Routine, reversible work with clear evidence: matching bank lines, coding matched invoices, reconciling sub-ledgers, chasing missing documents and preparing schedules. An agent can do these on its own once it has been tested on your books and stays within your limits. Anything that moves money, changes master data or needs accounting judgement still goes to a person.
What makes a finance task safe to hand over?
Three tests, and a task must pass all of them.
Is it reversible without anyone outside noticing? Re-coding an invoice to the right cost centre is an internal correction. A payment to the wrong supplier is not.
Is the right answer evidenced in a system of record? Matching a bank line to an invoice with the same amount and reference is evidenced. Estimating a bonus accrual is not.
Is the volume high enough that review is the bottleneck? If a person reviews ten items a month, keep them. Independence pays off where people are re-checking hundreds of routine items.
A task that passes all three is a candidate for depth 3: execute routine work within bounds. It still has to earn that depth on your own cases.
Which tasks usually qualify?
| Task | Can the agent do it alone? | Condition |
|---|---|---|
| Matching bank lines to ledger | Yes, at depth 3 | Exact amount and reference; others drafted |
| Coding matched supplier invoices | Yes, at depth 3 | Under the item cap, known supplier, known account |
| Sub-ledger to ledger reconciliation | Yes, prepares and posts routine clearing | Differences above tolerance go to a person |
| Chasing missing receipts and approvals | Yes | From a template the owner approved |
| Preparing close schedules | Yes, as drafts | A person reviews before close |
| Variance commentary | Drafts only (depth 2) | A person edits and signs |
| Accruals and estimates | Drafts only | Judgement stays with a person |
| Payment runs | Drafts only | Release stays with signatories |
The fuller picture of depth by capability is in the finance agent depth matrix. The mirror of this article, covering tasks that never leave a person, is which finance tasks should always stay with a person.
Why does "on its own" still mean bounded?
An agent acting alone is not an agent acting without checks. Each action still meets the agent's role scope, its certificate, its sources, your limits and the platform guardrails, and the lowest of those decides. "On its own" means no person approves each item. It does not mean nobody can see, sample or reverse what it did.
Ownership matters here. Someone must answer for the agent's independent work the way a finance manager answers for a team. Our answer on who should own an AI agent in a business sets out why that should be the process owner, not IT.
A worked example
Take a services company's month-end close. Using the month-end close checklist with agent tasks, the controller splits the work.
The agent matches 4,000 bank lines on its own, clearing the exact matches and drafting proposed matches for the 200 that differ by fees or rounding. It reconciles the payables sub-ledger to the control account, finds a 120 difference caused by a credit note posted in the wrong month, and drafts the correcting entry for the accountant. It sends reminders to budget holders who have not approved invoices, using the approved wording.
The accountant spends the close on the 200 near-matches, the credit note and the accruals, which is where judgement was always needed.
What tends to go wrong?
- Spreadsheets as hidden inputs. If a reconciliation depends on a workbook only one person understands, the agent cannot do it alone. Bring the workbook under control first; see letting an AI agent work in your finance spreadsheets.
- Tolerances set too wide. A generous rounding tolerance lets the agent clear items that deserved a look.
- No sampling. Independent work still needs a person to sample it each period.
Which controls make independent work safe?
- Depth granted per task on your evidence. Matching can be independent while accruals are not.
- Tolerances and caps enforced at the moment of action.
- Every action filed. The record shows what was done, under which bound, so sampling is quick.
- Depth that falls fast. If sampling finds errors, the task drops to draft the same day.
The topic guide AI agents for finance: books you can audit explains how this record supports the audit.
Questions people ask
Can the agent post journals on its own?
Routine, recurring journals with fixed logic, such as clearing entries, can reach depth 3. Journals involving estimates or judgement stay as drafts for a person.
How do we decide the tolerance for automatic matching?
Start with zero for amounts and an exact reference, then widen only where your history shows differences are always fees or rounding. Write the reason down.
Does independence ever extend to small payments?
Moving money is depth 4 and needs the strongest evidence. Most finance teams keep payment release with people even for small amounts, because fraud often starts small.