AI Assistant Workflow Fit Protocol

AI Assistant Workflow Fit Protocol

Evaluation principle

A demonstration establishes that a system produced an output under demonstration conditions. It does not establish permission, reliability, quality, economic value, accessibility, security, or fit for another workflow.

The evaluation should begin with one task and an accepted-work baseline. It should compare the assistant with the current process and a credible non-AI alternative.

Required test record

Define the task, trigger, sources, data authority, user, accepted output, reviewer, ordinary failure, harmful failure, system and version, configuration, representative cases, sample, baseline, time box, measure, stop rule, fallback, and decision date.

Include ordinary, ambiguous, missing-evidence, conflicting-source, restricted-data, unusual-language, accessibility, interruption, outage, and abstention cases as relevant.

Measurement

Measure accepted quality, consequential correctness, completeness, source support, active effort, elapsed time, review, correction, rework, harmful failure, data behavior, user and affected-person impact, accessibility, support, cost, and recovery.

The unit is accepted work, not conversational naturalness, token speed, or time to first draft.

Product-layer boundary

Separate the device, operating system, application, assistant interface, model, voice, connector, local or cloud execution path, data controls, and organizational policy. A product change in one layer can invalidate the result.

Decision

Approve the exact use with controls, constrain and retest, pause pending a named change, or stop. Approval does not extend to adjacent tasks, new data, another account, another model, another voice, or a new connector.

E019's [[How to Choose a First Workplace AI Task]] covers task selection. E042's [[How to Evaluate a Workplace AI Feature]] covers workplace measurement. E018's protocol focuses on separating the capability demo from the workflow and product layers it must fit.