Every finance vendor now claims AI. Most of it is a chat box on top of unchanged software. It is worth being specific about what AI genuinely does well inside an accounting system, and what it should not be allowed near.
Transaction matching. Reconciliation is pattern recognition over structured data with a clear right answer — close to an ideal application. Systems learn from your historical matches and handle variations that break rigid rules: a slightly different reference, a payment covering several invoices, a partial settlement.
Document extraction. Pulling structured data from invoices and receipts. Modern extraction handles unfamiliar layouts without templates, which is the main advance over older OCR.
Coding suggestions. Proposing the account, cost centre and tax code based on supplier and description history. Reliable for recurring suppliers, less so for genuinely new ones — which is the correct place for a human to look.
Anomaly detection. Surfacing transactions that differ from the established pattern. A duplicate payment, an expense well outside the norm for a category, a supplier invoice that jumped 40%. This is genuinely valuable because it finds things nobody was looking for.
Flux commentary drafting. Producing the first pass of "marketing spend rose 22% due to the campaign in March" by reading the underlying movements. A draft, reviewed by a person, saves real time in the close.
Natural language querying. Asking a question of the ledger and getting an answer without building a report. Useful for exploration; the answer still needs to trace to underlying data before it enters a board pack.
Judgement on accounting treatment. Whether an implementation fee is a distinct performance obligation is a judgement about your specific contract, applied under a standard, with your auditor's agreement. A system can surface the question. It should not decide it.
Estimates and provisions. Doubtful debt provisions, bonus accruals, impairment assessments. These are forward-looking judgements about uncertain outcomes. Automating them produces a number nobody owns.
Anything unreviewed at the boundary. Payments leaving the business, journals posting to the ledger, invoices going to customers. Automate the preparation; keep a human at the point of irreversibility.
Explaining anomalies. Finding an anomaly is pattern recognition and machines are good at it. Determining whether it is an error, a genuine business change, or fraud requires context the system does not have.
Not "do you have AI" — everyone says yes. Ask instead:
There is a genuine difference between AI built into the data layer and AI added on top.
A system where reconciliation, anomaly detection and coding are part of how transactions are processed behaves differently from one where a language model reads a database and answers questions about it. The first changes the work. The second is a better search box.
Campfire's approach is the former — Ember AI is built on a model trained specifically on accounting data and embedded in the workflows rather than layered over them. That is the distinction worth probing with any vendor.
AI in accounting is not replacing finance teams. It is removing the mechanical majority of the work — matching, extraction, coding, first-draft commentary — and leaving the judgement, the review and the explanation.
That is a meaningful shift in what a small finance team can carry. It is not the same as autonomy, and any vendor promising the latter is worth more scepticism, not less.
We configure these capabilities against real transaction patterns, set the confidence thresholds, and make sure the review points are in the right places — which is the difference between AI that saves time and AI that creates unreviewed risk.
Get in touch, or see what to automate first in reconciliations.