AI harness accounting refers to the environment where agents are built, deployed, reviewed and improved. They complete each line of the close, from reconciliations through accrued liabilities and commissions, returning a tied-out workpaper with its support attached for approval.
It is day three, and the checklist says intercompany is due. Nothing on that line says who will work the forty unmatched items, decide which ones are timing, fix the supporting schedule, and prepare the entry that clears the remainder. That job belongs to a person, and it usually runs past dinner.
AI harness accounting begins at the second sentence, with the work a due date assigns rather than with the tracking of it. The distinction matters because most teams have already solved the coordination half, and the calendar has barely moved. Owners are named, dependencies are visible, and the heavy preparation still happens in a spreadsheet nobody else opens.
What follows walks a period end in the order a controller works it, from the first export out of the ERP to a signed balance sheet, and names what gets completed at each stop. The ideas behind agent-led reconciliation run underneath all of it.
What Close Software Was Built to Do and Where It Hands Back
Close platforms standardize a checklist, assign owners, track dependencies, and report what is late. That is real value, and a large department coordinating on a shared spreadsheet would struggle without one. Anyone who has replaced that tracker with a proper system knows the improvement lands in the first month.
The limit shows up in what happens after the alert. A platform can report that the bank tie-out has not been signed, and resolving it requires someone to pull the statement, investigate the variance, and decide on treatment. The status record is accurate the entire time, while the underlying accounting sits exactly where it was.
That handoff repeats at every line with judgment in it. Standardized, repeatable steps stay inside the tool. Exception-heavy preparation, anything drawing on data outside the ledger, and anything whose treatment depends on last period all move back into Excel.
Reconciliation Splits Into Two Very Different Jobs
Controllers tend to say "reconciliations" as though it names one activity, and the close calendar disagrees. Two accounts can carry the same due date, the same preparer, and completely different failure modes, which is why a single automation rule rarely covers both.
Tie-Outs Against Schedules the Team Maintains
Prepaids, fixed assets and deferred balances are checked against a supporting schedule the department owns. Take a $1,000 computer depreciated over ten months, producing a $100 charge each period. Add an asset or dispose of one, and the amortization table, the monthly charge, and the accumulated balance all move together.
An agent rebuilds the support from the asset register, ties the ending figure to the general ledger, explains any variance, and prepares the journal entry for review. The preparer opens a tied-out workpaper with documentation attached, not a task marked ready.
You might also like: Agentic Reconciliation: How Preparation, Reconciliation, and Close Run as One Process
Matching Against Records a Third Party Controls
Bank and intercompany balances behave differently, because the counterparty controls half the evidence. Timing differences, partial settlements and truncated reference numbers create exceptions that no rule anticipated, and volume makes manual clearing slow.
Resolution here has three legitimate paths: correct the workpaper, post an entry, or do both. An agent that stops at "unmatched" has finished the easy portion. One that evaluates each item, applies the policy the team wrote down, clears what it can, and escalates the rest with its reasoning attached has finished the account.
Accrued Liabilities Turn on a Branch in the Treatment
Accruals are where standardization usually fails outright. Services are received in one month, the invoice lands in the next, and the correct treatment depends on whether it is expected or already in hand. Those two states carry different entries.
The reversal is the second half, and it is where errors surface two periods later. When the real invoice posts, the estimate has to unwind cleanly against it, and a duplicate expense in the following month is the usual symptom of a reversal that landed wrong.
Encoding that branch takes the policy the accounting department applies: which vendors accrue, what evidence counts as expected, what threshold is worth booking. Written down once, that logic runs every period, and the figure arrives with the support behind it.
Commissions Pull Data the Ledger Never Sees
The commission waterfall is the canonical example of work that could never be standardized. Amounts depend on deal records in the CRM, plan terms in a signed document, and amortization treatment in the general ledger, so the calculation lives in a workbook one analyst built and everyone else trusts.
Every change ripples. A deal closes late, an amount gets revised, a rep leaves, and the schedule has to be recalculated before the entry can be prepared. The workbook and the system of record then get updated separately, which is how two versions of the same number end up in circulation.
An agent ingests the deal file, recalculates amortization, updates the worksheet and the ERP in the same pass, and leaves a changelog behind. What the reviewer approves is one figure with one history.
Analysis Only Works Once the Balances Hold Still
Flux and variance review is scheduled last and compressed hardest, because the numbers keep moving until the final entry posts. Most teams sample the largest movements, write an explanation from memory, and move on, which leaves the rest of the population unexamined.
Testing the full population changes the reach of the exercise. An explanation written from a sample covers only the movements someone chose to examine, while reading every line can surface a driver nobody had flagged.
Flux produces no journal entry. The artifact is a narrative traced to the transactions underneath it, which is precisely what the audit request will ask for in four months.
Where AI Harness Accounting Fits Between the ERP and Sign-Off
Every stop above describes a gap between a system of record that holds balances and a close that someone has signed. Nominal is the environment where accounting agents covering that ground are built, deployed, reviewed and improved, and solutions are named for the activity they perform, as in Reconciliation, Accrued Liabilities, and Commissions.
Two conditions govern execution:
- A written procedure comes first, since the logic has to be grounded in how one company treats an activity rather than how the category generally treats it. Nominal's finance engineers encode it, and the controller's team reviews and approves the result.
- Approval is the second condition. Nothing reaches the ledger without a person signing it, and every action carries its evidence, an approach described in more depth in human-in-the-loop governance.
What Stays With the Controller
Plenty does: materiality judgments, the treatment of a genuinely novel transaction, conversation with the auditor, decision to change a policy, and final sign-off on every number all remain human. Agents work inside the procedure they were given, and changing it is an accounting decision.
What changes is the ratio. Reviewing prepared work with its support attached takes a fraction of the hours that preparing it consumes, and that is the part which genuinely requires ten years of technical experience. A department spending days one through five on preparation and day six on review can invert that shape.
Teams that already run month-end close automation across the coordination layer tend to see the change fastest, because due dates are reliable and the bottleneck has moved to the work underneath them.
A Close That Finishes on Its Own Schedule
An eight-day month end is rarely eight days of technical judgment. Most of it is preparation, chasing and formatting that expanded to fill the calendar. AI harness accounting compresses that surrounding portion by completing it, which leaves the hard calls sitting with the people qualified to make them.
The version worth aiming for is a department where capacity stops scaling with transaction volume, senior people spend the month on analysis rather than tie-outs, and audit support is assembled as the work happens. That is available now, one account at a time, starting with whichever line costs the most today.
See how accounting agents run against a real close. Book a demo with Nominal and walk one of your own accounts through, from the export out of the ERP to a signed workpaper.
About the writer
