Pick one recurring job
Choose a task with a clear beginning and end: a coordinated set of visual assets, a script-led video or an edited interview. “Make marketing better” is too broad to evaluate.
Write a small brief
Define the audience, input materials, desired output, tone and constraints. Include an example of work that meets your standard. Keep confidential material out of early trials.
Decide what success looks like
- What needs to be correct for the result to be usable?
- How much human correction is acceptable?
- What is the maximum cost per usable result?
- What requires approval before an action can happen?
Shortlist two or three candidates
Compare products in the same job context. A video creation agent and an editing agent solve different problems, even when both produce video. Check inputs, outputs, plan restrictions and data policies.
Try a contained pilot
Use the same brief across repeated runs. Keep the failures. Record the effort required to turn each output into work you would actually share. Expand only when the result is useful and repeatable.