Ideas to take with you.
- 01Start with a specific workflow, not a model.
- 02Keep a human accountable for consequential decisions.
- 03Measure failure modes alongside time saved.
A convincing demonstration is easy to mistake for a dependable workflow. This original demo essay considers a more useful starting point: write a job description for the task you want AI to perform.
Define the work
Specify the input, the intended output, and the point at which a person checks the result. A task such as “help our team with research” is too broad to evaluate. A task such as “summarize these approved source documents and attach a citation to each material statement” creates a clearer test.
Set the boundary
Identify information the system must not receive and decisions it must not make alone. Decide how the workflow handles missing evidence. An explicit “I could not find this” is often more useful than a polished guess.
- Use a representative set of tasks for evaluation.
- Include ambiguous and incomplete inputs.
- Record factual errors and unsupported claims.
- Keep a named person responsible for approval.
Evaluate the whole process
Measure the time spent checking and correcting output as well as the time spent generating it. A tool that moves work into a difficult review stage may not improve the process.
Start with a reversible, bounded use case. Document the conditions under which the team should pause it. Revisit those conditions as the product and the work change.
This is a general editorial exercise. It does not recommend a specific vendor or claim that a fictional company achieved a measured improvement.
Sources & transparency
This fictional editorial exercise contains no reported interviews or financial claims. No external sources are presented as verification of an invented event.
How we workSee an error? Tell the editorial team.



