Short answer: A business is ready when it can name the task, baseline, accountable reviewer, permitted data, failure limits, and decision that follows the pilot.
Readiness starts with a bounded business problem
“Use AI” is not a pilot objective. Name one repeated task, who performs it, what enters the task, what acceptable output looks like, and what happens next. A useful candidate has enough volume to matter and outputs that a responsible person can review.
Avoid beginning with decisions that can materially affect employment, credit, health, safety, legal rights, or access to essential services. These require deeper governance and specialist review than a first small-business pilot.
AI-readiness scorecard
Score each area as ready, partly ready, or not ready.
| Area | Ready means… |
|---|---|
| Business problem | The task, volume, delay, and desired improvement are written down |
| Accountable owner | One person owns the workflow, review, correction, and shutdown decision |
| Baseline | Current time, cost, error, quality, and exception levels are measured |
| Inputs | Representative examples exist and their use is permitted |
| Review | A qualified person can check every important output during the pilot |
| Risk limits | Prohibited data, unacceptable failures, and escalation paths are explicit |
| Integration | The pilot can run without uncontrolled access to production systems |
| Decision | Success, failure, pause, and next-step conditions are agreed in advance |
Check the data before choosing a model
Identify where examples come from, whether they contain personal or confidential information, who may access them, and whether they are representative of difficult cases. Remove data the task does not need. Confirm provider storage, training use, location, retention, deletion, and contractual terms before uploading business material.
If the business cannot explain which material the pilot may use, it is not ready to select a tool.
Keep the first boundary small
A first pilot should normally prepare a draft, classification, summary, recommendation, or retrieval result for human review. It should not silently update critical records, contact customers, make commitments, or trigger payments.
Use a separate test environment and a fixed evaluation set. Record prompts, model/version, settings, source material, outputs, corrections, latency, and cost so the result can be reproduced.
Go, pause, or stop
Proceed when the problem, owner, baseline, permitted data, review, and decision thresholds are all clear. Pause when missing data or process inconsistency prevents a fair test. Stop when the expected value is too small, a simpler rule or automation is better, or the pilot requires access and risk disproportionate to the task.
Readiness is the ability to run a controlled experiment and learn from it—not enthusiasm for a particular model.