AI Accounting Workflow: Client Upload to Reviewed Books, Step by Step

Five stages: the client uploads and the bookkeeper checks completeness, AI captures document data, the ledger's suggestions code and reconcile bank lines, a senior reviews anomalies against prior months, and the partner signs off judgement items. Each stage has a named owner and a check that must pass before the work moves on.

AI speeds stages two and three and drafts the questions in stage four. It does not remove the checks, and the checks are where a practice's liability sits: a suggestion accepted without a look becomes the practice's own entry. So every stage ends with something a person verifies.

Follow me on Instagram@sagnikteaches

The worked month below follows an electrical contractor with ten staff, one bank account, one credit card and about 120 bank lines a month. The figures are illustrations, and the tools are examples; each practice should swap in its own.

Connect on LinkedInSagnik Bhattacharya

Two habits carry most of the benefit: counting what should arrive before chasing what has not, and sampling accepted AI suggestions by risk instead of by mood. Both cost minutes, and both are easy to skip when the month is busy.

Subscribe on YouTube@sagnikteaches

The five stages with owner, tool and check

Three rules to write down before the first month

A workflow with no written limits drifts to the slowest client and the loudest exception. These three rules take an hour to agree and save arguments every month.

A cut-off calendar. Statements by day 3, documents by day 8, partner sign-off by day 12. A client who misses day 8 gets a one-line email saying the books will be reviewed on what is there and the rest carried into next month. Without a written cut-off, every month waits for the slowest upload.

Thresholds by client size. One set of numbers does not suit a ten-line sole trader and a 300-line contractor. These are starting points, set per client with the partner:

The contractor, at 118 lines, sits in the middle row, which is where the $200 and $150 limits in stages 3 and 4 come from. A file note per stage. One line in the practice's file records the date, the person, the tool and the counts. For stage 3 in the worked month it reads: "118 lines, 84 auto-matched, 26 sampled, 2 errors ($490), suspense $2,140 cleared, ledger agrees to statement after $64 duplicate removed, [initials]". If a client or an insurer ever asks what review took place, that line is the answer.

Stage 1: upload, and knowing what is missing

Chasing is cheaper when you know the number you are waiting for. Take the average count of bills and receipts from the last three months, round it, and call that the expected count. For the contractor it is 42 (the average of 41.7 across three months). Send the month-end request on day 1 with a short checklist the client can tick off.

• Bank statements for the current account and the credit card, to the last day of the month, by day 3.
• Supplier bills and receipts uploaded to the portal, card receipts photographed the same day.
• The month's sales list or completed-jobs list.
• The payroll provider's summary for the month.
• Loan and finance statements, including the fuel card.
• Anything unusual: a new vehicle, a new supplier, personal spending on the business card.

On day 5, 33 of the expected 42 documents are in. The missing list comes from a pivot of last quarter's bills by supplier against this month's uploads, which takes two minutes and does not need AI. The assistant's job is the message, drafted from a template and sent by a person:

Hi [first name], for [month] we have 33 of the 42 bills and receipts we usually receive. Still missing: [supplier list]. Please upload them by [date]. If a receipt is lost, tell us the date and amount and we will use the bank line.

By day 8, 41 are in and one receipt is genuinely lost, so the bookkeeper notes it against the bank line and moves on. The check for stage 1 is simple: statements reach month-end, and the document count is within one or two of expected, with every gap explained.

The difference shows in the first email. Before: "Hi, can you send everything for October?" brings a trickle of attachments over three weeks and no way to tell when it is finished. After: the same email carries the checklist, the expected count and the day-8 deadline, and the client can see for themselves what "everything" means.

Stage 2: capture, and the exceptions queue

Two capture routes cover most small clients. Xero's pricing page lists Smart Document Capture on all three plans, with automated bill entry on the Growing and Established plans and a five-bill limit on Early. Dext's business pricing shows a single configurable plan, starting at 250 documents a month and 5 users, at $302.50 billed annually. AI Assist, which Dext describes as noticing the calls you make on categories, tax rules and suppliers and offering to make them for you, with nothing applied until you approve it, is priced separately: an add-on at $10.50 a month on that plan until 31 January 2027, and $7 per client for custom guidance on the practice plans. For practices, Dext lists Essentials from $17.70 and Advanced from $19.20 per client per month, excluding tax, with a ten-client minimum. The page also shows higher totals at that minimum ($353.99 and $383.99), so ask for a quote at your own client count. Dext pricing for accountants compares the practice plans in more detail.

Whichever tool reads the documents, the exceptions queue is the check. Five rules catch most problems before they reach the ledger:

1. Duplicates: the same supplier, invoice number and total.
2. Line items that do not add up to the stated total.
3. A supplier with no history in this client's ledger.
4. A document dated outside the month or in the next accounting period.
5. Any total above a limit you set, say $1,000, opened and read against the image regardless of what the extraction says.

