(157 chars): Bookkeepers can use AI for capture, categorisation and reconciliation. Intuit's own research puts categorisation accuracy at 68.67%, so review

How bookkeepers can use AI safely in 2026

Last verified: 2026-07-28

How bookkeepers can use AI in 2026 comes down to four tasks: pulling data off receipts and invoices, suggesting categories for bank transactions, matching payments to invoices, and drafting reports. The judgment work stays human. Measured accuracy is well below what the marketing says, and Intuit’s own research team puts its best transaction categorisation model at 68.67% top-1 accuracy. Reviewing that output is now the main part of the job.

This article sets out how bookkeepers can use AI in 2026: which tasks to hand over, what it gets wrong, how to review it, what the rules require, how to work inside a client’s own file, and why the role is changing rather than disappearing.

Most accountants have used AI, and few trust it to work unsupervised. In Intuit’s 2026 Accountant Technology Survey of 725 US accounting professionals, 88% had used AI for at least one client service in the past year. Only 6% wanted AI to execute work autonomously.


The job is shrinking, not disappearing. The US Bureau of Labor Statistics projects bookkeeping, accounting and auditing clerk roles to fall 6% between 2024 and 2034, a loss of 94,300 positions. It also projects about 170,000 openings in the occupation every year over the same decade.



The three kinds of AI in bookkeeping

Three different things get called AI in bookkeeping, and they behave differently on cost, control and confidentiality. The first is built into the ledger you already use. The second is a separate platform you buy and connect. The third is a general chatbot like ChatGPT or Claude.

This is a new layer on old technology. Bookkeeping software has read documents since optical character recognition arrived, and category rules have existed for as long as bank feeds. What changed in 2025 and 2026 is that vendors moved from rules you write to models that predict, and then to agents that act without being asked each time.

AI built into QuickBooks and Xero

The AI built into QuickBooks and Xero arrives with the subscription, whether you ask for it or not. Intuit runs eight named agents in QuickBooks Online, including Accounting AI for categorisation, Payments AI, Sales Tax AI and Business Tax AI. The chat assistant, Intuit Intelligence, comes with every tier, subject to a monthly prompt cap, with 100 extra prompts available for $10 a month.

Which agents your client has depends on what they pay. Prices below are Intuit’s US list rates, checked on 28 July 2026, with rupee equivalents at 95.94 to the dollar.

QuickBooks tier Monthly (USD) Approx. INR AI you get
Simple Start $38 ₹3,645 Intuit Intelligence chat
Essentials $75 ₹7,195 Adds Accounting AI, Payments AI
Plus $115 ₹11,030 Adds Sales Tax AI, Customer AI, AI-powered reconciliation
Advanced $275 ₹26,385 Adds Finance AI, Project Management AI

One timing point worth knowing before you quote a client. Intuit announced on 27 June 2026 that monthly prices for Essentials, Plus and Advanced change for renewals on or after 1 August 2026, while Free, Lite, Ledger and Simple Start stay the same. Intuit has not published the new figures, so check the current rate rather than quoting the table above after that date.

Xero’s assistant, JAX, is priced differently. Xero says JAX chat is available to all Xero subscribers and users, with currently no additional charge, though it reserves the right to change that. Access follows the user permissions already set in the file.

The US plans are Early at $25 a month (about ₹2,400), Growing at $55 (₹5,275) and Established at $90 (₹8,635). Smart Document Capture, which reads source documents and extracts the data into Xero, is available now on all three. Automatic bank reconciliation is in beta on Growing and above and is opt-in rather than on by default, so check whether the client has switched it on before you assume it is running. Bill Protection, which inspects bills before payment for altered bank details and unfamiliar suppliers, is in beta.

If you are working in Xero rather than QuickBooks, you are less likely to be blocked from a feature by the client’s plan tier.

Purpose-built AI bookkeeping platforms

Purpose-built platforms sit on top of the ledger and take over a defined slice of work. Dext extracts data from documents and offers AI Assist for categorisation. Ramp automates approvals, codes each line of an invoice, and syncs to the ledger rather than replacing it. Digits and Zeni run close automation with varying degrees of human sign-off.

Platform US price Approx. INR What the AI does
Dext $17.70 per client per month (Essentials) or $19.20 (Advanced), 10-client minimum; AI Assist adds $7 per client ₹1,700 to ₹1,840, plus ₹670 Document extraction, categorisation
Docyt Plans start at $299 a month, then quoted on a demo From ₹28,685 Revenue reconciliation, expense management
Zeni $494 on annual billing, $549 month to month ₹47,395 to ₹52,670 Categorisation with reasoning shown, receipt matching

One of these is not like the others. Puzzle is an AI-native general ledger rather than an add-on: its AI sits inside the ledger itself, and it markets a direct switch from QuickBooks with free migration. Moving a client onto it is a migration decision, not a tool purchase.

One naming point: the tool many bookkeepers know as Keeper rebranded to Double on 23 October 2025 and now sits at doublehq.com.

Advertisement

General chatbots like ChatGPT and Claude

General chatbots are the only kind where you decide what data you send. ChatGPT and Claude do not connect to the ledger unless you connect them. They are useful for explaining a treatment, drafting a client email, checking a reconciliation approach, or turning a messy note into a clean journal description.

They are also the riskiest of the three, because of data handling rather than accuracy. On the consumer tiers of both products, your content may be used to train the model by default. The client data section below sets out what that means for what you are allowed to paste.


