Hand AI the repetitive, checkable jobs first: reading receipts into a list, suggesting categories for bank lines, spotting duplicates, and drafting payment reminders. Keep judgement jobs with yourself or your accountant: tax treatment, payroll, filing, owner pay and anything you cannot trace to a document. Review everything it produces before it reaches your books.
The test is simple. Can you open the source document, compare it with the AI's answer in under a minute, and say which is right? If yes, the job is a good one to hand over. If the answer depends on a rule you would need to look up, or on a decision about tax, keep it with a person. A faded fuel receipt passes that test in thirty seconds; the question of whether a client lunch is deductible never does.
The three-list checklist
Every task below is sorted by how safely it can go to AI. The right-hand column says how to verify the result, because a task you cannot verify is not safe to hand over.
List one: hand these over first
| Task | Why it suits AI | How you check it |
|---|---|---|
| Reading receipts and invoices into rows (date, supplier, total) | Repetitive, and the document is right there | Spot-check five in ten; confirm the totals add up to your pile |
| Suggesting a category for each bank line | Pattern recognition against your last 12 months | Sort by category and scan for oddities; review anything over a set amount |
| Finding duplicate payments and double-entered bills | Matching amounts and dates is quick | Open both entries before deleting either |
| Drafting payment reminders for late invoices | Wording, not numbers | You read the amount and due date before sending |
| Summarising a month's income and costs in plain words | Turns a report into sentences | Compare the totals with the report it was written from |
| Renaming and filing scanned documents | Clerical | Open a sample from each folder |
List two: AI prepares, you review line by line
| Task | Why it needs review | What to look at |
|---|---|---|
| Matching bank lines to invoices (reconciliation) | A wrong match hides a real error | Every match over a threshold; any match where the amounts differ by cents |
| Splitting mixed receipts (supermarket, hardware store) | Some items are business, some personal | The split lines and the total |
| Spotting unusual spending | Flags are questions, not findings | Look at the source line before assuming a mistake |
| Preparing a list of questions for your accountant | It cannot know your history | Delete questions that answer themselves |
| Cash-flow forecasts for the next quarter | Built on assumptions it cannot test | Each assumption, in writing |
List three: keep with you or your accountant
- Tax treatment and filing. Whether an expense is deductible, what rate applies, which period it belongs to. These rules differ by place and change, and a confident wrong answer costs penalties.
- Payroll, benefits and owner pay. Small errors compound and affect other people's pay.
- Anything that goes to a lender, investor or authority. A person signs these.
- Write-offs, bad-debt decisions and adjusting entries. Judgement calls with consequences.
- Choosing how to treat a one-off transaction such as a vehicle purchase or a grant.
An AI can draft a question about these things for your accountant, and that is a good use. It should not be the one answering it.
The main example: a photography studio's quarter
Everything here is illustrative. A three-person photography studio keeps its books itself and passes them to an accountant each quarter. The owner takes 140 bank lines, 62 receipts, and a folder of 11 supplier invoices, and decides to hand over the first list only for one quarter, with list two kept for herself.
Step one is the receipts. She photographs the 62 receipts and asks an assistant to return a table. Her prompt:
I am attaching 62 receipt photos for a photography studio.
For each, give: file name, date, supplier, total amount, and
what was bought in five words or fewer. If any value is hard to
read, write UNSURE and say why. Do not guess. Finish with the
sum of all totals.The illustrative output has 62 rows: 55 clean, 7 marked UNSURE (faded thermal paper, a crumpled corner), and a total of $4,291.15. She adds up the amounts actually paid on her own receipts and gets $4,318.40, so the table is $27.25 too low. Tracing the gap takes ten minutes: on 6 rows the assistant read the subtotal rather than the amount paid, so tips and service charges were missing. Fixing those and the 7 unsure rows takes about 12 minutes more. Her note for next time: "Give the amount paid, not the subtotal."