In the contractor's month, 41 documents produce 4 exceptions. Two are duplicates the client uploaded twice, one is a bill whose lines add up to $412.80 against a stated total of $421.80 (a transposed digit on the image, which the extraction faithfully copied), and one is a new supplier. After clearing them, 39 bills post. The bookkeeper also opens six documents at random against their extractions and finds all six correct, which is the evidence for trusting the reading on the other 33.

Stage 3: coding and reconciliation, sampled by risk

The ledger now offers to do the matching. Xero's pricing page lists auto-reconcile as an optional $5 a month on its Early plan and as included on the plans above, and Xero presents JAX as its AI finance partner. QuickBooks lists Intuit Intelligence on its paid plans, with Accounting AI from Essentials and transactions auto-categorised and verified on Plus. Xero JAX bank reconciliation, set up and checked covers the settings for the Xero route.

The sampling rule matters more than the tool. Accept low-risk suggestions without opening them, and open the rest:

• Every suggestion at or above $200.
• Every line coded to suspense, uncategorised or an owner's account.
• A random 15 of the remaining matched lines.
• Every line where the same amount appears on two open bills.

The contractor has 118 bank lines. The rules and auto-matching clear 84 (71%), leaving 34 for hand work: 22 matched to bills, 9 uncoded and 3 for a client query. The 9 uncoded lines total $2,140.00 and sit in suspense. Reviewing the 84 means 11 lines at or above $200 plus 15 random ones, 26 lines in all, about a third. The sample finds two miscodings: a $310 hardware-store purchase for personal use and a $180 streaming subscription, both posted to materials, $490 together. The client's answers clear the suspense: six lines, $1,380, go to materials, two lines, $520, to subcontractors, and one line, $240, to the owner's drawings. That is 1,380 + 520 + 240 = $2,140, so suspense ends at zero.

Then the control that nothing else replaces: the ledger balance must equal the statement's closing balance. In this month it was $64.00 too high until the bookkeeper sorted the bank lines by amount and found one imported twice. How accountants use AI to spot errors in client books shows further patterns worth sampling for.

Stage 4: anomaly review against prior months

The senior reviewer works from account totals, not transactions. Export each account's total for the month and its average over the last three, and paste only account names and amounts: no suppliers, no client name. Use an assistant on a plan whose data terms the practice has accepted; whether it is safe for an accountant to use ChatGPT with client data sets out the questions to settle first.

You are reviewing one month of account totals for a small trade business.
Columns: account, this_month, average_last_3_months.

Flag an account only if the difference is more than 25% AND more than
150 in absolute terms. For each flagged account give: difference,
percentage difference, and two or three questions I should ask the
client. Ask questions only. Do not state causes, because you cannot
know them. List the accounts you did not flag. Show your arithmetic.

Materials, 9880, 9300
Subcontractors, 4200, 3900
Fuel, 2010, 1225
Subscriptions, 412, 243
Insurance, 0, 380
Vehicle repairs, 760, 790

An illustrative result:

Not flagged: materials (+580, +6.2%), subcontractors (+300, +7.7%) and vehicle repairs (-30, -3.8%). Check the arithmetic yourself: 2,010 − 1,225 = 785 and 785 ÷ 1,225 is 64.1%. The usual defect is a first draft that adds a cause, such as "fuel prices rose this quarter". Strike it and keep the question, because an unsupported cause in a review note becomes the explanation the client hears first.

Stage 5: partner review and the client pack

The partner should open a file that is ready to judge, not to repair. The checklist below is filled in for the contractor's month; the practice keeps one blank version and the filled version goes on file.

Only the judgement row is the partner's own work, and it is a short list, which is the point. The commentary for the client pack is a place where an assistant earns its keep, but only from figures you hand it:

Write 120 words of plain commentary for the owner of a small trade
business, using only the figures below. Do not explain causes beyond
those I give. Do not mention tax. Use these headings: Cash, Costs to
watch, Action for you.
Cash: closing bank balance 18,642.17, up 4,310.00 on last month.
Costs to watch: fuel up 785 on the three-month average because of a
second van; one subscription charged twice, refund requested.
Action for you: confirm the insurance renewal date.

The draft will read well, which is the danger. Tie every number in it to the report before it goes out, and delete any sentence that offers a forecast or an opinion you did not give it. The partner signs the pack, so the partner's check is the one that counts.

One month, hour by hour

Here is the contractor's month in effort, before the workflow and after. The before figures are what an unstructured month tends to cost a practice that keys most documents by hand; the after figures include every check above. Both are illustrations.

At a blended internal cost of $55 an hour, 3.6 hours is $198 per client per month. Across 20 clients that is 72 hours and $3,960 a month. Subtract whatever the capture tool charges per client, and the assistant plan: ChatGPT Business Standard seats list at $20 each a month billed annually with a two-seat minimum, so $40 a month covers two people. The saving is real but it is not free, and a practice with five clients should price the setup time before assuming it pays.