The three kinds of AI in bookkeeping
KindWho paysWho controls itWhat data leavesWhat breaks
Built into QuickBooks or XeroThe subscription holder, usually the clientThe plan tier. Individual features cannot be turned off in QuickBooks OnlineNothing extra. The ledger already holds itCategorisation on unusual transactions, rule drift, add against match
Purpose-built platformYou or the client, per client per monthWhoever bought it. Needs app install rights on the ledgerLedger and document data, under a data processing agreementSync errors, duplicated entries, coverage gaps between the tool and the ledger
General chatbotYou, per seatYou, entirely. Including what you pasteExactly what you type. Consumer tiers may train on itConfident invention, and confidentiality if nothing was redacted
The three are not alternatives. Most bookkeepers touch all three in a week, and each one puts client data in a different place.
Skill Arbitrage

Which bookkeeping tasks AI can take over

AI can take over the mechanical parts of the bookkeeping cycle: reading documents, proposing categories, and matching payments to open items. It cannot take over decisions about how a transaction should be treated. The line between those two is not always obvious, which is why the accuracy numbers matter more than the feature list.

If you are starting out in US bookkeeping, none of these tools remove the need to understand the underlying treatment. You still need to know the correct treatment. AI just produces more entries to check.

Document capture and data entry

Document capture is the most reliable AI task in bookkeeping. The software reads a receipt, invoice or statement and extracts supplier, date, each line of the invoice, tax and total into structured fields. Dext claims 99.9% data extraction accuracy. BILL claims 99% on key invoice fields, though the same page also cites 95% day-one accuracy on the same task, and neither figure carries a substantiating footnote.

Those are vendor claims, published by the vendors, on their own test sets. They are plausible for clean, typed, single-page documents. They tell you little about a photographed receipt with a fold across the total.

Transaction categorisation

Categorisation is where AI is most heavily marketed and least reliable. The software looks at a bank feed line, compares it to how similar lines were treated before, and proposes an account. It improves as it sees more of a client’s history, and it is often wrong on anything unusual.

Both platforms learn from what you do, but not in the same way. Intuit says plainly that if you change a category, “QuickBooks will learn from your changes.” Xero splits it: reconciliations tagged Memory come from how you have treated similar transactions in that file, while those tagged Prediction come from a model trained across all Xero users, which your corrections do not retrain.

The mechanism runs both ways. Repeatedly accepting a wrong suggestion trains the file just as effectively as correcting one, which is why a rushed first month stays expensive for a year.

Bank reconciliation and payment matching

Reconciliation suits AI because it is a matching problem. The software proposes which bank line corresponds to which invoice or bill, and it is right most of the time on straightforward one-to-one matches.

The failure cases are partial payments, payments covering several invoices, foreign currency settlements, and anything involving a clearing account. Those are the cases where a wrong match is hardest to spot later.

Measured accuracy against vendor claims

Measured accuracy is far lower than the marketing implies, and the best evidence comes from the vendors’ own research teams.

Source Figure What it measures
Intuit and University of Notre Dame, arXiv:2506.09234 62.49% Production baseline, top-1, few-shot
Same paper, proposed model 68.67% Top-1 (88.04% top-5)
University of Warwick and SME Capital, arXiv:2508.05425 73.49% Overall, UK Open Banking data
Same study, high-confidence subset only 90.36% Predictions above 0.8 confidence
Digits, self-run benchmark 97.8% Vendor claim, 2,000 transactions
DualEntry Labs leaderboard 83.2% Best model, vendor-run test

The Intuit paper trained on three million transactions across 100,000 categories and 15,000 companies, so it is a harder task than a small business chart of accounts and the figure understates real-world performance on a simple file. The Warwick study’s 90.36% is a filtered subset tested on a single firm, with a confidence interval that reaches down to roughly 84%. Its authors concluded the technology suits semi-automated workflows with human review on low-confidence output.

Digits’ 97.8% came from its own benchmark against outsourced human accountants it scored at 79.1%. DualEntry, running a different test set, has had the best model on its public leaderboard rise from 66.0% at launch to 83.2%. One vendor scoring its own agent at 97.8% while another scores the field in the sixties and eighties is a good reason to trust neither number and build a review step instead.

Treat the research figures as an upper bound on error rather than an expectation. At 70% accuracy on a client with 1,000 monthly transactions, roughly 300 entries need correcting. At 90% it is still 100.


AI categorisation accuracy: measured against claimed
Measured in published research
Production baseline, few-shot62.49%
Intuit and University of Notre Dame, arXiv:2506.09234
Best model in the same paper68.67%
Intuit and University of Notre Dame, arXiv:2506.09234
UK SME bank transactions, overall73.49%
arXiv:2508.05425, August 2025
Same study, high-confidence predictions only90.36%
arXiv:2508.05425, August 2025
Claimed by vendors
Digits, self-run benchmark on 2,000 transactions97.8%
Vendor claim
Dext, document extraction99.5%
Vendor claim
BILL, key invoice fieldsabout 99%
Vendor claim
DualEntry Labs, best of 19 models on 101 tasks66.0%
Vendor-run test, contradicts the figures above
At 70% accuracy, a client with 1,000 monthly transactions needs about 300 corrections. At 90% it is still 100. Neither supports leaving the ledger unattended.
Skill Arbitrage

What AI gets wrong on bank feeds

AI gets bank feeds wrong in patterns, not at random, so you can learn where to look. Four patterns account for most of the damage.