The AI took 62 receipts in about four minutes. Typing them would have been two hours. Her checking took 25 minutes, so the saving is real but not 95%; plan on it taking a quarter of the original time, not none of it.
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 nowCategorising the bank lines without trusting the first pass
Step two: give the assistant her chart of accounts, which is the list of categories her accounting software uses, and last year's bank lines with their categories as examples. Without examples, models categorise by what a business of that type usually spends, not by what this studio does.
Here is my chart of accounts (list). Here are 40 bank lines from
last year with the category I chose for each. Now categorise the
attached 140 lines from this quarter. Rules:
- Use only categories from my list.
- Add a confidence: HIGH, MEDIUM or LOW.
- If a line could be personal, mark it PERSONAL?
- Do not categorise anything over $500 as an expense;
mark it CHECK.
Return a table with the original line, category, confidence.Illustrative result: 104 lines HIGH, 21 MEDIUM, 9 LOW, 4 PERSONAL?, 2 CHECK. She reviews the 36 non-HIGH lines (about 20 minutes) and spot-checks 12 of the 104 HIGH ones. She finds one error in the spot-check: a payment to a camera repair shop categorised as "equipment" that was actually a repair, which belongs elsewhere. That one error tells her to look at all equipment lines, and she finds one more. The transaction categorisation tutorial goes through this method in detail, and the receipt capture set-up shows how a bookkeeper automates it.
The rule for spot-checking
Review all lines that are not HIGH, every line above a dollar limit you choose (say $300), and a random 10% of the rest. Then sort by category and look for categories with one or two odd lines. Errors cluster, so one wrong line means look closer at that category.
Reconciliation: let it match, then check the matches
Reconciliation means confirming that what the bank says matches what your books say. AI is good at proposing matches and poor at knowing which unmatched lines matter. Xero's auto-reconcile (included on its Growing and Established plans, a $5-a-month add-on on Early) and QuickBooks' bank feed, which learns how you categorise and then matches transactions, propose matches inside the software, which keeps the data in one place; if you use a chat assistant instead, you export both lists and ask it to pair them. Either way, the check is the same.
The studio's check, in numbers. The bank statement closes at $12,806.15. The books, after the AI-assisted matching, show $12,944.75. She needs to explain a $138.60 difference, and she lists three items recorded in her books that have not reached the bank: an invoice payment received on the last day but not yet cleared (+$420.00), a supplier cheque not yet cashed (-$185.00) and a card payment that posts two days later (-$96.40). Add them up: $420.00 minus $185.00 minus $96.40 is $138.60, so the books and the bank agree. She asks the assistant to confirm that each of the three appears on the next statement, then ticks them off.
This is the stage where a wrong match hides a real error. If the assistant had matched the $420 to an unrelated invoice of the same amount, the difference would have vanished and nobody would have looked. So she reviews every match above $300 and any match where the dates are more than a week apart. For a full method see the reconciliation tutorial.
A quick sum on what it costs
The costs are small compared with the time. A paid chat plan lists at about $20 a month (ChatGPT Plus $20; Claude Pro $17 billed annually, $20 monthly). If your accounting package has built-in AI, check what your plan includes: Xero's pricing page lists Smart Document Capture as included on all three plans, with list prices of $27, $59 and $97 a month, and QuickBooks Online lists $38, $85, $140 and $340. Dext, a receipt-capture service, lists its Business plan at $302.50 a year for 250 documents a month and five users. Verify the current figures on each vendor's page, because features and prices move.
The studio's time: before, about 7 hours a quarter on data entry and categorising. After, about 2 hours of AI work and 1.5 hours of review. The saving is roughly 3.5 hours a quarter. At an illustrative $40 an hour of the owner's time, that is $140, against $60 of subscription over the same three months. The sum works, but not hugely, which is why the first-list rule matters: take the jobs where you save most per check.
