AI capability

What happens when it is not sure

Everyone evaluates agents on what they do when they are confident. The property that actually determines whether automation works in a finance function is what happens the rest of the time — because that is the part that lands on a person, and a bad escalation costs more than doing the work.

Act inside rules your controller wrote.6 of 12 actions permitted

Answer a question from your dataread onlyL0
Run a report or reconciliationread onlyL0
Draft a vendor billheld for reviewL1
Draft a journal entryheld for reviewL1
Send an invoice reminderinside dunning policyL2
Post a coded bill under thresholdmatched PO, within toleranceL2
Post a bill over thresholdneeds named approverL3
Schedule a payment runproposal only, never releaseL3
Release fundshuman only, any amountnever
Close an accounting periodhuman onlynever
Reopen a closed periodhuman only, loggednever
Change its own permissionsadministrator onlynever
Routed by rule, not by queueContext handed over, not a linkDecisions fed back

What it does

Six things, specifically.

Routed by rule

To the person authorised for this decision at this amount in this entity, not into a shared queue that whoever is least busy drains.

Full context attached

The document, the contract, the vendor history, the budget, the candidates considered, and why each was rejected — so the reviewer decides rather than investigates.

The reason it stopped

Low confidence, a policy boundary, a conflicting signal, or missing information — each is a different problem and stating which changes what the person does.

Deadlines and reassignment

An escalation nobody actions is a stalled transaction. Time limits, reassignment, and visibility of what is waiting on whom are part of the mechanism rather than a report.

The decision is fed back

What the person chose, and where they overrode, becomes evaluation data. Escalations that recur identically are a configuration problem rather than a workload.

Rates are reported

Escalation rate by workflow, by vendor, by entity, tracked over time. A rising rate is a signal about the data, not about the agent.

Escalation is where automation is won or lost

Consider a workflow running at eighty-five percent straight-through. The fifteen percent that escalates is now the entire job of whoever receives it, and it is the hard fifteen percent — the ambiguous, the unusual, and the genuinely wrong.

If those arrive as a link to a record and a note saying review required, the person does the investigation from scratch and the automation has saved nothing on the cases that cost the most. Escalation quality, not straight-through rate, is what determines whether the numbers translate into hours.

Straight-through rate measures the easy work. Escalation quality decides whether the hard fifteen percent got cheaper or just moved.

Naming why it stopped

The four reasons are genuinely different. Low confidence means the agent had the information and could not decide. A policy boundary means it decided and is not permitted to act. A conflicting signal means two reliable inputs disagree. Missing information means the answer is not in the system.

Each implies a different action, and lumping them into one exception queue means the person works out which every time. Stating it is close to free and it is the single change that most reduces handling time.

Routing to the authorised person

A shared exception queue routes work to whoever is available, which means decisions are regularly made by people who are not authorised to make them and are then rubber-stamped afterwards.

Routing by the same authority model that governs everything else means the person who receives it can actually decide it. That is faster and it is the difference between a control that operates and one that is documented.

Recurrence is a signal

The same escalation appearing weekly is not a workload problem. It is a missing policy, a vendor record needing attention, or a threshold set wrong.

Recurring escalations are grouped and surfaced as configuration suggestions rather than delivered individually forever, which is how the escalation rate falls over time rather than plateauing.

Limits

Where it does not help.

Every capability page on this site carries one of these, because a feature described without its boundaries is a claim rather than a description.

It cannot escalate to an undecided owner

If nobody is authorised for a decision, escalation has nowhere to go. That gap is an organisational one and configuration surfaces it rather than solving it.

Some escalations will always be wrong

An agent that never escalates unnecessarily is an agent escalating too little. We tune toward over-escalation early and pull it back on evidence.

Context has a ceiling

Where the missing information lives in somebody’s head or in an email thread, no amount of context assembly finds it. The escalation says so rather than guessing.

Questions

What people ask.

What escalation rate should we expect?
It is the inverse of the straight-through rate for that workflow — roughly nine percent on PO matching, thirty-one on multi-line allocation. Our benchmarks publish both.
Who receives an escalation?
The person authorised for that decision at that amount in that entity, routed by the same model that governs everything else rather than into a shared queue.
What is attached?
Document, contract, vendor history, budget, the candidates considered, and why each was rejected — so the reviewer decides rather than investigates.
What if nobody actions it?
Time limits, reassignment, and visibility of what is waiting on whom. A stalled escalation is a stalled transaction.
Does the escalation rate fall?
Yes, if recurring escalations are treated as configuration problems. Grouped and surfaced as suggestions is how that happens rather than by tuning the model.

Look at the fifteen percent.

Tell us how exceptions are routed today. That is usually where the hours actually are.