Practitioners have documented these in public. One QuickBooks Community thread, titled “QBO Suggested Categorization broken”, reports suggestions being wrong almost every time, with credit card transactions assigned to another credit card account. Another describes whole batches of unselected transactions being recategorised during tax preparation.

Rule drift

Rule drift is the hardest failure to notice and usually the most expensive. QuickBooks and Xero bank rules fire on a text match against the description or payee. A supplier whose bank descriptor changes by a word stops triggering the rule and reverts to whatever the model guesses. A rule written too broadly keeps catching transactions it should not.

Two settings make it worse. A QuickBooks rule with auto-add enabled posts straight into the register without appearing in your review queue, so nobody sees those transactions at all. And rules only act on items still awaiting review, so correcting a rule fixes nothing it has already posted.

By the time someone notices, the correction spans several closed periods and possibly a filed return. Check any rule you changed in the last month, any rule that suddenly stopped firing, and any rule you did not write that has auto-add switched on.

Add versus match

Add and match are two buttons that do opposite things, and AI confuses them in both directions. Pressing Add on a bank line already entered creates a second copy of the same transaction. Pressing Find match and ticking an unrelated open item settles a record that was never paid, while the transaction that actually happened is never recorded.

A duplicate inflates the expense or income and usually surfaces at reconciliation, because the extra copy stays uncleared. A false match never surfaces. The bank line is cleared, the reconciliation ties to the penny, and the damage sits in the subledger as an aged invoice showing as paid.

QuickBooks also offers Record as transfer as a third option, which is the correct answer more often than either of the other two.

Revenue recognition

Revenue recognition is where AI is confidently and structurally wrong. AI books cash as revenue on the date it lands, because that is what the bank feed shows. For an accrual-basis business with multi-month contracts, that overstates revenue in month one and understates it across the rest of the term.

A hosted software subscription billed annually in advance is the clearest case. The cash arrives in January, and under ASC 606 the revenue is recognised across twelve months with the unearned portion sitting as a contract liability.

Two caveats stop that being a rule you apply blind. A term licence the customer installs and runs on its own hardware is generally recognised at a point in time, so the same bank line can be right exactly as booked. And ASC 606 only binds if the client reports on US GAAP: a large share of small bookkeeping clients report on a cash or tax basis, where the January receipt genuinely is January revenue. Establish the basis before you call anything an error.

Judgment calls AI cannot make

The judgment calls AI cannot make start with capitalise against expense, and the first question is which rule applies. If a contractor worked on an asset the client already owns, the test is what the work did to the asset under Treasury Regulation section 1.263(a)-3(d), which is not in the transaction description. If the payment bought a new item, it is an acquisition and the improvement rules never come into it.

Three safe harbours decide most of these and none is visible in a bank feed. The de minimis safe harbour lets a taxpayer expense items up to $2,500 per invoice or item, or $5,000 with an applicable financial statement, but only where a written capitalisation policy was in place at the start of the year. The routine maintenance and small taxpayer safe harbours cover much of the rest. Whether the client made the election is a fact about their tax return, not about the transaction.

Depreciation sits on top of that and is a book against tax difference rather than a coding decision. Section 179 is an affirmative election the client makes on Form 4562. Bonus depreciation under section 168(k) works the other way: it applies automatically to qualifying property unless the client elects out for that class, so the default is that it has already been claimed.

The same applies to intercompany transfers, which AI routinely books as revenue or expense in one entity because the other side is not in the file it can see. The correction is not automatic either, because money moving between commonly owned entities can be a loan recorded as due to or due from, a capital contribution or distribution, or a genuine management fee that is real revenue in the standalone books. E-commerce clients bring their own version, because sales tax nexus by state determines the treatment and no bank line carries it.


How to review AI’s work

Reviewing AI’s work means checking in a fixed order rather than reading every line. Reading every line defeats the point of the automation.

Sort by risk rather than by date. AI is most reliable on small, repeated, familiar transactions and least reliable on large, new or unusual ones. Reviewing in date order checks the safest transactions first.

The weekly pass

The weekly pass takes about twenty minutes per client. On a client with 400 categorised transactions waiting on a Monday, four filters cover most of the risk.

Sort by absolute value, largest first, so refunds and credits do not sort to the bottom, and check the top twenty regardless of what the software thinks of them. Then filter to payees appearing for the first time, because a new payee has no history and the model is working from the description string alone. Then check anything sitting under a rule you changed this month.

Neither platform shows you a confidence score, so the fourth filter is indirect. QuickBooks groups its highest-confidence suggestions into a Ready to post batch on Essentials and above, which means the useful signal is whatever falls outside it. Xero’s JAX auto-reconciles only where it is confident and leaves the rest as suggestions, and on the lines it did reconcile the Method column tells you whether the decision came from a rule, a document match, your own history, or a prediction drawn from other Xero users. Treat Prediction as the one to check.

The month-end pass

The month-end pass looks for pattern damage rather than individual errors. Run a profit and loss comparison against the prior month and investigate any account that moved more than about 20% and more than a dollar figure you set for that client, so a 20% swing on a $50 account does not outrank a 5% swing on the largest expense line. Then run the same comparison on the balance sheet, because most of the failures above land there rather than in the profit and loss.

