Reconciliation is where finance agents earn trust

Reconciliation is the best place for a finance agent to start because every result can be checked. Either two records agree, or the agent explains the difference and a reviewer can test the explanation. That makes reconciliation a steady source of evidence about the agent's accuracy, which is exactly what any later authority in the books should be built on.

Why is reconciliation such checkable work?

Reconciliation compares two sets of records that should agree and accounts for any differences. The bank statement against the cash book. The supplier statement against the payables ledger. The sub-ledger against the control account. The balance sheet account against its supporting schedule.

What makes it ideal for building trust is that the answer is knowable. At the end of a reconciliation, every item is either matched, explained or outstanding. A reviewer can check each match against the two records, test each explanation against the evidence, and see whether outstanding items are truly unresolved or simply missed. There is no room for a persuasive but vague answer.

Compare that with variance commentary or forecasting, where a plausible explanation can be wrong without anyone noticing for months. Reconciliation errors surface quickly: an unmatched item reappears next month, a balance does not agree, an auditor's confirmation does not match.

What does the agent actually do?

Three things, in increasing order of difficulty.

Match. The agent pairs items across the two records by amount, date, reference and counterparty. Simple matches are easy. Harder cases include one-to-many matches, such as a single bank receipt paying several invoices, timing differences across period ends, and references that are truncated or mistyped.

Explain. For items that do not match, the agent proposes an explanation: a payment in transit, a bank charge not yet recorded, a duplicate entry, a misposting to the wrong account. Each explanation should point to the evidence that supports it.

Propose. Where an explanation implies an adjustment, such as recording the bank charge, the agent drafts the journal at depth 2 for a person to review.

A bank reconciliation is usually the first candidate: high volume, clear rules, sealed data on both sides. The bank reconciliation template and balance sheet reconciliation template give the structure the agent's output should follow, so reviewers see it in a familiar form.

What makes an explanation good?

An explanation is only useful if a reviewer can verify it quickly. A good one:

• Names the cause specifically, such as "cheque 1042 issued on the 29th, not yet presented", rather than "timing difference".
• Cites the evidence: the ledger entry, the bank line, the document.
• States its confidence: confirmed by evidence, likely from pattern, or unexplained.
• Says what will resolve it: the cheque clearing next month, a journal to be posted, a query to the bank.
• Flags anything unusual: an item outstanding for longer than usual, an amount that does not fit the pattern, a counterparty the agent has not seen before.

The confidence field matters most. An agent that labels every explanation as certain is less useful than one that distinguishes "matched to cheque register entry" from "probably a timing difference; no supporting document found". The second gives the reviewer somewhere to look.

Where does the agent stop?

At anything that changes the books or settles a question that needs judgement:

Reconciliation also brings the agent into contact with one of the most dangerous areas of finance. A new or changed bank account appearing in a supplier payment, or a receipt from an unfamiliar account, should be flagged and held, not explained away. The reasons are set out in supplier bank detail changes are where finance agents must stop.

How does reconciliation evidence feed certification?

Every reviewed reconciliation produces labelled data. For each match and each explanation, the reviewer has decided whether the agent was right. Over a few cycles, that becomes a clear measure of accuracy by type of item: simple matches, one-to-many matches, timing differences, errors, unexplained items.

That evidence supports certification directly. A labelled set of past reconciliations, with known correct outcomes, lets the agent be scored gate by gate: grounding (did explanations cite the right evidence?), refusal (did it decline to explain what it could not?), escalation (did it flag unusual items?), and drift (does accuracy hold as transaction patterns change?).

If the evidence is strong, the agent may earn depth 3 for narrow tasks, such as posting routine bank charges below a set amount. More often, the value of the evidence is wider: it shows the finance team how the agent behaves on checkable work, which informs how far to trust it on work that is harder to check.

Why does this matter more than speed?

Agents do make reconciliations faster, sometimes dramatically. But speed is not what builds trust. A fast reconciliation that a controller has to redo because they cannot rely on it saves nothing.

What builds trust is a record of being right, on work where being right can be proven. Reconciliation produces that record month after month. A finance team that starts its agents here learns, on evidence, where they are reliable and where they are not, before giving them anything harder.

Where Pryme fits

In Pryme, bank feeds and ledger entries reach a reconciliation agent as sealed facts through the Pryme Context Protocol, so matches rest on systems of record rather than copies. Reviewed reconciliations can feed Attest, building the evidence for any later depth on the customer's own cases. This is the starting point for AI agents for finance: books you can audit.