A year ago, most finance leaders looking at AI in the back office were running small pilots. An invoice-reading model here, a variance-analysis prototype there, and a lot of caution about touching anything that hit the general ledger. The posture in 2026 looks different. Agents now read documents, code transactions, chase receivables, and draft close checklists as a matter of course, and vendors are shipping product with that assumption baked in. What has not moved is who owns the number that gets posted.
Agents doing the work up to the ledger, humans owning the entry into it. That split is quickly becoming the operating model finance teams are settling on, and it is the decision every controller, CFO, and audit committee is being asked to make right now, whether they realize it or not.
Decide What an Agent Is Allowed to Post on Its Own
The first choice is the sharpest one: does an AI agent get to write to the ledger without a human in the loop, or does every posting still cross a person's desk? Enthusiasts point to speed. Auditors point to the fact that a journal entry is a legal record, and once it is in the books, someone has to be able to explain how it got there.
The answer taking shape is a tiered permission model. The agent handles high-volume, low-judgment work end to end, and anything that hits revenue, accruals, reserves, or intercompany requires a review gate. Recent ERP.io coverage on usatoday.com describes exactly this kind of split — agents doing the read-code-match-route work on a vendor bill, with the policy check and approval gate shown in the flow rather than hidden inside the model.
Decide How the Audit Trail Has to Read
An agent that saves your team ten hours a week is worth nothing if the auditor cannot follow what it did. Under SOX, control owners have to be able to validate the outputs of any system involved in financial reporting, and "the model decided" is not a validation.
Grant Thornton's SOX guidance on AI in controls makes the point plainly: AI belongs in the assistant seat, with humans owning the judgment and the sign-off, and every step needs to be reconstructable. You are not buying an automation feature. You are buying an evidence chain.
Decide Where to Spend Your Human Review Budget
Every review gate has a cost. If a controller has to personally sign off on every AP transaction in a month, the agent has bought the team nothing. Skip sign-off entirely and the ledger becomes a black box. The decision is where to spend the finite attention of the people who understand the business.
The pragmatic move is to route review by risk and materiality rather than by volume. Small, repetitive, well-matched transactions can be sampled. Anything above a threshold, anything with an anomaly flag, and anything touching a period-end estimate gets a person. That approach lines up with how Deloitte's CFO Signals survey describes the current mood among large-company finance chiefs — confident enough in AI to deploy it, cautious enough about governance to keep hands on the wheel where it counts.
Decide What the Close Actually Looks Like When Agents Do the Prep
The month-end close is where the line between agent work and human ownership gets tested every thirty days. Agents can pull subledger balances, propose reconciling items, flag unusual variances, draft accrual entries, and lay out the checklist. What they should not do is close the books.
The controller has to decide which prep tasks become agent-owned by default, and which stay human because the judgment call is the whole point. A reasonable split looks like this:
- Agent-owned by default. Pulling and tying out subledger balances, drafting standard recurring journal entries, running reconciliations on high-volume accounts, and assembling the checklist and supporting schedules.
- Agent-proposed, human-approved. Accrual entries, variance explanations, reclasses, and any reconciling item above a materiality threshold. The agent does the work, a person signs the entry.
- Human-owned, full stop. Reserves, revenue cutoff judgments, intercompany eliminations that require policy interpretation, and the final sign-off that the books are closed.
Decide How You Will Explain This to the Auditor and the Board
The last decision is a communication one, and it is the one most teams put off. Boards want to hear that AI is producing value. Auditors want to hear that controls have not weakened. Both groups get nervous when the answer is vague.
Write a one-page policy that says three things: what the agents do, what humans still own, and how you can prove either at any time. Bring it to the audit committee before they ask for it. Bring it to your external auditor before fieldwork. The teams that move fastest on AI in finance are not the ones with the loosest controls. They are the ones who can describe their controls in a sentence and defend them on the next line.
The ledger is still the ledger. Agents can do a great deal of the work that leads up to it, and increasingly they should. The books remain a human responsibility, though, and the finance teams drawing that line clearly right now are the ones who will get the productivity without giving up the trust.
Headlines Team 
Daniel