Clear the suspense accounts to zero: Uncategorised Income, Uncategorised Expense, Uncategorised Asset and Ask My Accountant. These are where everything AI could not place ends up, and they are the single most useful number in the file.

Confirm Undeposited Funds holds only genuinely undeposited items. A deposit added straight to income while the invoices behind it sit in Undeposited Funds counts the same revenue twice, and it is the most common bank feed error of its kind. Then agree the receivables and payables ageings to their control accounts and the loan balances to the lender’s schedule, and reconcile balance sheet accounts on the supporting detail rather than the closing balance, because a false match still balances.

Set a closing date with a password once the month is signed off, so neither the AI, a rule, nor the client can post into it.

Categories to spot-check every time

The categories to spot-check every month are revenue, fixed assets and repairs, and owner transactions.

Revenue carries the recognition problem and a second one: any client paid through Stripe, Square, PayPal or Shopify receives deposits net of processor fees, so booking the deposit as revenue understates gross revenue and drops the fee expense entirely. Fixed assets and repairs carry the capitalise-against-expense judgment. Owner transactions depend on the entity, because a draw in a sole proprietorship is an equity reduction while the same bank line in an S corporation is a distribution sitting alongside a reasonable-compensation requirement.

Also check deferred and prepaid accounts, intercompany transfers, loan payments where principal and interest split against an amortisation schedule, credit card payments where the error is booking the payment as an expense a second time instead of as a transfer against the card liability, sales tax collected which is a liability rather than revenue, payroll where the net bank draw hides gross wages and withholdings, foreign exchange gain and loss, and inventory and cost of sales for any product business.

A prompt for checking a transaction list

The prompt below checks a transaction list you have already categorised. Do not ask the chatbot to categorise. Ask it to find what looks wrong, with the client’s identity stripped out.

You are reviewing a bookkeeping categorisation for a US client on the accrual
basis. The client is an S corporation with a December year end.
Below is a list of transactions with the direction of the money and the
category each was assigned.

Flag only entries that look wrong or need a human decision. For each one, give:
the line, the reason it is doubtful, and what you would need to know to decide.

Pay particular attention to:
- amounts that suggest a capital purchase booked as an expense
- payments that look like they cover more than one period
- transfers that may be between related entities
- entries where the direction of the money contradicts the account type

Do not suggest categories for entries that look correct. Do not guess where the
description is ambiguous, say what you would need to know.

2026-06-03 | OUT | 4,120.00 | VENDOR-A equipment purchase | Repairs & maintenance
2026-06-07 | OUT | 1,000.00 | VENDOR-B annual software    | Software subscriptions
2026-06-11 | OUT | 8,500.00 | TRANSFER to ENTITY-2        | Consulting income
2026-06-19 | OUT |   240.00 | VENDOR-C monthly internet   | Utilities

A good reply flags the first three. Line 1 is equipment sitting in a repairs account, which is the wrong account whatever the tax treatment, and at $4,120 it is above the $2,500 de minimis threshold that applies without an applicable financial statement, so it is probably a fixed asset. Line 3 is wrong on its face before the related-party question arises, because a payment out coded to an income account debits Consulting income and understates June revenue by $8,500. Line 2 is the one to ask about rather than correct: strictly, eleven twelfths of a twelve-month prepayment is a prepaid asset, but $1,000 sits below the prepaid threshold most firms operate. A weak reply also rewrites line 4, a recurring monthly overhead correctly booked, and invents a reason for it.

Three things make that prompt work. It names the accounting basis and the entity type, so the model is not guessing. It shows the direction of the money, without which no category can be assessed. And it tells the model to say when it does not know, so it stops inventing answers.

The same structure works in other fields. Lawyers use role, then governing rules, then facts, then the required output format, which is why how lawyers write prompts for the same problem reads the same way.

Note what is missing from the transaction list. No client name, no bank account number, no tax identification number, no addresses.


The review pass
Weekly, about 20 minutes per client
  • Sort categorised transactions by amount, largest first. Check the top twenty regardless of confidence
  • Filter to payees appearing for the first time. No history means the model guessed from the description
  • Check everything the software marked low confidence
Month end
  • Compare profit and loss against the prior month. Investigate any account that moved more than about 20% without a known reason
  • Check any categorisation rule you changed in the last month, and any rule that stopped firing
  • Reconcile balance sheet accounts on the detail, not the balance. A false match still balances
Always, whatever the confidence score
  • Revenue, and any deferred or prepaid account
  • Fixed assets and repairs, because capitalise against expense is a judgment
  • Owner drawings and personal expenses
  • Intercompany transfers
  • Loan and credit card payments that split principal from interest
  • Foreign exchange gain and loss, and payroll clearing
Sort by risk, not by date. AI is most reliable on small repeated transactions and least reliable on large, new or unusual ones.
Skill Arbitrage

How bookkeepers can use AI on client data

Bookkeepers can use AI on client data only after deciding which tool tier they are on and what has been stripped out of the input. Some of the rules covered here carry criminal penalties. AI output is a draft, not a deliverable, and client financial data is not yours to distribute.

Data training on consumer and business tiers

The consumer tier of every major chatbot may train on what you type into it, and the business tier does not.

