How to clear suspense account balances with an AI agent

Treat each suspense item as a small investigation. The agent breaks the balance into its individual postings, traces each back to its source, works out where it belongs, and drafts the reclassification journal with the evidence attached. A named accountant approves and posts. Write-offs, and items the agent cannot trace, stay with a person. The agent also reports why items keep arriving.

What is a suspense account, and who owns clearing it?

A suspense account holds postings that could not be allocated when they were made: a bank line with no matching record, a payroll element with no mapping, an intercompany charge nobody recognised, a system interface that rejected a code. It is meant to be temporary. Left alone, it becomes a store of unexplained amounts that auditors ask about and nobody can explain.

The financial controller owns the suspense accounts and the clearance policy, including how old an item may get before it is escalated. A named accountant clears each account. The owners of feeding systems, such as treasury, payroll and the interface team, fix root causes. The wider finance position is in AI agents for finance: books you can audit.

How does the agent clear the balance, step by step?

1. Agent, depth 1: breaks the balance into items. Every posting into suspense not yet cleared, with date, amount, source system and description. Evidence: the item list reconciled to the balance.
2. Agent, depth 1: traces each item to its source. The bank statement line, the payroll run, the interface log, the intercompany invoice. Evidence: the source record linked to each item.
3. Agent, depth 2: proposes where it belongs. The correct account, cost centre and, for receipts, the customer. It states the evidence and how sure it is. Evidence: the proposal and its support.
4. Agent, depth 2: groups items that clear each other. A posting and its reversal, or two halves of a split payment. Evidence: the pairing and why.
5. Agent, depth 2: drafts reclassification journals, one per item or per group, with the evidence attached and the suspense balance after posting.
6. Person, accountant: reviews, approves and posts. Evidence: the posting with the approver's name and the draft it came from.
7. Agent, depth 2: prepares the residue. Items it could not trace, with what it tried, and items past the age limit. Evidence: the residue list.
8. Person, controller: decides the residue. Further investigation, escalation to the source owner, or write-off. Evidence: the decision and its authority.
9. Agent, depth 2: drafts the root-cause note. Which sources keep feeding suspense and why, for the owners to fix.

Why does the agent draft but not post?

A reclassification moves an amount from one account to another in the books. If it is wrong, the error is now in a real account, where it is harder to spot than in suspense. Suspense at least announces that something is unexplained. A confident wrong reclassification hides it.

That is why drafting first is the default for finance agents. A person reviewing a well-evidenced draft takes a minute; finding a wrong posting at year-end takes far longer. Writing off a residue is a further step, because it removes the item from view entirely.

"Clearing" here means resolving with evidence, not making the item disappear. The same word is used for alerts in compliance, with the same discipline: see can an AI agent clear a sanctions alert.

Which controls matter, and what evidence do they leave?

Item list reconciles to the balance. If the items do not add up to the balance, something is missing before work starts. Evidence: the reconciliation.

Evidence on every draft. No journal without a linked source. Evidence: attachments on each draft.

Separate drafter and poster. Evidence: the agent named as preparer, the accountant as approver.

Age limit and escalation. Evidence: item ages and escalations against the controller's policy.

No write-offs by the agent. Evidence: refused verdicts on any write-off attempt.

What should you measure?

• Suspense balance and item count, at each month-end.
• Ageing: items by age band. Old items are the audit risk.
• Drafts approved unchanged, which shows the quality of proposals.
• Items arriving per source, and the trend after root-cause fixes.
• Write-offs, by value and reason.

Where does it usually go wrong?

Netting hides problems. The balance is near zero because large debits and credits offset. Clear by item, not by balance. A small suspense balance can still hold large unexplained items.

Receipts are mixed in. Unidentified customer receipts often land in suspense. They belong in the cash application process, with its own rules; see clearing unapplied cash.

Customer consequences are triggered. An unidentified payment sits in suspense while the customer is chased, or even put on stop, for an invoice they have paid. Before any account action, check suspense for likely matches; the stop decision is discussed in can an AI agent put a customer account on stop.

The same root cause every month. An interface keeps rejecting one code and nobody fixes it. The agent's root-cause note only helps if a source owner is required to respond.