Designing an exception queue for work an AI agent cannot finish

Every useful agent will encounter cases that require a person. Put those cases in a queue someone already watches, and make each item answer three questions. What stopped the work? Who can fix it? What should happen after the fix?

Show the smallest useful packet

For a company that checks supplier invoices against purchase orders, a mismatch could arise from quantity, currency, delivery timing or a missing order number. “Agent failed” sends an accounts payable clerk on a new investigation. “Invoice shows 18 units, purchase order shows 12, receiving record unavailable” gives the clerk a specific discrepancy to resolve.

Include references to the original records, a short reason, the proposed next action and a link to the approved system. Avoid pasting complete documents into channels where people lack access to the originals. Let the reviewer choose a clear outcome such as request information, correct the mapping, mark duplicate or pass to a supervisor.

Decide how a resolved case resumes

A reviewer might supply the missing purchase order. The workflow should return to the comparison step with the correction recorded. If it restarts from the beginning, check that it cannot create a second case or reissue an earlier notification. Give closed exceptions a visible outcome and timestamp.

Track the oldest open item and the most common reason each week. If half the queue consists of missing order references from one supplier, fixing that supplier’s submission form may outperform more model tuning. Make one person responsible for that follow-up.

Before launching, ask a new reviewer to resolve three test exceptions using only what the queue provides. Each time they have to ask the builder what to do, add the missing information or simplify the handoff.

Put this into practice

Bring a sample of the work and the tools your team uses. We can help define a first project, who reviews it, and how to tell if it works.

Explore AI implementation · Discuss your project with Escencion