Accounts payable · Exception review
Automation should stop before it guesses.
Holdfast handles the reconciliation exceptions automation cannot safely resolve. It applies typed payment holds, assembles field-level evidence, and asks a human for the one thing a model should not decide: who acts next.
- Payment held
- Field evidence
- Routed to an owner
No approve/reject. No confidence score. No model prose.
The problem
Reconciliation tools compete on how much they clear. The harder question is what happens when they are wrong.
- A wrong match is worse than an unmatched transaction.
- An unmatched item stays visible and gets worked. A wrong one settles quietly and is found much later, if at all.
- Reviewers should inspect evidence, not model prose.
- Fluent explanation invites agreement. A field, its reference, and the delta between them invite a decision.
- Exceptions need routing, not approve/reject.
- A binary control forces the real work into spreadsheets and email, outside the record entirely.
How it works
One exception, from reconciliation to a recorded decision.
- 01
Reconcile
Invoices are matched against payments and purchase orders.
- 02
Apply typed hold
Where a condition is breached, a typed hold blocks payment.
- 03
Assemble evidence
Every compared field is captured with its reference value.
- 04
Route next action
A reviewer chooses the resolution path and the next owner.
- 05
Record control changes
A tolerance change is written down like any other decision.
- 06
Preserve the trail
Each action lands on an append-only timeline for the case.
The workstation
Three panes, and no screen the reviewer has to leave.
An analyst reviewing a held invoice normally rebuilds context from five or six screens. Holdfast puts the queue, the evidence, and the decision in one place.
Intake
Money-at-risk queue
Diagnostic ledger
Field-level evidence
Why Holdfast
Typed holds
A held invoice cannot progress as though nothing happened.
Holds are typed, not a boolean flag, and they carry the amount they are blocking. Releasing one takes a named reviewer and a stated reason, and the release is recorded.
Field-level evidence
Observed values, reference values, and the delta between them.
The reviewer reads the comparison the engine actually made — what the record said, what it was checked against, and by how much it differs. There is no narrative summary to agree with.
Recorded control changes
A tolerance change is a decision, not a setting.
Widening a threshold records what it was, what it became, the scope it applied to, who changed it, why, and which holds the change affected.
Control philosophy
The model may propose.
Deterministic code verifies.
A generated suggestion is a nomination, not a verdict. It is scored by the same deterministic rules as every other candidate, and it is discarded if it does not clear on its own merits. No generated output can release a hold or clear a transaction by itself.
Automation can assemble the evidence.
Accountability stays explicit.
Review the exceptions automation shouldn’t guess on.
The workstation opens on the demo run, ordered by money at risk.