Review every entry
Not a sample. Each journal entry checked against the pattern for that account, vendor, department, and period — which is a review depth no human team performs at volume.
AI agents · finance
It does the review pass nobody has time for: every journal entry checked against pattern, every schedule prepared, every variance investigated to the transaction that caused it. It stays at Level 1 by design — this is judgement work, and the agent’s job is to bring you the judgement already framed.
What it does
Each of these is work a person does today. The agent does them in sequence and stops at the first thing it is not confident about.
Not a sample. Each journal entry checked against the pattern for that account, vendor, department, and period — which is a review depth no human team performs at volume.
An amount outside pattern, an unusual account pairing, a duplicate, a posting to a rarely used account, or a manual entry where an automated one was expected — each with the comparison that triggered it.
Prepaid amortisation, accruals, deferred items, and depreciation calculated and staged for review before the period ends rather than after.
AR and AP detail against control accounts continuously, so a break surfaces on the day it occurs rather than during the close.
Period-over-period and budget movements resolved to the specific transactions responsible, with a draft explanation you edit rather than write.
Tracks what is done, what is blocked, and who owns it, and follows up without anyone having to feel awkward about following up.
Authority
Reviewing, comparing, and reporting change nothing and need no approval. Most of what this agent does lives here.
Schedules, accruals, and adjusting entries are drafted with their supporting calculation and wait for a person to commit.
Closing a period is an assertion by a person. It is absent from the permission model entirely rather than defaulted off, at any authority level.
We are asked reasonably often whether the Controller Agent can be raised to Level 2 for routine adjusting entries. It can technically, and we advise against it for a reason worth stating plainly.
The value of a review control is that a second party looked. An agent that both prepares an accrual and posts it is not a review — it is the same party doing both halves, which is exactly the segregation problem that made the control necessary. Keeping it at Level 1 preserves the thing you were trying to buy.
A human controller reviews by sampling and by instinct — the entries that look odd, the accounts that have caused trouble before, whatever fits in the time available. That is a reasonable strategy under a time constraint and it misses things systematically rather than randomly.
Reviewing every entry against its own history changes the shape of what gets caught. The errors it surfaces are usually not dramatic: a recurring vendor coded to a different account this month, an accrual that reversed twice, an intercompany entry with no matching side. Individually small, collectively the difference between a clean close and a restatement conversation.
It is not a fraud detection system and we do not sell it as one. It flags what is unusual relative to your own history, which catches error far more often than intent. Deliberate fraud that is designed to look normal will look normal to it, and any vendor claiming otherwise is overselling.
Questions
A month of journal entries is enough to show you what it flags, and for you to judge whether it is right.