An AI accounting glossary for finance people

Vendor conversations are harder than they need to be because the vocabulary is unsettled and some of it is deliberately vague. Here is what the terms mean, with a note on which ones tell you something and which are decoration.

Agentic. Software that completes a sequence of actions and decides when to involve a person, rather than proposing each step. Meaningful, and it raises the stakes: ask what hard limits constrain it and how you reverse a run that went wrong. See agentic AI in the finance function.

Audit trail. The record of what happened to a transaction — what was posted, on what evidence, by whom or by which process, and when. In an automated system it must also capture the model or rule version and the confidence level. Ask: can my auditor read it without your help?

Confidence score. A number expressing how sure the system is about a proposal. This is one of the most useful things in the interface, because it is what allows automatic handling of the certain cases and human review of the rest. Rule-based systems cannot produce one — a rule either fires or does not — so its absence tells you something about what is underneath.

Continuous close. Keeping the books effectively current so that period-end is a review rather than a construction project. A real operational goal rather than a product feature; achievable to the extent your transactions are captured and reconciled as they occur.

Document AI / intelligent document processing. Extracting structured data from unstructured documents — invoices, statements, receipts. The mature end of accounting AI. Ask: does it need a template per supplier? If yes, it is the older technology under a newer name.

Exception queue. The list of transactions the system could not handle confidently. Not a failure list — it is the control that makes the rest safe, and it is the single most informative screen in any accounting automation product. Ask to see it, populated with real messy data, not a demo dataset.

Explainability. Whether the system can show why it proposed something. In accounting this is not philosophical: it is the difference between an entry you can defend and one you cannot. Ask: click into this figure and show me the transactions behind it.

Extraction accuracy. The share of document fields read correctly. Quoted constantly, and less useful alone than it sounds — a system that is 99% accurate and silent about the 1% is worse than one that is 96% accurate and flags what it was unsure about. Ask what happens to the ones it got wrong.

Fine-tuning. Adapting a general model to a specific domain or customer using additional examples. Mostly an implementation detail. It matters to you only in one respect: whether your data is used to train models that serve other customers. Ask that directly.

Hallucination. A model producing fluent, confident, fabricated output. In finance this is the most dangerous failure mode, because a wrong number in a well-argued paragraph passes casual review. The defence is architectural — figures must be computed by the system and retrieved, never generated by the model. See AI hallucinations in financial data.

Human in the loop. A person reviewing or approving before an action completes. Meaningful when it names which actions and which person; meaningless as a general reassurance. Ask: which specific steps require a human, and what happens if they do not respond?

Large language model (LLM). A model trained to predict and produce language. Excellent at reading documents and explaining figures, poor at arithmetic. What it changed in accounting is covered in what language models changed in accounting.

Machine learning. Software that infers rules from examples rather than being told them. In accounting it is what proposes coding and matching, and what improves when you correct it. Distinct from rules — see rules versus machine learning in accounting.

Natural language interface. Asking the system questions in words instead of building a report. Genuinely useful; the value depends entirely on whether answers are traceable to records.

OCR. Optical character recognition — turning an image of text into text. A component, not a capability. Modern document AI includes OCR but the interesting part is what happens after: understanding which number is the tax and which is the total.

Predictive analytics. Forecasting from historical patterns — cash flow, demand, payment behaviour. Useful for the questions where history genuinely predicts, less so where the future differs from the past for reasons the data cannot see.

Prompt. The instruction given to a model. Relevant to you when a product is configured by describing what you want, which is increasingly how process changes are made.

RPA (robotic process automation). Software that mimics clicks and keystrokes to move data between systems. Pre-AI technology, still widely sold. It is brittle — it breaks when a screen layout changes — and its presence often indicates the systems involved lack proper integrations. Ask whether the connection is a real integration or a robot clicking through a screen.

Reconciliation rate. The share of transactions matched without human intervention. Genuinely informative if defined honestly. Beware two things: whether unmatched items are excluded from the denominator, and whether "matched" includes items dumped into a balancing figure. We treat around 98% as a design target for what should match automatically, deliberately short of 100% because the remainder is real ambiguity a person should see. Separately, full traceability means every line is accounted for whether or not it matched.

Straight-through processing. A transaction completing end to end without human touch. A useful measure of automation depth. Ask what percentage genuinely goes straight through, and what the exceptions have in common.

Structured / unstructured data. Structured arrives in defined fields — a bank feed. Unstructured does not — a PDF invoice, an email. The whole value of document AI is turning the second into the first.

Training data. The examples a model learns from. In accounting, usually your own transaction history. Ask: where is it stored, who else can see it, and what happens to it if we leave?

Three phrases that mean nothing on their own

"AI-powered." Applies equally to a language model and a rule someone wrote in 2011. Always ask what specifically the AI does.

"Fully automated." Nothing in accounting is fully automated, because exceptions exist by definition. A vendor claiming it either has not met your data or is not counting the exceptions.

"Learns your business." True of a well-built system, but also the easiest sentence in software to say. The test is concrete: does correcting it change its future behaviour without anyone editing configuration?

Common questions

What is the difference between OCR and document AI?

OCR converts an image of text into machine-readable text and stops there, whereas document AI interprets what the text means — identifying which figure is the total, which is the tax, and which supplier issued the document. Products that are effectively OCR plus a template per supplier are older technology, and the test is whether the system copes with an invoice layout it has never encountered.

What is a good reconciliation rate?

It depends on how honestly the figure is defined, so ask whether unmatched items are excluded from the denominator and whether anything is being absorbed into a balancing figure. As a design target, matching around 98% of lines automatically and surfacing the remainder as genuine exceptions is a reasonable expectation for high-volume reconciliation; a claim of 100% usually means the exceptions are being hidden rather than resolved.

Should I worry about my data training someone else's model?

It is a fair question to ask directly, and the answer should be specific rather than reassuring. What you want to know is where your data is stored, whether it is used to train models serving other customers, who at the vendor can access it, and what happens to it if you terminate the contract.

Is RPA the same as AI accounting?

No. Robotic process automation mimics human clicks and keystrokes to move data between systems and breaks whenever a screen changes, whereas AI accounting interprets content and generalises to cases nobody configured. The presence of RPA in a proposal often indicates the underlying systems lack real integrations, so it is worth asking whether a connection is a genuine integration or a robot operating a screen.

Related: AI in accounting · how to evaluate AI accounting software · questions to ask an AI accounting vendor