oprill / original field notes

Decide what evidence would change your mind.

An original planning guide for evaluating workplace AI. Define the task, access boundary and failure cases before comparing demonstrations or discussing a purchase.

1. Write a one-task brief.

Name one recurring workflow and the person accountable for its output. Describe the current input, the desired result and the decision that still requires review. “Help a support reviewer draft an answer from approved help articles” is easier to evaluate than “improve productivity.”

For the first exercise, use synthetic help articles and fictional questions. Write expected answer elements before running a demonstration so a persuasive response cannot redefine success afterward.

2. Choose cases that can fail.

Include a straightforward answer, an outdated article, a contradiction between sources and a question that the source set cannot answer. Add an access case in which the evaluator must not see one of the documents.

For each case, define acceptable behavior in advance. A missing source may require an explicit unknown. A contradiction may require surfacing both versions and asking an owner to resolve them. Do not score confident wording as correctness.

3. Inspect the operational boundary.

Ask where content is processed, what is retained, how access changes propagate and who can suspend the workflow. Request evidence appropriate to the proposed environment rather than assuming a marketing statement establishes the control.

Separate what is available now from what would require implementation. Record unanswered questions and assign them to an owner before widening access. This guide does not make a security or compliance determination for your organization.

4. Record effort as well as output.

During the exercise, record the review and correction work required for each result. Use the same cases to compare approaches. Keep setup effort separate from recurring work, and keep measured observations separate from estimates.

Do not extrapolate a small synthetic exercise into an organization-wide savings claim. If you later use representative workplace cases, document the sample, limitations and any changes to the process.

5. Write the decision memo.

Summarize which cases passed, which failed and what additional evidence is needed. Name the next bounded step: stop, revise the exercise or conduct a limited approved evaluation.

Include an exit condition and a way to stop access. A good decision memo can recommend against expansion; its purpose is to support a justified choice, not to make every trial look successful.