Picture two errors, both made by the same software, in the same month. In the first, it miscodes a client dinner as office supplies. In the second, it decides a twelve-month software contract should be expensed rather than capitalised.
The first is caught at reconciliation, in days, by controls you already run. The second balances perfectly, ties out cleanly, and gets discovered by a tax authority or an auditor at a time of their choosing — by which point the same treatment has been applied all year.
That asymmetry is the entire policy. Automate the layer where mistakes are detectable and reversible: transaction categorisation, receipt and invoice extraction, bank reconciliation matching, anomaly flagging. Do not automate the layer where mistakes are silent: accruals, revenue recognition, capitalisation calls, tax positions, and anything with your name on it.
The detectability principle, and why it beats a feature list
Every accounting task sits somewhere on a spectrum defined by one question: if the machine gets this wrong, how and when do I find out?
At the safe end, the error announces itself. A bad categorisation makes the month-over-month variance look strange. A mis-keyed invoice amount fails to reconcile against the bank feed. A duplicate payment shows up in the AP ageing. The controls that catch human error catch machine error too, because they were never checking who made the entry — they were checking whether the numbers agree.
At the dangerous end, the error is structurally invisible. If a model decides a twelve-month software contract should be expensed rather than capitalised, nothing fails. The books balance. The reconciliation ties. The error is a judgment, and judgments do not throw exceptions. They accumulate quietly and are discovered by someone external, at a moment of their choosing.
Which is why the useful frame is not "is this tool accurate?" It is "does this task have a control that catches the tool being wrong?" If yes, automate aggressively. If no, the tool can prepare, summarise and suggest — but a human decides, and the human is the one on record.
The layers, and what to automate in each
| Layer | Representative tools | Automate? | How the failure shows up |
|---|---|---|---|
| Document capture (receipts, bills, statements) | Dext, Booke, Ramp | Yes, aggressively | Wrong amount fails to reconcile; wrong vendor is visible |
| Transaction categorisation | Booke, QuickBooks/Xero built-ins | Yes, with review of low-confidence items | Variance analysis and month-end review catch it |
| Bank reconciliation matching | Booke, native ledger tools | Yes | Unmatched items sit there, loudly, until cleared |
| AP / invoice approval | Vic.ai, Ramp | Suggest, then human approves | A wrong approval is money out the door — irreversible |
| Spend policy enforcement | Ramp | Yes, as flagging | Flags are reviewable before the card is charged |
| Anomaly and duplicate detection | Ramp, Puzzle, analytics layer | Yes — this is pure upside | False positives cost you a minute; misses cost you a client |
| Analysis and variance commentary | Julius, Hex | Draft only | Confident, wrong narrative — must be checked against source |
| Accruals, cut-off, revenue recognition | None. Seriously | No | Silent. Surfaces in audit |
| Tax positions, filings, client advice | No | No | Silent, and it is your licence |
Swipe the table sideways to see every column →
That last row is not a stylistic preference. It is the line between bookkeeping software and unauthorised practice.
Capture and categorisation: the layer worth every penny
This is where the hours actually are, and where the technology is genuinely mature. Optical extraction of a total, a date, a vendor and a tax line from a photographed receipt is a solved problem in a way that "should we recognise this revenue" is not.
Booke is built specifically around the bookkeeping grind: pulling documents in, categorising transactions, learning from your corrections, and — the part practitioners actually rate — automating the client chase for missing receipts and unanswered queries. That last function is doing more work than the AI, honestly. Half the pain of bookkeeping is not the ledger; it is the eleven emails to get a coffee receipt.
Who should use it: bookkeepers and small firms with a lot of clients and a lot of transaction volume. Who should not: anyone whose real problem is advisory capacity rather than data entry. It will not give you that.
Dext and similar capture tools remain the reliable backbone for extraction, and most firms end up with a capture layer plus a ledger rather than one product doing both.
Spend, cards and AP: automate the flagging, never the approval
Ramp sits on the corporate-card and spend-management side, using automation to match receipts, enforce policy, and flag things that look wrong — duplicate subscriptions, out-of-policy spend, a vendor being paid twice.
The distinction that matters: flagging is safe, approving is not. A flag is a suggestion that costs a minute to dismiss. An automated payment approval is money leaving the building, and money leaving the building is the one accounting error that does not reverse cleanly. Vic.ai and similar AP tools are good at reading and coding invoices; let them prepare the entry and let a person press the button.
Who should use Ramp: companies where employee spend and card sprawl are the mess. Who should not: a bookkeeping practice looking for a general ledger. It is not that.
The AI-native ledger
Puzzle takes the more ambitious position — an accounting system built around automation from the start, aimed largely at startups, with the pitch that the books stay close to real-time rather than being assembled in a panic each quarter.
The honest trade: the further a tool automates into the ledger, the more it depends on your inputs being clean and your business being ordinary. Complex revenue arrangements, multi-entity structures, unusual inventory — these are exactly the situations where a system designed for the common case makes confident errors. Startups with straightforward SaaS economics are the sweet spot. Anyone with genuinely awkward accounting should be more suspicious the smoother the demo looks.
Analysis: draft-only, permanently
Julius will run analysis on a spreadsheet in conversation, and Hex will do the same for teams working in SQL. Both are useful for the "why did marketing spend jump" question at month-end.
