# Include the person who resolves the AI exception

Automation cost includes the work handed back to staff. Measure whether an exception arrives with enough evidence to resolve it efficiently.

By Cobnex editorial. Published 2026-09-10. Updated 2026-09-11.

## Follow a failed task past the model

An assistant cannot reconcile an invoice and sends it for review. The model usage may be small, but a person now has to identify the document, inspect the fields and decide what to correct.

If the exception contains only "validation failed", the reviewer repeats much of the original work. If it includes the conflicting values and source locations, the same task may be resolved quickly. The design of the handoff affects cost as much as the model's initial price.

Measure the full route from automated attempt to accepted business result. Keep waiting time and active human effort separate because they suggest different improvements.

## Categorise exceptions by cause

Distinguish missing evidence, ambiguous input, policy denial, model error and downstream failure. Some cases need a user answer, some need an authorised reviewer and some need engineering repair.

Routing every case to the same queue wastes expertise and delays resolution. A finance reviewer should not have to diagnose a connector timeout, while an engineer should not invent the meaning of an ambiguous invoice adjustment.

Track how often the recipient can resolve the case from the supplied package. Repeated requests for basic context indicate a product gap that can often be fixed without changing the model.

## Compare configurations with human effort included

A cheaper model that sends more difficult cases to review may increase total cost. A more expensive model may still be unjustified if it does not reduce meaningful intervention.

Use observed handling time by exception category and clearly labelled assumptions for any monetary estimate. Avoid assigning a precise saving to unmeasured work merely because the spreadsheet needs a number.

Include rework after false acceptance. Errors discovered downstream can involve more people than the original review queue and may have consequences that deserve separate reporting.

## Improve the handoff before raising automation targets

Provide the unresolved question, relevant evidence, proposed result and reason for escalation. Preserve completed work so the person does not start again. Keep sensitive data limited to what the recipient is authorised to inspect.

Review whether common exceptions can be prevented through clearer input requirements, better source extraction or deterministic validation. Increasing automatic acceptance without understanding those cases can move cost from review into correction.

The useful measure is cost per correctly completed task across both machine and human work. That view makes it easier to choose an automation boundary that saves effort without hiding the work it transfers to someone else.

## Sources

- [Microsoft Learn: evaluation and observability](https://learn.microsoft.com/en-us/azure/foundry/concepts/observability)
- [OWASP: authorisation guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html)