Tier Price Approx. INR Trains on your content
ChatGPT Plus $20/month ₹1,920 Yes, by default. Opt-out in settings
ChatGPT Business $25/user/month, $20 annual, 2 seats minimum ₹2,400 No, by default
Claude Pro $20/month, $17 annual ₹1,920 Yes, unless you opt out
Claude Team $25/user/month, $20 annual, 2 to 150 seats ₹2,400 No, by default

OpenAI’s data-use policy states that for individual services such as ChatGPT it may use your content to train its models, and that for business users it does not train on any inputs or outputs by default. Anthropic’s commercial terms draw the same line, and on the consumer side allowing training extends data retention to five years against 30 days otherwise.

The ledgers differ too. Intuit’s Global Privacy Statement, updated 9 March 2026, says it analyses customer content to train its artificial intelligence and machine learning models. Xero says its third-party LLM providers “process the data you enter, which may include personal data, but this data is not retained and isn’t used to train the LLMs.”

What to redact before pasting

Before pasting, strip the client’s legal name and trading name, all bank and credit card account numbers, the EIN or SSN, physical addresses, and any customer or employee names inside transaction descriptions. Replace them with tokens like VENDOR-A and ENTITY-2, as in the prompt above.

Amounts and dates can usually stay, because they carry the analytical content and identify nobody on their own. If a single transaction is distinctive enough to identify the business, round it or leave it out.

Redaction is an operational control, not a legal defence. Tax return information means information furnished in connection with preparing a return, and removing the client’s name does not automatically take a figure outside that definition. A connected tool under a signed data processing agreement is a controlled disclosure; pasting into a personal chat window has no contract and no record behind it.

What Circular 230 and section 7216 require

Circular 230 and section 7216 require verification and consent, and neither has an exception for general-purpose AI. In June 2026 the IRS Office of Professional Responsibility issued guidance on responsible AI use in federal tax practice stating that practitioners must thoroughly review all AI-created documents and language incorporated into writings before delivery to a client or submission to the IRS.

The bulletin routes through existing Circular 230 duties. Section 10.22 requires due diligence in preparing documents. Section 10.35 requires competence, which the bulletin reads as understanding a system’s operational mechanics, limitations and risks. Section 10.37 means you cannot rely on generative AI for written advice without independent verification. Both reliance provisions carry a detail worth noticing: the presumption of due diligence applies where a practitioner relies on the work product of another person, and a model is not a person.

Section 7216 carries the criminal penalty. Knowingly or recklessly disclosing or using tax return information without consent is a misdemeanour carrying a fine of up to $1,000 and up to a year of imprisonment. Section 6713 adds a separate civil penalty of $250 per disclosure capped at $10,000 a year, and it has no intent requirement at all, so an accidental disclosure is enough.

The regulations permit some disclosure without consent. Treasury Regulation section 301.7216-2(d) covers disclosure to a preparer located in the United States for tax return processing, and to contractors doing software programming, maintenance or testing. The second of those applies only to the extent necessary and only if every individual receiving the data gets written notice of the sections 7216 and 6713 penalties, which is not something you can arrange with a public chatbot. There is no general carve-out for AI vendors.

Where a US firm sends tax return information to staff or preparers located outside the United States, section 301.7216-2(c)(2) requires the taxpayer’s written consent before the disclosure. The duty sits on whoever makes the disclosure, normally the US firm, and consent must be obtained beforehand rather than ratified afterwards.

There is an exception worth knowing. Under section 301.7216-2(c)(3), where the taxpayer furnished the information to the offshore firm directly in the first place, that firm’s own internal use does not require consent. So an Indian firm engaged directly by a US client sits differently from an Indian contractor receiving work from a US firm.

For individual clients there is a harder limit. Section 301.7216-3(b)(4) bars a US preparer from obtaining consent to send a Form 1040 filer’s social security number outside the United States at all, unless an adequate data protection safeguard is in place and the consent records it, in the wording prescribed by Revenue Procedure 2013-14.

Your own exposure is separate. A contractor who receives tax return information is a tax return preparer under section 301.7216-2(d)(2), because they perform auxiliary services in connection with return preparation. That makes sections 7216 and 6713 apply to your onward disclosures, including into an AI tool.

The AICPA reached a similar conclusion by a different route. Its Statements on Standards for Tax Services, effective January 2024, added a section on reliance on tools whose definition expressly includes artificial intelligence. Using a tool does not absolve the member of professional obligations, and tools should enhance rather than supplant professional judgment.

Deloitte Australia delivered a review to a federal department that contained fabricated citations and a fabricated court quotation traced to a generative model, and refunded part of the fee. The IRS bulletin cites that incident.

What to put in the engagement letter

An engagement letter needs to name the AI you use, the data it receives, and who checks the output. Whether it can also carry the section 7216 consent depends on the client. For a Form 1040 filer it cannot: Revenue Procedure 2013-14 requires a separate document containing only that consent, in at least 12-point type, with prescribed wording, and uses and disclosures cannot sit in the same document. For business entities filing 1120, 1120-S or 1065, section 301.7216-3(a)(3) allows consent in any format, including the engagement letter itself.

Use of artificial intelligence tools

We use software with artificial intelligence features to assist with document
data extraction, transaction categorisation and payment matching. These
features are built into the accounting platform you have licensed. We may also
use AI assistants under business or enterprise terms that contractually
prohibit training on customer content.

All AI output is reviewed by a person before it is entered in your books or
provided to you. We do not treat AI output as final work.

We do not enter your identifying information, account numbers or tax
identification numbers into general-purpose AI tools. Where such a tool is
used for analysis, identifying details are removed first. Removing them is an
internal control and does not by itself remove the information from the scope
of Internal Revenue Code section 7216.

