AI in Accounting Examples: Six Worked Jobs for a Small Practice

A small accounting practice can use AI today for six jobs: capturing receipts and bills, suggesting bank reconciliation matches, drafting month-end commentary, chasing missing records, answering routine client questions, and turning calls into file notes. Each works as a first draft that a person reviews before it reaches the ledger or the client.

The pattern is the same every time: the software or assistant does the typing and the first pass, a person checks a defined sample, and the judgement stays with the practice. In a five-person practice with 40 monthly bookkeeping clients, for instance, capture alone can take a document-heavy client from about 135 minutes of keying a month to roughly 50 minutes of checking. Every figure here is illustrative, so swap in your own.

Follow me on Instagram@sagnikteaches

A five-person practice with 40 bookkeeping clients

Picture a five-person practice: two partners, two bookkeepers and one part-timer. It looks after 40 clients on monthly bookkeeping, mostly small trading businesses, with year-end accounts and tax work on top. Most clients use Xero, a handful use QuickBooks Online, and about half send receipts through a capture tool such as Dext or the capture feature inside the ledger.

Connect on LinkedInSagnik Bhattacharya

The bookkeepers spend roughly 70 minutes per client per month on routine processing: keying or checking bills, clearing bank feed lines, chasing what is missing, and writing a note for the partner. That is about 47 hours a month across 40 clients. The six jobs below each target a slice of that time. The figures are illustrative, so time a few real clients before you plan anything.

Subscribe on YouTube@sagnikteaches

One rule applies throughout. Nothing the AI produces goes to a client, a ledger or a filing without a named person having looked at it. If you want the wider list of tools and what they cost, the tool roundup for small accounting firms covers that.

Job 1: Capturing receipts and supplier bills

This is the oldest AI job in accounting, and still the safest to start with. A tool reads a photographed receipt or an emailed supplier bill, pulls out the supplier, date, total and tax, suggests an account code, and creates a draft transaction. Dext is a widely used standalone tool. Xero launched its own Smart Document Capture on 2 July 2026, and Xero's pricing page lists the feature on all three plans (Early, Growing and Established) with no add-on charge; on 18 September 2026 Xero renamed its Files area Documents. Xero's own posts disagree on tax: the September post says capture pulls the local tax rate from uploaded documents, while the July launch post, updated on 28 October, still lists tax extraction under "coming next". Test it on a dozen real bills from your clients before you rely on the tax field.

Dext's plan for businesses is listed at $302.50 billed annually, shown as $25.21 a month, for 250 documents a month and five users; its AI Assist add-on is priced separately, at $10.50 a month until 31 January 2027. Whether that is worth paying depends on how many documents your clients send, and whether the ledger's built-in capture already does enough.

The plumbing supplier's 90 documents a month

Take one client, a small plumbing supplier, who sends 90 documents a month by email and phone photo. Keying each one takes about 90 seconds, so 135 minutes. With capture, the bookkeeper reviews drafts at about 25 seconds each, 38 minutes, and fixes perhaps 12 of them for another 12 minutes. The month drops from 135 minutes to roughly 50. That is 85 minutes saved on one client.

Checking capture drafts before approval

Do not approve in bulk. Sort drafts by supplier and look at three things on each: total against the document, the tax amount, and the account code. Capture tools are good at totals and weak at two cases, a delivery note that looks like an invoice and a credit note scanned as a positive bill. Add a monthly rule: open the 10 highest-value documents and check them against the images.

A realistic failure: a supplier changes its invoice layout, and the tool begins reading the account number as the invoice number. Every bill from that supplier then looks like a duplicate or goes unmatched. You spot it because the same odd reference appears on several lines. Fix it once in the supplier rule and tell the tool's support.

Job 2: Suggested matches for bank reconciliation

Reconciliation is where most practices spend their routine hours. In Xero, JAX automatic reconciliation clears a bank line by itself when it is highly confident and suggests a match for the rest. Xero's reconciliation page places it on Growing and above, and the pricing page also offers it on Early as an optional $5.00 a month add-on. Every reconciled line carries a label, Rule, Match, Memory, Prediction or Manual, so you can see why it was matched, and a subscriber or admin switches the feature on or off for each bank account. QuickBooks Online takes a different route. Intuit's help centre describes an Accounting Agent whose Accounting AI, on Essentials, Plus and Advanced, gathers confident categorisations into a Ready to post list governed by a Transaction posting policy you set.