When a smooth category column hid a real problem
A florist tried the same method on a quarter's card payments. The assistant categorised a $2,400 payment to a wholesale supplier as "stock purchases", a reasonable guess that sailed through at HIGH confidence. The payment was actually a deposit on a refrigeration unit, a purchase of equipment. It would have been written off as stock in the quarter, which understates profit and gets the books wrong for the asset.
The error was caught only because she had set the rule "anything over $500, mark CHECK" and so saw it. Without the rule, it would have ridden through. Her accountant saw the entry, asked for the invoice, and moved it. The cost was five minutes. The lesson: dollar thresholds catch the expensive mistakes that confidence scores miss.
Late invoices: the wording is AI's job, the numbers are yours
Chasing payment is dull, emotionally awkward and a good fit for drafting. Keep a list of unpaid invoices from your software, with the customer, invoice number, amount, due date and days overdue. Ask the assistant for a three-step sequence in your tone, and nothing else.
Write three payment reminders for an overdue invoice.
Tone: friendly and direct, never threatening.
Reminder 1: 3 days overdue. Reminder 2: 14 days overdue.
Reminder 3: 30 days overdue, asks them to call me this week.
Use placeholders: [first name], [invoice number], [amount],
[due date], [payment link]. Under 70 words each. Do not
mention fees, interest or legal action.The "do not mention fees" line matters, because a model may invent a late fee your terms never included. An illustrative reminder two: "Hello [first name], invoice [invoice number] for [amount] was due on [due date] and I cannot see a payment yet. Could you check it for me? You can pay here: [payment link]. If anything is unclear about the invoice, reply and I will fix it today." Before sending, you fill the placeholders from your software, so the amounts never pass through the model's memory. For a studio with eight overdue invoices, that is ten minutes instead of forty.
A plain-words summary of the month
Owners often skip reading their own reports because the layout is dense. Export the profit and loss for the month and the same month last year, and ask: "Compare these two reports. List the five biggest changes in dollars, say whether each is income or cost, and flag anything that looks like a one-off. Do not explain causes you cannot see in the data." An illustrative answer for the studio: income up $1,900, mostly wedding bookings; studio hire costs up $640; equipment repairs $410 against none last year; software subscriptions up $85. You then check each figure against the report, which takes five minutes, and add the causes yourself because only you know them. The summary becomes a page you can show your accountant, so quarterly questions arrive with context.
One caution: ask it never to round or "estimate" figures. A summary that says "roughly $2,000" when the report says $1,900 is harmless once, but a habit of approximate numbers makes real errors hard to see.
A month-end checklist you can paste into the notes app
- All receipts for the month photographed and read into the table; UNSURE rows resolved.
- Bank lines categorised; every non-HIGH line and every line over your threshold reviewed.
- Reconciliation run; every matched pair over $300 opened; outstanding items listed with dates.
- Overdue invoices listed; reminders sent with figures filled in by you.
- Monthly summary compared with last year; causes added in your words.
- Questions for the accountant written down, each with the source document attached.
- A note of what AI did this month, and what you changed, saved with the records.
The last line is the one people skip, and the one an accountant will thank you for. It is a plain record of what was automated and how it was checked.
Splitting a mixed receipt: a before and after
Mixed receipts are where a quick categoriser goes wrong, because one line holds two purposes. Say the studio's owner buys a printer cartridge, two lunches and a birthday card at one supermarket for $47.30. Before: the bank line reads "SUPERMARKET 47.30" and the assistant files it under "office supplies". After a prompt that asks it to read the receipt and split by item, the table shows cartridge $21.00 (office supplies), lunches $19.10 (staff meals or personal, your decision), card $7.20 (personal). The three parts add to $47.30, which you confirm before accepting anything.
The decision on the lunches is yours or your accountant's, because the treatment of meals depends on the rules where you operate. Ask the assistant to mark the line PERSONAL? and add it to your list of questions. A good habit for any receipt that mixes purposes is to photograph it the day you buy it, because thermal paper fades within weeks and an unreadable receipt is a record you cannot defend later.
