# Price the exception queue before buying an AI automation

An automation demonstration can process a tidy invoice in seconds. The business decision is harder: what happens to the invoice that arrives twice, lacks a purchase-order reference, or names a supplier whose bank details have changed?

Those exceptions create work for someone. Before buying an AI workflow, describe that person's queue as carefully as the successful demonstration. Otherwise, the automation may move effort to a less visible place instead of removing it.

## Count the work that remains

Consider a hypothetical team handling 400 documents in a month. Assume its current handling takes six minutes per document, or 40 hours. These are chosen example inputs, not industry averages.

Suppose a proposed process needs a one-minute review for each document, while 80 exceptions need another five minutes each. That leaves about 13.3 hours of review and exception handling. The arithmetic suggests roughly 26.7 hours of handling capacity before ongoing maintenance, oversight and other costs.

That isn't automatically a cash saving. Staff may still be employed for the same hours, and the freed time may or may not have a valuable use. Actual exception rates and handling times must be measured. A sample that excludes difficult documents cannot establish the performance of the real queue.

## Give exceptions a destination

For each failure, specify who sees it, what evidence they receive and what they are allowed to do. Missing information should have a visible state. A workflow that silently drops an item can make a dashboard appear healthy while work accumulates elsewhere.

A useful review screen might show the source document, extracted fields, validation problems and the proposed action. Keep the distinction between “the model suggested this value” and “an authorized person approved this action.” Avoid copying sensitive documents into unrelated tools for convenience.

If the exception needs a person with purchasing authority, routing it to a generic inbox doesn't complete the design. Someone must own that inbox and have a way to resolve, reject or escalate the item.

## Decide which actions can happen automatically

Extracting a draft description is different from changing supplier payment details. Route high-consequence or uncertain actions through a defined approval process. Decide who may change the rules and how those changes are reviewed.

NIST's voluntary AI Risk Management Framework organizes risk work around governing, mapping, measuring and managing AI risk. It doesn't certify a particular workflow or provide a universal acceptable error rate. Use it as a reference for the risk questions your actual process requires.

Source: [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework).

## Test failures deliberately

Agree a representative, authorized evaluation set with the operator. Include duplicates, ambiguous records, unavailable downstream systems and interrupted runs. Describe expected behavior before running the evaluation.

For a duplicate invoice, the expected result might be a held item for review rather than a second payment instruction. For an unavailable destination, it might be a visible retryable state with an accountable owner. The appropriate behavior depends on the real system's contract and business risk.

An AI model's confidence label doesn't replace this evaluation. Ask whether a wrong value can reach a consequential action, what prevents it and how an operator notices the failure.

## Buy a decision before buying broad autonomy

A first engagement can be limited to one document type or one internal handoff. Agree the input, exclusions, approval boundary, exception destination, operating costs and acceptance evidence. Include a decision to stop or retain the existing process if the pilot doesn't justify expansion.

StackEndBox's [AI automation scope](https://www.stackendbox.com/services/ai-automation-services#automation-handover) covers implementation and handover questions for that conversation. A deterministic rule, an existing platform feature or a simpler form may be enough. Choosing a smaller intervention is a valid outcome.

Ask the provider to demonstrate the exception queue as well as the successful run. That's where you can see who will operate the system after the presentation ends.
