks what, and let the evidence from your own files decide what moves up.

Questions practice owners ask about AI and review

Which accounting jobs should I never give to AI alone?

Anything that needs tax judgement, an opinion on treatment, a signature or a filing on the client's behalf. AI can prepare the working papers and a draft explanation, but the accountant decides, signs and takes responsibility.

How long should review take compared with doing the job by hand?

Measure it on your own files. As a rough aim, review of well-structured output such as receipt coding should be a fraction of the time of coding by hand. If review takes as long as doing it, the tool or the setting is not helping and you should stop.

Do I need to tell clients I use AI?

Many practices say so in the engagement letter, covering what is processed, by whom and with what safeguards. Check your professional body's guidance and your tool's data terms, then tell clients in plain words before you process their data.

Where should a small practice start?

With receipt capture and bank-transaction categorisation on a few cooperative clients, because the output is easy to check and the errors are visible. Add summaries and client emails next, and leave tax positions for last.

Further reads

• Can AI Do My Bookkeeping? What Still Needs an Accountant — What still needs an accountant when AI does bookkeeping.
• How Accountants Use AI to Explain Tax to Clients in Plain English — Explain tax to clients in plain English, accurately.
• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven AI mistakes that cost an accounting firm trust.
• AI Bookkeeping Agents: What They Do Alone and What You Must Review — What bookkeeping agents do alone and what you review.
• Best AI Note Taker for Accountants: Which Fits a Small Practice? — Choose a note taker for client meetings.
• How to Choose an Accountant Who Uses AI Well — See the questions clients ask about your AI use.
• AI Agents for CPA Firms: Starting Jobs and a Review Gate for Each — A small CPA firm's first four agent jobs, scored and sequenced, with a review gate, pass rule and stop condition for each, plus a 90-day start.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Xero pricing page, USD list prices (read 4 November 2026); no other vendor figures are relied on.