Protecting the data you hand over
Bank statements and receipts contain account numbers, customer names and card fragments. Before upload, crop or redact what the task does not need, and replace customer names with codes. Check the privacy controls on your plan, since consumer ChatGPT and Claude plans have a setting that stops chats being used to improve the models, and business plans do not train on business content by default. Anthropic advises against using its browser agent for managing financial accounts, so do not hand over log-ins; give it exports instead. The same caution applies to any tool that asks to be connected to your bank.
The expense management tutorial covers receipts and approvals if you have staff claiming costs, and the Google Sheets prompts tutorial has ready-made prompts for tidying a bank export before you upload it.
A small routine for the first month
- Week 1: time your current bookkeeping tasks. Write down hours for receipts, categorising, matching and chasing invoices.
- Week 2: pick one task from list one, usually receipts. Run it side by side with your normal method on one month's receipts and compare totals.
- Week 3: add categorising for one month of bank lines with the confidence rule and the dollar threshold. Record how many lines you changed.
- Week 4: show your accountant the output and your checks. Ask what they want attached to each entry, and whether the way you named categories matches theirs.
After a month you have two numbers: hours saved and corrections made. If corrections are more than a tenth of the lines, tighten the prompt or give more examples before adding another task. Never move a job from list two to list one because it felt fine for a month; move it only after your accountant has reviewed a quarter of the output and found nothing material.
Tax and authority rules: where the caution is
This checklist stays out of tax rules on purpose. The rules differ by place, change, and depend on your business form. A model asked "is this deductible?" will give a fluent answer that may be wrong for your situation, and your tax authority will hold you responsible for the return, not the tool. Use AI to draft the question and to gather the receipts, then put the question to a qualified person and write down their answer with the date. If you want help working out where AI belongs in your own routine, that is what my AI implementation consultation maps, and you can reach me through the contact page.
Questions owners ask before handing over the books
Can AI do my bookkeeping for me?
It can do the repetitive parts: reading receipts, suggesting categories, matching bank lines and drafting reminders. It cannot take responsibility for the figures. You or your accountant still review what it produced, especially anything that affects a tax return, payroll or a lender's view of the business.
Is it safe to upload bank statements to a chat assistant?
Only after you have checked the privacy settings on your plan and removed what the job does not need, such as full account numbers and customer names. Business plans do not train on business content by default; individual plans have a setting you can turn off. Use your accounting software's own AI features where you can, because the data stays in one place.
Which accounting package has the best AI?
The one your accountant already uses is usually the right start, because they can review what it did. Xero includes document capture and auto-reconcile; QuickBooks learns how you categorise and matches bank transactions. Features change monthly, so check the vendor's page for what your plan includes before you pay for an upgrade.
Should I tell my accountant I am using AI?
Yes. Tell them which jobs, what you check afterwards, and whether anything goes through a chat assistant rather than the accounting software. Most accountants will happily work with AI-sorted records as long as the supporting documents are attached and the sums can be traced.
Further reads
- Xero JAX Bank Reconciliation: Set It Up and Check the Matches — Sets up Xero's automatic bank matching with checks.
- Catch-Up Bookkeeping With AI: Rescue Months of Untouched Books — Recovers months of untouched books.
- Best AI Tools for Small Accounting Firms, Sorted by Task (2026) — Tools your accountant may already use.
- Is AI Worth It for a One-Person Bookkeeping Business? — Costs and gains of AI for a one-person practice.
- ChatGPT for Accounting With Excel: Seven Tasks, Each With Its Check — Shows how accountants check each AI task.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: xero.com pricing plans, quickbooks.intuit.com pricing, dext.com business pricing, claude.com/pricing, openai.com ChatGPT pricing, Claude Help Centre (Use Claude in Chrome safely) (all checked 5 November 2026). Tax rules differ by place and are not quoted.