# AI pilot readiness check

Twelve questions across the three phases from pilot to production. Answer Yes, Partly or Not yet to see a readiness band for each phase, the gaps to close first and the right level of human oversight for your outputs.

## Before the pilot

- Can you state the user, task, output and a measurable quality threshold in one paragraph?
- Does the data already exist, does the output feed a workflow people use today, and will value show within 90 days?
- Is every production input traced to a source, a maintainer and a refresh schedule?
- Do you know what it will cost to run for a year at production volume, and what return justifies that?

## During the pilot

- Are you measuring adoption and the commercial outcome, not just accuracy?
- Does each type of output have an agreed level of human review, with the reason written down?
- Is the pilot running with real users, live data and the awkward cases included?
- Have you agreed what result would make you stop?

## Into production

- Is one named person accountable for the system's cost, quality and business result?
- Does each go-live decision have a date and an owner?
- Are support, a fallback and the switch-off of the old process planned?
- Is there a fixed cycle for reviewing quality, oversight and cost after launch?

## Oversight tier finder

- Is the output a judgement a qualified person is accountable for, such as credit, legal or clinical?
- Does the output reach a customer, move money or affect someone's rights?
- Could a regulator, auditor or court ask how it was produced?
- Can a mistake be spotted and reversed before it causes harm?

1. **Automatic, logged.** The AI acts without review. Every action is recorded and can be reversed. Examples: Classifying and routing incoming documents, Tagging records, Filling metadata. Review: Weekly report of volumes and exceptions, with a monthly spot audit.
2. **Sampled.** Outputs go live and a reviewer checks a sample. The sample rate rises if quality drops. Examples: Extracted fields below a value limit, First-pass summaries for internal use. Review: A fixed share checked each day or week against the quality threshold.
3. **Approved before release.** A named reviewer approves each output before it takes effect. The record shows who approved what, on which evidence. Examples: Client-facing letters, Payments or decisions above a limit, Anything a regulator may ask about. Review: Every case, with turnaround tracked as a service level.
4. **Advisory.** The AI recommends. A person decides and records their reasons. Examples: Credit and eligibility decisions, Legal or clinical judgements, Complaints. Review: Every case. Overrides feed back into evaluation.

Guide: [How to run an AI pilot that actually makes it to production](/insights/how-to-run-an-ai-pilot-that-actually-makes-it-to-production)

HTML: https://rodan.io/tools/ai-pilot-readiness