A three-client pilot before you roll it out

Run the new workflow on three clients before the whole book: one simple (under 60 bank lines), one typical (around 120) and one awkward, such as a cash-heavy business or one with dozens of suppliers. For the first month, rebuild a closed, already reviewed month with the new workflow and compare account totals with the reviewed books. Every difference above the client's variance threshold needs an explanation, and the explanations tell you which stage to tighten.

Measure four things per client: hours by role, exceptions per 100 documents, the error rate in sampled suggestions, and queries sent to the client. The error rate is errors divided by lines sampled. The contractor's 2 errors in 26 lines is 7.7%, which is too high to relax the sampling. The cause was clear, because both errors were the owner's personal spending, so the fix is a rule: suppliers that sell both business and personal goods are never auto-coded. A reasonable starting rule is to widen automation for a client only after two clean months with the error rate under 5%, and to review that number every quarter.

Where the workflow breaks

• A wrong rule that keeps repeating. Accepting an automation, such as "always code this supplier to materials", applies it to every future line. Review accepted rules each quarter, and look first at suppliers that sell both business and personal goods.
• Two open bills of the same amount. A matcher can pair a payment with the wrong one. The risk sample includes every line where the amounts collide.
• Poor photographs. A crumpled receipt read in poor light produces a plausible wrong total. Ask clients for flat, lit photos, and keep the line-sum rule.
• Cut-off. A bill dated next month but delivered this month drifts into the wrong period. The date rule in the exceptions queue is there for this.
• The sample shrinks when busy. The month the team is stretched is the month the errors arrive. Write the sampling rule into the checklist so skipping it is a visible decision.
• Client detail in a chat assistant. Keep review prompts to account totals, as in stage 4.

Putting AI in the engagement letter

Clients mind surprises more than software. One paragraph in the engagement letter covers it, and the practice's regulator or insurer may expect more, so ask them:

We use software with AI features to read documents, suggest bank-line matches and draft questions and commentary. A person at our practice reviews the output before it is relied on, and we remain responsible for the books we prepare. Your records are processed only in the tools listed in our data-handling schedule, and we do not enter client or supplier names into general-purpose chat assistants.

The paragraph is only true if the practice follows it, so write it after the rules above exist, and update the schedule when a tool is added. The accountants' AI risk checklist is a useful companion when the partner reviews that schedule.

Questions practices ask about AI in the accounting workflow

Which stage should a practice automate first?

Start with capture and the exceptions queue, because its checks are mechanical: duplicates, line totals that disagree with the invoice total, and new suppliers. Reconciliation suggestions come next, with risk-based sampling. Leave partner review and judgement items alone until the earlier stages have run clean for a few months.

How much of the AI's coding should a reviewer check?

Check every suggestion at or above a dollar threshold you set, every line coded to suspense or a personal account, and a random sample of the rest. In the worked month that meant 11 large lines plus 15 random ones, about a third of the matched lines, and it found two errors worth $490.

Can I paste client ledger data into ChatGPT?

Not as it comes out of the ledger. Use account totals without supplier or client names for review prompts, keep transaction detail inside ledger tools your clients have agreed to, and use a plan whose data terms the practice has accepted. Say in the engagement letter how you process client data.

Do clients need to know AI is involved?

Say so in the engagement letter in a sentence: which tasks use AI assistance, that a person reviews the output, and that the practice remains responsible for the books. Clients mind surprises more than the technology, and your regulator or insurer may have its own expectations.

What does this cost per client?

Xero lists Smart Document Capture on every plan, so Xero clients need no extra capture tool, and its auto-reconcile is an optional $5 a month on the entry plan. Dext's practice plans start from $17.70 or $19.20 per client per month with a ten-client minimum. Add an assistant plan, which lists at $20 a seat a month billed annually.

Further reads

• Claude for Bookkeepers: Month-End Checks and Client Queries — Month-end checks and client queries written as prompts.
• ChatGPT for Accounting With Excel: Seven Tasks, Each With Its Check — Seven spreadsheet tasks, each with its own check.
• Claude for Excel for Accountants: Reconciliations a Manager Can Sign Off — Reconciliations a manager can sign off.
• AI Accounting in New Zealand: GST Returns, IRD and Xero Fit — A return-specific example of the review stage in Xero.
• AI Accounting in Singapore: GST, ACRA Filings and Software Choices — How to use an AI assistant for GST F5 coding checks, a filing calendar and software questions, while a named person files with IRAS and ACRA.
• Does ChatGPT Charge VAT? Reverse Charge Steps for UK Firms — A practical invoice-first guide to ChatGPT VAT, reverse charge entries, app-store receipts and accounting checks.
• 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, QuickBooks Online pricing page, and the Dext home page with its business and partner pricing information, all read in November 2026.