Where your tax return information is disclosed to any person located outside
the United States, including our own personnel, we will obtain your written
consent before the disclosure in the form the IRS requires. If you file in the
Form 1040 series, we will not send your social security number outside the
United States unless an adequate data protection safeguard is in place and
your consent records it.

We maintain the security safeguards our engagement with you requires,
including those we are required to maintain by contract under the FTC
Safeguards Rule.

How the FTC Safeguards Rule reaches contractors

The FTC Safeguards Rule reaches offshore bookkeepers through the client’s contract rather than through direct regulation. The rule lists tax preparation firms among the financial institutions it covers. Section 314.3(a) requires a written information security programme, and section 314.4 sets out what it must contain, including encryption in transit and at rest, multi-factor authentication, staff training, and a written incident response plan.

The element that matters to a contractor is section 314.4(f), which requires a covered firm to contractually oblige its service providers to maintain safeguards. That is how the duty reaches an Indian bookkeeping vendor with no direct FTC exposure. If you handle US tax data, expect those terms in your engagement and expect to be asked which AI tools you use.

India’s own regime is not yet the binding constraint. The Digital Personal Data Protection Rules were notified in November 2025, with obligations on notice, consent, breach reporting and security safeguards commencing on 13 May 2027. On the wider question, see data-protection duties when AI handles personal data.


How bookkeepers can use AI inside a client’s books

Bookkeepers working inside a client’s own QuickBooks or Xero file use AI on the client’s terms, not their own. This is the normal situation for offshore and contract bookkeepers, and almost every guide written on this topic ignores it. That advice assumes you hold the subscription. Frequently you do not.

You log into a file you do not own, on a plan you did not pick, with apps you cannot install.

Turning off AI in QuickBooks Online

Individual AI features cannot currently be turned off in QuickBooks Online. Intuit’s own help documentation states that “currently, there isn’t a way to turn off AI features individually.” The features arrive with the plan.

One QuickBooks Community thread is titled “How do I turn off AI on my QuickBooks Online account? I did not ask for and do not want this feature”. Xero differs mainly because JAX only acts when you open it and ask, and access follows existing user permissions.

Assume the AI is on, it is touching the file between your sessions, and its output will be waiting for you. That makes your review pass the only control you have.

What you control on a client’s subscription

Working inside someone else’s file means what you control moves from tools to process. You usually cannot buy Dext for a client, but you can write and document the categorisation rules, so the software follows your rule instead of guessing. You can also ask for the tier that has the agent you need, with a reason attached.

Know the tiers before that conversation, and note that Intuit also sells Free, Lite and Ledger below Simple Start, so a client may be on a plan with almost no AI at all. The ask itself is short:

Categorisation on this file is generating about 30 corrections a month, most of them sales tax. Moving from Essentials at $75 to Plus at $115 adds Sales Tax AI, which should cut that. Happy to review after two months and drop back if it does not.

Keep your own record of what the AI decided against what you decided. If a categorisation is later questioned by the client’s CPA or in an examination, you need a documented answer. A monthly note listing the entries you overrode, and why, takes ten minutes.

Escalation across time zones

Working an overnight shift against US business hours changes where you draw the line between deciding and parking. You review AI output when the client and their CPA are asleep, so the cost of a question is a full day rather than five minutes. That pushes people to decide things they should have asked about.

Set the rule in advance, before you are mid-shift. Anything touching revenue recognition, a capital purchase, an owner transaction, an intercompany movement, payroll, sales tax, or a period already closed gets parked with a note rather than decided.

Park it in a suspense account rather than leaving it unrecorded, so the bank reconciliation still ties:

Parked 2026-06-11, $8,500 payment OUT to ENTITY-2. Posted to Ask My Accountant
so the June reconciliation still ties. Not left unrecorded.

AI categorised it as Consulting income. Wrong in direction as well as account:
an outgoing payment coded to an income account debits Consulting income and
understates June revenue by $8,500.

ENTITY-2 shares a registered address with the client, per the January file. A
shared address is not evidence of common ownership on its own, since registered
agents and shared accountants produce the same thing, so this needs confirming.

Need: (1) who owns ENTITY-2; (2) if commonly owned, whether this is a loan, a
distribution or a genuine charge for services, and whether a due to / due from
account exists; (3) how ENTITY-2 has recorded the other side.

It takes about ninety seconds to write.


Why AI is not replacing bookkeepers

AI is not replacing bookkeepers, but it is shrinking the clerical half of the job while the judgment half grows. The numbers describe a migration. This article does not repeat the wider argument about whether AI will automate accounting, which is covered separately.

BLS projections for clerks and accountants

The BLS projections show one occupation shrinking and the adjacent one growing.

Measure Bookkeeping clerks Accountants and auditors
Jobs, 2024 1,613,400 1,579,800
Median pay, 2024 $49,210 (about ₹47.2 lakh) $81,680 (about ₹78.4 lakh)
Projected change, 2024-34 -6%, or -94,300 +5%, or +72,800
Annual openings About 170,000 About 124,200

The Bureau of Labor Statistics names technology as the cause of the clerk decline, and still projects about 170,000 openings a year in that shrinking occupation, all arising from people moving on or retiring. Accountants and auditors grow over the same period. The gap between the two medians is $32,470, or roughly ₹31.2 lakh a year.