The difference matters for review. Apart from a few transaction types it can post automatically, such as payroll run through Intuit's own payroll and bills paid with QuickBooks Bill Pay, QuickBooks waits for a person to select Post, and Intuit says AI never causes your bank rules to auto-post a transaction. Xero reconciles high-confidence lines without waiting, so your check happens after the event; you can change the account code or remove and redo any line. Switch it on for one well-behaved bank account first, not the whole client file.

A café's 320 bank lines

A café client has 320 bank lines in a month. Before automation, a bookkeeper clears about 60 per cent (192 lines) through bank rules and codes the rest by hand: 128 lines at 40 seconds is about 85 minutes. Suppose JAX reconciles 70 of those 128 on its own and suggests matches for another 30. Checking an auto-reconciled line by its label takes about 8 seconds and accepting a suggestion about 12, roughly 15 minutes together, plus 28 manual lines at 40 seconds, about 19 minutes. The total is near 34 minutes against 85.

Reviewing reconciled lines by label

Check by label, not by line order. Rule and Match lines need a quick scan; read every Memory and Prediction line properly, because those rest on past patterns, your own or other Xero users', rather than on a bank rule or an existing record. Then run a spot check on anything over $1,000 and anything involving a director, a loan or a tax payment, because those carry the greatest cost if miscoded. A sample bookkeeper note might read:

Reconciled 318 of 320 lines. 192 by existing rules, 70 auto-reconciled by JAX, 30 accepted from suggestions, 26 coded manually.
Held back: 2 lines over $1,000 (equipment deposit, unknown payee) - queried with client.
Checked: every Memory and Prediction line, plus 15 random Match lines. 1 auto-reconciled line recoded (card fee coded as supplies).

If you are weighing which ledger to standardise on, the vendors' pricing pages in November 2026 listed QuickBooks Online at $38, $85, $140 and $340 a month across its four paid tiers (Simple Start, Essentials, Plus and Advanced, alongside a free plan), and Xero at $27, $59 and $97. Ready to post needs Essentials or above. Both change, so check before quoting a client.

Job 3: Draft month-end commentary for the partner

Partners rarely want a data dump; they want three sentences on what moved and why. A general assistant such as ChatGPT, Claude or Copilot is good at turning a summary table into a plain-English paragraph, as long as you give it the numbers and tell it not to invent causes.

A landscaping client's jump in fuel costs

Say a small landscaping client shows revenue of $41,200 against $36,800 last month, up 12 per cent, and fuel costs of $3,150 against $2,300, up 37 per cent. Paste a clean, anonymised table of the ten accounts that moved most, and use a prompt like this.

You are helping a bookkeeper write a month-end note for a partner.
Below is a table of the ten accounts with the largest movement against last month.
Write: (1) a three-sentence summary, (2) a list of movements above 10% with the size in dollars,
(3) a list of questions to ask the client, one per large movement.
Do not state a reason for any movement unless it appears in the data. Where the cause is unknown, say "cause not in the data".
[table pasted here]

A realistic output (illustrative) would say that revenue rose $4,400, or 12%, that fuel rose $850, or 37%, with the cause not in the data, and ask the client whether a second vehicle was added, whether the fuel card changed supplier, or whether a refund posted last month distorts the comparison. That final line, the questions, is where the value is. The AI cannot know why fuel rose, but it will not forget to ask.

Recalculating before the partner reads it

Recalculate every percentage and dollar movement yourself or in a spreadsheet. Assistants are fluent but can slip on arithmetic, and a wrong figure in a partner note erodes trust quickly. Delete any sentence that offers a reason you did not supply. If you do this in a spreadsheet, the tutorial on Copilot for data analysis shows how to have the formulas do the sums so the assistant only writes the words.

Job 4: Chasing missing records

Missing receipts, unanswered queries and absent statements are the biggest drag on a monthly cycle, and the least skilled. The job is to write polite, specific chasers in the client's tone and to keep a list of who owes what. There is a dedicated tutorial on chasing missing client records with AI before deadlines; here is the short version with an example.

