AI Pre-Accounting Agent: What Should It Do?

Your finance team does not need another chat box that can name an expense account. It needs a system that turns a receipt, invoice or bank line into a proposed accounting record, shows what it inferred, and knows when to stop.

That is my answer to the question behind the current search interest in AI pre-accounting agents. The agent is worth using when it reduces complete units of reviewable work. A category suggestion by itself is not a unit of work. Neither is an impressive accuracy percentage whose denominator you cannot inspect.

This distinction matters because vendors are moving quickly from assistants to operators. Moss describes a pre-accounting agent that codes expense accounts, cost centres and due dates, and advertises accuracy above 98%. On 7 August 2026, freee announced accounting agents that automate high-confidence entries and matching, while sending lower-confidence cases to people. On 26 August, Meridian by Pilot launched Meridian AI, which performs ad hoc accounting work, asks when information conflicts and returns a reviewable summary.

The industry conversation has moved past whether AI can suggest a code. The useful question is whether the proposed entry survives the path from evidence to ledger.

What is an AI pre-accounting agent?

Pre-accounting is the work between receiving evidence and accepting an accounting entry. It includes collecting documents, extracting dates and amounts, identifying suppliers, splitting line items, applying VAT treatment, proposing accounts or cost centres, matching payments and preparing exceptions for review.

An AI pre-accounting agent coordinates those steps. It differs from OCR because OCR reads fields. It differs from a chatbot because it advances a workflow. It also differs from an autonomous accountant because a company still owns its accounting policies, approvals and filing responsibility.

The best mental model is a junior operator with a narrow mandate and unusually good memory. It may prepare a lot, but it should not invent policy. When the evidence is incomplete, it should produce an explicit question or waiting state, not a plausible answer.

That is also why we separate this role from the broader AI accounting agent. The pre-accounting agent prepares trustworthy inputs. Later workflows may reconcile, report or file them, but only after the evidence and authority are clear.

Which tasks should an AI pre-accounting agent automate?

Buyers often compare products by counting AI features. I would compare the handoffs instead. Every stage should leave an output that the next stage can verify.

This table exposes a common product gap. A tool can be excellent at extraction and still leave a person rebuilding the transaction. It can be excellent at coding and still attach the wrong payment. It can send everything to review and technically remain safe while saving almost no time.

The whole path matters. Our guide to AI expense management software makes the same point from the buyer side: scanning is the start, not the outcome.

How accurate does an AI pre-accounting agent need to be?

“98% accurate” sounds precise, but it is not a complete metric. Ninety-eight percent of which fields? On what documents? Before or after deterministic validation? Does a wrong supplier count the same as a wrong VAT rate? Does the sample include ten-line receipts, foreign currencies and company-specific cost centres?

The number can be legitimate and still fail to describe operational risk. A model may populate 98 of 100 fields correctly while the two wrong fields determine the tax treatment. Another system may automate only 70 cases, park 30 honestly, and create fewer bad entries. Coverage, field accuracy, accepted-entry accuracy and straight-through automation are different measures.

I would ask for four denominators: documents attempted, fields proposed, entries accepted without edits and entries later corrected. Then split the result by field and document type. If a vendor cannot provide that shape, run a representative trial and measure it yourself.

This is why we have argued before that AI bookkeeping accuracy needs a denominator. A headline percentage becomes useful only when you can connect it to a business action and its correction path.

Why do line items and VAT matter?

A document is not one label. It is a set of accounting claims that happen to arrive in one file.

We can see that in Norman's production system. From 9 June through 7 September 2026, more than 10,000 active invoices and receipts uploaded by users or through chat produced persisted multi-line extraction. Together they contained more than 50,000 lines. The median document had three extracted lines; the upper tenth had ten or more. Roughly one in twelve contained more than one VAT rate.

Almost every extracted line received a category proposal. That is more than 99% suggestion coverage. It is not 99% accuracy, and we do not present it as such. It tells us the system usually has an answer. The mixed-VAT rate and line-count distribution tell us why accepting that answer at document level would be unsafe.

The architectural consequence is simple: preserve the lines. A German supermarket receipt can combine reduced-rate food and standard-rate household goods. A services invoice can mix reimbursable costs with professional work. One account, one VAT rate and one confidence score cannot represent either document faithfully.

In Norman, line extraction keeps the printed order, amount and VAT rate. Category proposals are resolved against the relevant company's available accounts. Invalid foreign rates are retained as evidence while the allowed local treatment is applied separately. These are shipped controls, not claims that the model never makes a mistake.

What should happen when the agent is uncertain?

Uncertainty should become workflow state. It should not disappear inside a score.

There are at least four useful outcomes: accept a reversible low-risk proposal, ask for a missing fact, offer a small ranked set of candidates, or route the exact case to review. “Could not complete” is sometimes the correct result. It is better than a confident entry built from missing evidence.

The review screen should show the original document, the proposed line, the rule or evidence behind it and the effect of accepting it. A correction should change the working entry without erasing the original proposal. Repeated corrections may become a company-specific pattern, but only after the system knows they represent policy rather than a one-off exception.

This is where observability becomes a product feature. Our approach to tracing AI agent runs records the workflow shape and decision fingerprints without duplicating private document content. A buyer does not need our implementation, but should demand the same outcome: an operator can reconstruct what happened without searching logs or trusting memory.

How do you compare AI pre-accounting software?

Start with your messiest normal month, not a polished vendor demo. Include receipts with several VAT rates, recurring suppliers whose coding changed, foreign invoices, partial documents, duplicates, delayed bank lines and a few items that should remain unanswered.

Measure time to a reviewable entry, not time to a suggestion. Record how many documents are complete, how many are parked with a useful reason, how many need edits and how many wrong proposals cross into the ledger. Then repeat the same set after corrections to see whether learning is scoped and durable.

Also verify the boundaries around the agent. Can it write directly, or only prepare? Can permissions differ by action? Does approval bind to the exact payload you reviewed? Can a later bank event trigger safe rematching? Can you export the resulting accounting record and its evidence?

An AI pre-accounting agent should make the month smaller. It should turn raw files into complete, inspectable proposals and concentrate human attention on real ambiguity. If it merely replaces typing with checking every field, it has changed the interface, not the economics of the work.

The winning systems will not be the ones that always answer. They will be the ones whose answers can become accounting records without losing the evidence, the lines or the right to say “not yet.”