--- title: "The exception is part of the workflow" description: "Plan for delivery discrepancies with clear failure states, useful escalation records, and a manual recovery path that keeps authority with people." slug: "the-exception-is-part-of-the-workflow" url: "https://withlodge.com/writing/the-exception-is-part-of-the-workflow" published: true published_at: "2026-09-22T12:00:00.000Z" updated_at: "2026-11-02T15:18:45.883Z" image_url: null image_alt: null image_width: null image_height: null word_count: 771 --- A delivery discrepancy deserves its own path through the business. If the ordinary process expects a matching order, dispatch record, and receipt, an agent needs instructions for what to do when those records disagree. Leaving that path undefined pushes the decision onto whoever notices the problem first. Imagine a hypothetical distributor using an agent to reconcile delivery paperwork. A receiving record says fewer cartons arrived than the dispatch note lists. The agent can prepare a discrepancy record, but a warehouse supervisor decides what was received and an authorized commercial owner decides any customer adjustment. The exception crosses those responsibilities without erasing them. ## Name the state before choosing the response A missing document, contradictory evidence, unavailable system, denied permission, and uncertain action outcome are different states. Each needs a visible label and a defined next step. A generic failure message gives the next person very little help. In the discrepancy, missing proof of receipt calls for someone to obtain the document. Conflicting counts call for a receiving check. A denied request to read the receipt record calls for an access decision by the appropriate person. Repeating the same action cannot resolve all three. Uncertain outcomes need special care. If a request to create a discrepancy record times out, the record may already exist. The next step is to inspect the destination through an authorized method and establish what happened. Repeating the write blindly could create competing records for the same delivery. Define which states stop further activity. The agent should never interpret a blocked step as permission to use another person's access, search a broader collection of records, or approve its own proposed adjustment. ## Give the next person a usable case An escalation should explain what decision is needed. Include the delivery identifier, the conflicting observations, their permitted sources, the relevant times, and what the workflow has already attempted. Distinguish a verified fact from an inference and from information that remains unavailable. For this delivery, the warehouse supervisor needs to see which count came from receiving and which came from dispatch. The supervisor also needs to know whether the workflow merely drafted a discrepancy or actually created one. That detail changes the recovery work. Assign the case to a named role with an available person behind it. Specify when an unanswered case should receive further attention and who can take over. Sending a notification without assigning responsibility leaves the business waiting for a volunteer. Keep the escalation within the recipient's access. A convenient summary can reveal restricted material just as readily as an attached document. When a reviewer lacks permission, route the decision to an authorized person or prepare an approved, limited account of the issue. ## Make manual recovery a normal operation People need a documented way to finish interrupted work. Explain how to inspect current state, identify any action already taken, preserve the evidence, and record the resolution. The recovery path should be usable by someone who did not build the automation. In the distributor, a supervisor might confirm that the receipt count needs correction. A different authorized person might need to decide how to handle the remaining shortage. The agent can prepare the supporting material within its grants. Each person retains responsibility for the decision belonging to their role. Record who resolved the exception and what happened in the business systems. If the agent resumes, it should start from that confirmed state. It should not reconstruct the outcome from a conversational hint or repeat every step from the beginning. A retry can be appropriate for a known temporary failure when the action's rules permit it. Define the limit and what happens after it. While the case waits, show that it is unresolved rather than leaving an apparently successful record behind. ## Turn exceptions into reviewed improvements Closed cases can reveal a recurring missing field, an ambiguous supplier document, or an instruction that needs revision. Review those patterns with the person responsible for the process. Some call for changing the source process; others justify changing the workflow. A single recovery does not automatically establish a new general rule. The person resolving this shortage may have relied on circumstances that do not apply elsewhere. Capture that reasoning and test a proposed rule against other permitted examples before adopting it. Share lessons only with an authorized audience, removing details the wider group does not need. Keep the original case's access restrictions intact. Take an exception your team expects to encounter and write its recovery record in advance. Include the state, evidence, owner, next decision, and restart condition. If a colleague cannot use that record to continue the work, the exception path needs another pass.