Fourteen clients and 61 missing items

At month end, 14 of the 40 clients have unexplained transactions or missing documents, 61 items in total. A bookkeeper pastes each client's item list (date, amount, payee, no names beyond the business) into the assistant and asks for one message per client. Writing 14 messages by hand takes about 6 minutes each, 84 minutes. With drafts, review and light edits take about 2 minutes each, 28 minutes. That frees 56 minutes monthly.

Write a short, friendly email to a client asking about the items below.
Group them under "Receipts we need" and "Payments we cannot identify".
Give a reply-by date of the 8th. Ask them to reply to this email with the answers or a photo.
Keep it under 150 words. Sign off as [your name]. Do not mention penalties.
Items: 12 Mar $212.40 fuel card; 15 Mar $1,800 transfer to unknown payee; ...

The sample output opens with a greeting using [first name], lists the items in two short groups, and asks for replies by the 8th. Check that the item amounts match your list, since an assistant can drop a line from a long list. Also check the tone against your relationship with that client, and never let it threaten a penalty you have not decided to apply.

Job 5: First drafts of replies to routine client questions

Clients ask the same questions: when will my accounts be ready, can I claim this, why is my balance different from my bank. The assistant drafts a reply from the facts you give it and your usual wording. The point is not to answer tax questions automatically; it is to produce a draft that the person with the knowledge corrects and sends.

A $460 gap between bank and ledger

A client writes: "My bank says $18,420 but your report says $17,960. Which is right?" The bookkeeper gives the assistant the reconciliation summary: bank balance $18,420, outstanding payments of $460, and a statement balance matching the ledger. The draft explains that the gap is $460 of cheques issued but not yet cleared, lists them, and says the ledger figure will match the bank once they clear. The bookkeeper confirms the two cheques, adds a sentence of reassurance and sends it. Time: 4 minutes rather than 12.

Keeping tax answers off the letterhead

Never let the assistant answer a tax or legal question from its own knowledge on your letterhead. Give it the facts and the position you have already decided, and ask only for the wording. If a client question is really advice, route it to a partner. Writing this rule down in your engagement letter wording for AI use also helps set client expectations about what AI touches.

Job 6: Turning a client call into a file note and action list

Partners finish a call and forget to write it up. A meeting recorder or a pasted transcript can produce a file note, decisions and follow-up tasks. The tools range from the Otter and Fathom type of recorder to the meeting summaries in Teams or Zoom; Microsoft's intelligent recap in Teams needs a Teams Premium or Microsoft Copilot licence, so check what you already have. Always tell the client you are recording, and keep recordings under the same retention rules as other client records.

A 30-minute year-end planning call

A 30-minute year-end planning call with a tutoring-centre owner produces a transcript. Writing the note by hand takes 15 minutes. The assistant drafts a note in about a minute; the partner spends 5 minutes correcting it. The sample note reads (illustrative):

Call: year-end planning, 30 mins. Present: [client first name], [partner name].
Decisions: client will buy a laptop before year end; partner to confirm the correct treatment before purchase.
Actions: client sends mileage log by the 20th; practice sends draft accounts by the 30th.
Open question: whether to move to monthly management accounts - client to decide after seeing the draft.

The review step is simple and strict: check every decision and every date against what you remember or the recording. Summaries sometimes turn a maybe into a decision. If a note records advice you gave, make sure it says precisely what you said, and file it with the matter.

What went wrong in the first two months (an illustrative case)

Suppose the practice above rolled out all six jobs in a single month. Three things would probably go wrong, and each is common. First, the bookkeepers would approve capture drafts in bulk during a busy week, and a duplicated supplier bill would slip through, found only when a client queried their payables. Second, the commentary job would produce a confident sentence blaming a sales rise on a seasonal promotion the client never ran. Third, a chaser would go out listing 11 items when 12 were missing.

None of these is dramatic, and all come from skipping the review step. The fix is to roll out one job a fortnight, give each a named owner, and write the spot check into the month-end checklist: top 10 documents, every Prediction line, figures recalculated, item counts matched. For the wider set of ways AI can damage client trust, see the tutorial on AI mistakes that put client trust at risk.

Putting the six jobs on a timetable

Do not start with the most impressive job. Order them by how easy a mistake is to catch. Capture and chasers come first because every output is checked line by line against a source. Call notes and client replies come second, because a person reads them before they go anywhere. Reconciliation suggestions and commentary come last, because errors there are subtle and live inside the numbers.

