|
EXPLAINER |
AI FOR ACCOUNTANTS · PART 2 OF 16
Why AI agents need approval gates the same way people do
The control is not the agent - it is the human checkpoint that sits between the agent's work and the general ledger.
When a staff accountant prepares a journal entry, someone else approves it before it posts. That is not bureaucracy - it is a basic internal control, and most accounting systems enforce it structurally. An AI agent that can draft, validate, and submit journal entries is doing the same work that staff accountant does. The segregation-of-duties logic that applies to the person applies equally to the agent, and skipping it because the agent 'already checked everything' is exactly the reasoning that breaks controls.
The mechanism matters here. An agent operates by reading inputs, making inferences, and then acting on those inferences - posting, filing, flagging, or moving data. The agent's confidence in its own output is not a substitute for a second set of eyes. In an accounts payable workflow, for example, an agent might match an invoice to a purchase order, calculate the payment amount, and queue the disbursement. Each of those steps is a judgment call. The match could be a near-miss on a vendor name. The amount could reflect a price that changed after the PO was issued. An approval gate after the match and before the disbursement is the point where a human with context - not just data - can catch what the agent cannot know it does not know.
The same logic applies in reconciliation work. An agent reconciling a bank statement to the general ledger will resolve clear matches automatically, which is appropriate. What it surfaces as 'unresolved' is only as good as its matching rules. A competent reviewer does not just approve the resolved items - she looks at what the agent decided not to flag, because that is where the miscategorized entry or the duplicate payment hides. The gate has to be structured so the reviewer sees the agent's reasoning, not just its conclusion. A workflow that shows 'reconciliation complete, approve?' is weaker than one that shows the exception list and the match confidence alongside each resolved item.
Designing these gates is not complicated, but it does require intentional decisions about where in the workflow human review lands. The rule of thumb is the same one that works for staff: the higher the irreversibility of the action, the earlier the gate should appear. Posting to a locked period is more consequential than staging an entry in a draft. Sending a client deliverable is more consequential than updating an internal workpaper. Agents should be scoped to low-irreversibility actions by default, with explicit approval required before anything that would be hard to unwind. That is not a technical limitation - it is a control design choice, and it belongs in the workflow documentation the same way a preparer-reviewer sign-off belongs on a workpaper.
The piece 'Desktop. Cloud. AI. Don't Hedge Your Future.' argues that firms are actively re-evaluating their core technology stacks in a way that sets the course of a practice for years - which makes control design at this stage more consequential, not less. Firms that wire agents directly into posting workflows without approval gates are not moving faster; they are accumulating control gaps that will surface in an audit or a review cycle at the worst possible time. The gate is not friction. It is the control.
WORKED EXAMPLE
In practice
A controller is piloting an agent on the month-end bank reconciliation for a single operating account. The agent reads the bank feed and the GL trial balance, matches transactions, and produces a draft reconciliation workpaper.
What came back. The agent returned a table of the 4 unresolved items with probable GL accounts and a note on each explaining the gap - one was a timing difference on an EFT, one was a duplicate payment candidate, and two were amounts with no corresponding PO. It also surfaced 11 matched items where it had matched on partial vendor name, listing the GL entry description alongside the bank memo text. The duplicate payment candidate was flagged correctly, but the agent had categorized one of the unresolved items as a timing difference when it was actually a voided check that had not been reversed in the GL.
How it was checked. The controller traced each unresolved item back to the bank statement PDF and the GL detail report, confirming transaction dates and amounts against the agent's output line by line before approving the workpaper.
A constructed example. The prompt is usable as written; the figures show the shape of a result, not a measured one.
WHEN TO USE IT
| WHEN NOT TO
|
WHAT TO TAKE FROM THIS
| Scope agents to draft and stage actions only; require explicit human approval before any irreversible post. | |
| Show reviewers the agent's reasoning and exception list, not just a pass-fail summary. | |
| Document agent approval steps in workflow notes the same way preparer-reviewer sign-offs appear on workpapers. |
SPONSORED
QUESTIONS THIS ANSWERS
What is an approval gate in an AI agent workflow?
A structured checkpoint where a human reviews and authorizes the agent's output before the next action executes - analogous to a preparer-reviewer sign-off on a workpaper.
Does an AI agent replace segregation of duties controls?
No. An agent is a preparer, not an approver. The same segregation-of-duties logic that applies to a staff accountant applies to any agent that can initiate, modify, or submit accounting entries.
Where should approval gates land in a reconciliation workflow?
After the agent produces its exception list and match results, before any resolved items are written back to the general ledger. The reviewer should see reasoning and confidence, not just a completion status.
SOURCES
Where this comes from
What the accounting job market is actually asking for.
GO DEEPER
Go deeper
IN THIS SERIES
Previously: What a model context window means for a workpaper
Next: What an MCP server is and why vendors ship one (coming)
An explainer, not a study: it carries no statistics on purpose. Examples are illustrative.
How accountants are using AI, automation and smarter workflows to close faster, audit cleaner, and free up time for real work.
Accounting Stack · Audit Friendly Data · Remote A&F Jobs
Audit Friendly · modernaccounting.ai