The move from clerk work to accountant work

Moving from clerk work to accountant work means being paid for decisions rather than entries. In Intuit’s 2026 survey, 64% of the professionals using AI said high-stakes client moments remained entirely or primarily human-driven. Across all 725 respondents, 41% named trust and liability as the main reason clients keep paying people.

The profession’s confidence in AI output is low. A July 2026 ACCA and CA ANZ study of 1,600 finance professionals found 93% reporting at least moderate concern about the integrity and verifiability of AI-generated insights, with 67% highly concerned. That concern is what clients pay a reviewer for.

Governance has not kept up. Karbon’s January 2026 study found 98% of surveyed firms using AI while only 21% had an AI policy or strategy. In KPMG’s May 2026 survey of 1,013 senior finance leaders, 33% of organisations able to produce AI audit evidence efficiently reported significant improvement in error reduction, against 6% of those that could not.

What happens to your hourly rate

Your hourly rate falls with the hours, if you bill by the hour. A file that took ten hours and now takes three cuts your revenue by 70% while you get better at the work. The loss comes from the billing method, not the tool.

Advice to pivot to advisory is usually presented as a skills problem. It is a pricing problem first. The usual transition is to price a fixed monthly fee at your current average monthly hours times your rate, define the scope in writing, then hold that fee as the hours fall. That makes how you charge US clients worth revisiting before you adopt any of these tools.

If your pitch is that you are cheaper at data entry, AI competes directly with that pitch and keeps getting cheaper. If your pitch is that you catch what the software got wrong and can explain it, AI creates the work instead.


Frequently asked questions

What is AI bookkeeping? AI bookkeeping is the use of machine learning to read documents, propose transaction categories and match payments, inside or alongside a ledger like QuickBooks or Xero. It automates the mechanical steps of the bookkeeping cycle. It does not make accounting decisions, and its output requires review before it is booked.

Can ChatGPT do my bookkeeping? ChatGPT cannot do your bookkeeping, because it does not connect to your ledger or your bank feed unless you build that connection. It is useful for explaining a treatment, drafting client communication, and reviewing categorisations you already have. Pasting client financial data into a consumer account also raises the confidentiality problems set out above.

Can AI handle industry-specific bookkeeping? AI handles industry-specific work poorly, because the treatment usually depends on information the bank feed does not carry. Construction progress billing, real estate escrow, restaurant tips and e-commerce sales tax all require knowing something about the transaction beyond its description and amount. Expect to write manual rules for these and to check them every month.

Do I need to record which entries AI suggested? No rule requires a dedicated AI audit trail, but the June 2026 IRS Office of Professional Responsibility bulletin tells practitioners to document AI usage and verification, and to show the steps taken to comply with Circular 230 section 10.36. A monthly note of what you overrode and why is the cheapest way to do that, and it also supports a reasonable-cause defence if a position is questioned.

Which tools have genuine AI and which are marketing? A useful test is whether the vendor explains what the model does when it is uncertain, and whether the product improves from your corrections. Neither QuickBooks nor Xero shows a confidence score, but both describe what happens to uncertain items and both learn from what you do. Tools that describe AI only in benefit terms, with no account of error handling at all, are usually applying rules with a new label.

Who is responsible if AI miscategorises something on a filed return? The taxpayer carries the tax, interest and any accuracy-related penalty. A preparer who signs the return can additionally face a penalty under Internal Revenue Code section 6694 where a position lacked substantial authority and they knew or should have known, though that penalty has a defence where there was reasonable cause and good faith, judged partly on normal review practice. Software vendors are not free of exposure either: a contractor receiving tax return information is itself a preparer for section 7216 purposes.

How long does AI take to learn a client’s books? Categorisation accuracy improves over the first two to three months as the model accumulates history for that file. It improves fastest when you correct errors rather than working around them. Accuracy on unusual and one-off transactions does not improve much with time.

Should I use AI on a client’s first month of books? Use it for document capture, and be sceptical of its categorisation. In month one the model has no history for that client, so it is working from the description string and general patterns, and it treats your month-one coding as the signal for month two. The bigger first-month risk sits outside the AI entirely, in opening balances, the conversion trial balance and the outstanding receivables and payables at the changeover date.

How do I choose between QuickBooks and Xero for AI features? Xero gives every subscriber the JAX assistant at no extra charge. QuickBooks gates its agents by tier, so Accounting AI needs Essentials at $75 a month and Sales Tax AI needs Plus at $115, with prices for those tiers changing for renewals from 1 August 2026. Xero’s US plans run $25, $55 and $90 a month. If the client already holds a subscription, the choice is usually theirs and not yours.

Do I need the client’s permission to use AI on their books? For US tax return information, consent is not merely good practice. Section 7216 requires the taxpayer’s written consent, obtained before the disclosure, where a US firm sends that information to a preparer or staff member located outside the United States. For a Form 1040 filer that consent must be a separate document in the wording set by Revenue Procedure 2013-14, and the social security number cannot go abroad at all unless an adequate data protection safeguard is in place.

How much do AI bookkeeping tools cost? Costs range from included to several hundred dollars a month. Xero includes JAX at no additional charge, and QuickBooks runs from $38 to $275 a month by tier, or about ₹3,650 to ₹26,500. Separate platforms run from $17.70 per client per month for Dext, on a 10-client minimum, to $299 and above for Docyt, with Zeni starting near $494 a month on annual billing.