Add a standing check at the end of six weeks. Compare the actual minutes per client against your pre-rollout timing, count the corrections your team made, and ask each bookkeeper which job they would stop. A job that saves ten minutes but generates three corrections is not saving much. Use the saving to fund advisory work clients pay for and to price the service line, a subject covered in the tutorial on pricing AI bookkeeping as a service line.

A second client: the same jobs on a different shape of work

The plumbing supplier and the café are document-heavy and transaction-heavy. A different client shows where the jobs help less. Take a one-person yoga studio with 35 bank lines a month, six supplier bills and a handful of card payments. Capture saves perhaps ten minutes, reconciliation suggestions save five, and the commentary is two lines. The saving is real but small, and it may not justify a paid capture subscription for that client alone.

The lesson is to switch jobs on per client, not per practice. A simple rule works: turn capture on for any client who sends more than 30 documents a month, turn suggested reconciliation on for any client with more than 150 bank lines, and keep the others on bank rules and a short manual check. Review the list every quarter, because clients grow. Revisit it too whenever a vendor changes a plan, since an add-on that made sense for three heavy clients may cost more than it saves once the heaviest of them leaves.

Log the actual minutes for two or three clients for a month before and after, in a plain spreadsheet with one row per client per job. That log is what lets a partner say the rollout paid for itself, and what you show a client who asks whether AI will lower their fee.

Checking the claims your software makes

Product names and plan boundaries move quickly, and a vendor's own pages do not always agree. In November 2026 Xero's reconciliation page put automatic reconciliation on Growing and above while its pricing page also sold it to Early users as an add-on, and its two capture posts disagreed on whether tax extraction was live. Intuit's AI features also step up by plan: Intuit Intelligence appears from Simple Start upwards, Accounting AI from Essentials and AI-assisted reconciliation from Plus. So before you promise a client anything, open the vendor's page, read what is live on your plan today, and note the date you checked.

The same caution applies to costs. Quote list prices in USD only from the vendor's own pricing page, and remember that promotional discounts, such as Xero's offer of 80 per cent off the first three months or QuickBooks' half price for three months, end. Build your own business case on the full price.

If you would rather map this with someone, that is what my AI implementation consultation covers: finding the jobs where AI pays off in your practice and handing over a process the team can run.

Questions practices ask before trying these

Which of the six jobs should a practice try first?

Start with document capture or missing-records chasers. Both have a clear before and after, the output is easy to check line by line, and a mistake is caught before anything reaches a client or a filing. Leave reconciliation suggestions and review commentary until your team trusts the review habit.

Do these tools replace the bookkeeper?

No. Each example moves typing, matching and first drafts to the software, and keeps coding decisions, judgement calls and client conversations with a person. The saving shows up as fewer keystrokes per client, which most practices spend on review and advisory work instead.

Is client data safe in a chat assistant?

It depends on the plan and your engagement terms. Business plans such as ChatGPT Business and Claude Team do not train on your content by default, but free personal accounts may differ. Check the vendor's data terms, tell clients in your engagement letter, and strip names and account numbers where the job does not need them.

What does this cost for a ten-client practice?

Software you already run often includes the features. Dext lists $302.50 a year for 250 documents a month and five users, and a standard seat on ChatGPT Business or Claude Team is $20 a month billed annually or $25 monthly, with at least two seats. Check each vendor's current page, because plans change.

Further reads

• How Small Accounting Firms Use AI Day to Day: Real Examples — More day-to-day examples from small firms.
• How Much Does AI Cost a Small Accounting Firm? — Itemised monthly and one-off costs for a small firm.
• What an AI Implementation Looks Like in a Small Accounting Firm — How a whole implementation unfolds in a small firm.
• How to Automate Client Onboarding in a Small Accounting Practice — Apply the same approach to onboarding new clients.
• AI Receptionist for Accounting Firms: Tax-Season Calls and Cost — Handle tax-season calls with an AI receptionist.
• Why AI Adoption Stalls in Accounting Firms, and How to Restart It — Why teams stop using these tools, and the restart.
• 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, bank reconciliation and JAX pages and Xero product-update posts of 2 July and 18 September 2026; Dext business pricing; QuickBooks pricing and Intuit help articles on Accounting Agent features and Ready to post; Microsoft Learn on Teams intelligent recap; ChatGPT Business and Claude pricing pages (all checked 4 November 2026).