Both will also produce a fluent, well-structured, entirely wrong explanation if the question is ambiguous or a column means something other than its name suggests. Treat the output as a first draft from a fast, confident junior: the calculation is usually right, the interpretation needs checking against the source, and the commentary should never leave the building unread. See the AI spreadsheet tools shelf for the wider field here.
Never: the chatbot as a technical authority
Asking a general-purpose model for a tax treatment, a standard reference, or a filing deadline is the single most dangerous thing an accountant can do with AI. Language models produce citations that look exactly like real citations. They will name a code section, quote a paragraph number, and be wrong in a way that reads as authoritative — and unlike a bad categorisation, nothing in your workflow catches it.
Use a model to help you phrase an explanation you already understand. Never to establish what the answer is. Go to the primary source, every time.
Audit trail: the non-negotiable
Before any tool touches a client's books, it has to answer four questions about every entry it makes:
- Who or what made this change? The tool must be identifiable as the actor, distinct from a human user. If automated entries are indistinguishable from staff entries, you have lost your trail.
- When, and on what basis? Timestamp, plus the source document that drove it.
- Can I see the source? A one-click path from the ledger line back to the original receipt or invoice image. If the tool extracts data and discards the document, walk away.
- Can I reverse it, and does the reversal show? Corrections should be visible as corrections. A tool that silently overwrites its own earlier entries is destroying evidence, whatever the marketing says.
Add one more thing, which many firms miss: confidence thresholds. Any categorisation tool worth using can tell you how sure it is. Route the uncertain items to human review, and let the confident ones through. A tool that gives you a single undifferentiated stream of guesses has pushed all the review cost back onto you, which is exactly the cost you were trying to remove.
Client confidentiality, said out loud
Accountants operate under professional confidentiality duties — the AICPA code in the US, equivalent rules through ICAEW, ACCA, CPA Canada and others, plus statutory obligations covering tax data specifically. Pasting a client's trial balance into a consumer chatbot is not a grey area.
The practical guardrails:
- Check the training terms. Consumer tiers of general chatbots may use inputs to improve models. Business and enterprise tiers generally commit not to. That distinction is the whole ballgame, and it lives in the terms, not the homepage.
- Get a data-processing agreement from any vendor touching client records, and know where the data is stored.
- De-identify by default. Most analysis works fine with the client's name stripped out. Do that as a habit, not as an exception.
- Read your engagement letters. If subcontracting or third-party processing needs consent, an AI vendor is a third-party processor, whatever it is called in the pitch deck.
- Write the policy before the tools arrive. Every firm that has had this go wrong had a staff member using something the partners had not heard of.
A rollout that will not embarrass you
- Pick the single most repetitive task in the practice. It is almost certainly receipt capture or transaction categorisation. Start there and nowhere else.
- Run the tool in parallel with your existing process for one full month-end. Compare outputs. Do not skip this to save the time.
- Look specifically at what it does with the awkward transactions — the ones that are half personal, the ones with no receipt, the ones from the vendor with three trading names. Anyone can categorise a Slack invoice.
- Turn on confidence thresholds. Review the uncertain pile. That pile is your training data, and correcting it is what makes month two better than month one.
- Keep the judgment layer human, in writing, in your policy. Suggest, review, approve — in that order, with a name attached to the approval.
- Tell your clients you are using it. They will find out; better it is from you.
FAQ
Will AI replace bookkeepers?
It is compressing the data-entry portion of the job hard, and that portion was already shrinking through bank feeds and rules engines. What it does not touch is the part clients actually pay for: knowing which questions to ask, catching the thing that looks fine but isn't, and being a person who can be held responsible. Practices billing for keystrokes are exposed. Practices billing for judgment are, if anything, better off.
Is it safe to put client financial data into ChatGPT?
Not on a consumer tier, and not with identifying details, unless you have checked the terms and your professional obligations allow it. Business and enterprise tiers typically carry no-training commitments and offer a data-processing agreement; the free tier is a different product with different terms. De-identify anything you can, and treat the chatbot as an outside party — because that is what it is.
What should accountants never automate with AI?
Anything where a wrong answer is silent. Accruals and cut-off, revenue recognition, capitalisation decisions, tax positions, filings, and client advice. The common thread is that no downstream control catches the error — the books still balance while being wrong, and the discovery happens during an audit or an enquiry rather than during your review.
Can AI do a bank reconciliation?
The matching, yes, and well — this is close to an ideal use case, because unmatched items are loudly visible and nothing is destroyed by a bad guess. What AI cannot do is decide what an unexplained item is. It will surface it. Somebody still has to know that the £4,000 transfer was a director's loan and not revenue.
Do AI bookkeeping tools keep an audit trail?
The good ones do, and it is the first thing to verify rather than assume. You need automated entries attributable to the tool rather than to a person, a link from each ledger line back to the source document, and corrections that appear as corrections rather than as silent overwrites. If a vendor cannot demonstrate all three in a demo, that is your answer.
Where to go next
The full field lives in AI finance tools, with the analysis side in AI spreadsheet tools and the plumbing in AI automation tools — worth a look if half your process is moving files between systems. Firms serving owner-managed clients should also read the small business hub, and our wider AI finance tools rundown covers the products this piece only names in passing. Using something in practice that is genuinely earning its keep? Submit it.