Back to the notes

Designing AI automation around its exceptions

A workflow connecting email, spreadsheets, and a CRM also needs to be judged by what happens when information is missing or contradictory.

A branching workflow model diverts an irregular item to a separate inspection loop

The work beyond the connection

When designing an assistant or automated workflow, I suggest mapping exceptions alongside the normal path: an incomplete request, a duplicate message, or an unavailable system. Those cases shape the connections as much as the successful path does.

Work through one example

In a proposed scenario, an incoming message first becomes a draft request. Missing fields stay visible, and the record is completed after customer and date details are checked. If processing stops halfway through, a duplicate check should be considered before retrying.

Treat possible errors as a design input

NIST’s Generative AI Profile describes how systems can present incorrect information confidently. A workflow therefore needs clear ways to inspect sources, send uncertain cases for review, and correct results. Its scope can grow within a process where the team can see what happened and intervene when necessary.