Do AI bookkeeping tools replace QuickBooks or Xero? Most do not replace the ledger, they connect to it. Dext, Ramp and similar tools push data into QuickBooks or Xero rather than becoming the book of record, and Ramp syncs to more than 30 accounting systems. Puzzle is the clearest exception, because its AI is built into a general ledger of its own and it markets a direct switch from QuickBooks. Check which kind you are buying before you commit a client to it.

Will AI replace bookkeepers? AI is reducing the clerical portion of bookkeeping rather than eliminating the occupation. The BLS projects a 6% decline in clerk roles to 2034 alongside about 170,000 openings a year, and 5% growth in accountant roles. The displacement question is covered in more depth in our piece on whether AI will automate accounting.

Is there still demand for bookkeepers who do not use AI? There is demand, but it is concentrated in smaller clients and it competes on price. The BLS projects about 170,000 clerk openings a year, so the roles exist. The practical risk is that a bookkeeper who does not use AI takes longer on the same file and has to price against people who do.

What should a bookkeeper learn to stay employable? Learn the treatments AI cannot infer from a bank feed: revenue recognition, capitalisation against expense, intercompany transactions and multi-entity work. Add the review discipline itself, plus enough data-handling knowledge to answer a client’s security questionnaire. Those skills move a clerk role towards the accountant role, where the BLS median is $32,470 higher.


References

Official and primary sources

  1. US Bureau of Labor Statistics: Bookkeeping, Accounting, and Auditing Clerks: 2024 median pay, employment, and 2024-34 projections
  2. US Bureau of Labor Statistics: Accountants and Auditors: comparative median pay and growth projection
  3. IRS Office of Professional Responsibility, Introductory Guidelines for Responsible AI Use in Federal Tax Practice: Issue 2026-19, 24 June 2026
  4. Treasury Department Circular No. 230: due diligence, competence and written advice duties
  5. 26 U.S.C. section 7216: penalties for disclosure or use of tax return information
  6. Treasury Regulation section 301.7216-2: permissible disclosures, including the offshore rule at (c)(2) and the exception at (c)(3)
  7. Treasury Regulation section 301.7216-3: consent requirements, including the social security number restriction at (b)(4)
  8. Revenue Procedure 2013-14: mandatory form and wording of section 7216 consents
  9. AICPA Statements on Standards for Tax Services, Nos. 1-4: section 1.4, reliance on tools
  10. FTC Safeguards Rule: covered institutions and required security programme elements
  11. Digital Personal Data Protection Rules 2025: Ministry of Electronics and Information Technology, notified November 2025

Research and studies

  1. Transaction Categorization with Relational Deep Learning in QuickBooks: Intuit and University of Notre Dame, arXiv:2506.09234
  2. Categorising SME Bank Transactions with Machine Learning and Synthetic Data Generation: University of Warwick, SME Capital and Navrisk, arXiv:2508.05425, August 2025
  3. ACCA and CA ANZ, Enabling finance insight: July 2026, 1,600 finance professionals
  4. KPMG Global AI in Finance 2026: May 2026, 1,013 senior finance leaders
  5. Karbon, State of AI in Accounting 2026: January 2026
  6. Intuit 2026 Accountant Technology Survey: 725 US accounting professionals, fielded May 2026

Vendor documentation and policies

  1. Overview of Intuit AI in QuickBooks Online: the eight agents, and the statement that AI features cannot be turned off individually
  2. QuickBooks Online August pricing changes: published 27 June 2026
  3. Xero, JAX: availability, pricing and LLM data handling
  4. Intuit Global Privacy Statement: updated 9 March 2026
  5. OpenAI, How your data is used to improve model performance
  6. Anthropic commercial products and model training

Practitioner sources

  1. QuickBooks Community: QBO Suggested Categorization broken
  2. QuickBooks Community: turning off AI in QuickBooks Online

Notes on figures

Vendor accuracy claims (Digits, Dext, BILL, Docyt) are self-published and based on the vendors’ own benchmarks and test sets. Two of the survey sources also have a commercial interest in their findings: Karbon sells practice-management software to the firms it surveyed, and KPMG sells AI assurance services. References 12 and 13 are arXiv preprints that have not been peer reviewed, and neither is free of commercial interest: reference 12 is co-authored by Intuit researchers, and reference 13 by authors at SME Capital and Navrisk.

Pricing was checked on 28 July 2026 at US list rates, excluding promotional discounts. Rupee equivalents use 95.94 rupees to the dollar as at 28 July 2026 and are indicative only. QuickBooks Essentials, Plus and Advanced prices change for renewals from 1 August 2026 and Intuit has not published the new figures.

The IRS bulletin at reference 3 is distributed through the US government’s official GovDelivery platform on the IRS account; at the time of writing it does not appear on the irs.gov alerts index. Several sites, including bls.gov, irs.gov, openai.com and quickbooks.intuit.com, block automated access and may return errors to link checkers while opening normally in a browser.


This article is for informational and educational purposes only and does not constitute professional, financial, tax, legal or career advice. Software features, pricing, accuracy figures and regulatory guidance change frequently and vary by jurisdiction and plan. Verify current details with the IRS, the FTC, the relevant vendor, or a qualified professional before making decisions about client data handling, tool selection or tax positions.

Comments

No comments yet. Why don’t you start the discussion?

Leave a Reply

Your email address will not be published. Required fields are marked *