Choose a few measures that reflect the task: useful completion, quality, total time, corrections, and the experience of the people involved. Counting generated documents or messages can be misleading if those outputs create more work than they save.
Record a normal example before the trial, then compare similar work during it. Include time spent checking and fixing the result. Ask the people using the output whether it helped. Keep the conclusion narrow: a successful trial on one task is a reason to consider that use, not proof that every process should be automated.
Further reading
- AI Risk Management Framework
National Institute of Standards and Technology · Publication date not established
Context and limits
Guidance on assessing AI risks, protecting information and keeping people accountable for decisions.
The framework was released in January 2023 and its generative-AI profile in July 2024. These are the resources' release dates, not the webpage's publication date. Citing this guidance does not mean TideProof has been certified by NIST.
Sources explain the background. Examples and suggested next steps are TideProof's own guidance.