An AI readiness assessment has one useful output: a decision about what to do next. A score with no evidence is a mood. This checklist is built as a scored worksheet, with a pass criterion for each item, a named owner and a decision band at the end. It is designed for an SME that already has a candidate use case and wants to know whether it is ready to pilot, needs a foundation project or should stop.
How the checklist works
Score each item from zero to three. 0 means absent. 1 means informal or dependent on one person. 2 means documented and assigned. 3 means documented, tested and reviewed within the last twelve months. Write the evidence next to the score: the document, the system, the test result or the person who confirmed it.
Assess one use case at a time. A general assessment of AI maturity produces a comfortable number and no decision. Weight the areas for the decision in front of you. A first pilot should weight use-case definition, data and economics heavily. A scale-up decision should add more weight to governance, security and operating ownership.
1. Sponsorship and decision rights
- An executive sponsor has allocated time, budget and attention, not only a verbal blessing.
- A pilot owner has authority to stop the pilot and change the scope.
- A success metric and a baseline are written down before the pilot starts.
- A stop rule states what result would end the work, and when it will be reviewed.
- The budget covers integration, human review, monitoring and run cost, not only the model or licence.
- An escalation path exists for a harmful, biased or incorrect output.
2. Use case and acceptance criteria
- The user, the process and the moment of use are named.
- Current volume, cycle time and cost are measured from records, not estimated from memory.
- The consequence of an error is understood, with a tolerance for false positives and false negatives.
- The human review point is defined: who checks what, with what authority.
- Inputs and outputs are specified, including format, language and length.
- Acceptance criteria are measurable and agreed with the user, not with the vendor alone.
- A non-AI alternative has been considered and rejected for a stated reason.
3. Data readiness
- The data exists in a system or a controlled repository, not only in email or personal drives.
- Records are complete enough for the task, and duplicates or missing fields are understood.
- Labels, ground truth or a review sample exist for evaluation.
- The lawful basis and consent position are documented for this use.
- The pilot team can access the data without copying it into unmanaged locations.
- Retention and deletion rules are defined, including backups and derived copies.
- Residency and vendor access are acceptable for the data classification.
- A named data owner can answer quality questions during the pilot.
4. Technology and security
- The source systems expose an API, a supported export or a documented integration path.
- Identity is centralised, with single sign-on and role-based access where possible.
- Development, test and production environments are separated.
- Prompts, outputs, tool calls and human decisions are logged for review.
- Secrets and model keys are stored in a managed secret store, not in code or spreadsheets.
- Model access has a fallback or a documented failure path.
- Vendor terms cover data use, training, retention, subprocessors and deletion.
- An evaluation harness and a rollback path exist before live traffic is enabled.
5. People and process
- A named owner will run the workflow after the pilot team moves on.
- Users receive training on the tool, its limits and the review step.
- Exception handling is defined, including what to do when the model is unsure.
- The team has capacity for the change, not only enthusiasm for the demo.
- Staff whose work changes have been consulted and their concerns recorded.
- Performance measures for the people using the tool do not reward unsafe shortcuts.
6. Governance and compliance
- An AI policy covers acceptable use, data handling, disclosure and human oversight.
- A named group or person approves a new model, vendor or material change.
- A risk assessment records the likely harms, affected people and mitigations.
- Quality, bias and harmful-output review is assigned and periodic.
- An incident path covers data leakage, a harmful output and a failed automation.
- Customer disclosure is defined where the use of AI affects a decision or a conversation.
- Records are sufficient for an audit, a customer review or a regulator request.
7. Economics
- Cost per task includes model usage, integration, human review, monitoring and rework.
- The current cost of the process is measured on the same basis.
- Volume and peak assumptions are written down.
- Review labour is costed, not treated as free capacity.
- A run and change budget exists for the first year after launch.
- The cost of exit, including data export and replacement, is understood.
Score sheet
| Area | Weight | Score | Evidence | Owner | Action |
|---|---|---|---|---|---|
| Sponsorship and decisions | 10 | 0-3 | Sponsor note, budget, stop rule | Executive sponsor | Enter |
| Use case and acceptance | 20 | 0-3 | Baseline, criteria, review point | Product owner | Enter |
| Data readiness | 20 | 0-3 | Sample, data map, lawful basis | Data owner | Enter |
| Technology and security | 15 | 0-3 | Integration test, logging, terms | IT lead | Enter |
| People and process | 10 | 0-3 | Owner, training plan, exceptions | Operations lead | Enter |
| Governance and compliance | 15 | 0-3 | Policy, risk assessment, incident path | Compliance lead | Enter |
| Economics | 10 | 0-3 | Unit cost, baseline, run budget | Finance | Enter |
Decision bands
Calculate the weighted score as a percentage of the maximum. These bands are a management convention, not an industry benchmark:
- 0-25 per cent: not ready. Fix sponsorship, use-case definition or data before buying anything.
- 26-50 per cent: foundation first. A data or process project should precede the AI pilot.
- 51-75 per cent: pilot with conditions. Proceed only if the gaps are named, owned and funded.
- 76-100 per cent: ready to pilot or scale, subject to the evidence being real and current.
Do not average away a zero in a critical area. A missing lawful basis or no human review point should block a pilot regardless of the total.
A 30-day evidence plan
- Week 1: confirm the sponsor, the use case, the success metric and the stop rule. Collect the baseline from records.
- Week 2: pull a representative data sample, test its quality and confirm the lawful basis and access route.
- Week 3: test the integration path, review the vendor terms and define the logging and human review steps.
- Week 4: build the unit-cost model, review the governance answers and hold a sign-off meeting with named actions.
Sign-off should record the score, the decision, the conditions, the owner and the next review date. Re-score after the pilot, using the same evidence standard. If the score rises because the evidence improved, that is a useful result. If it rises because the questions were softened, the assessment has failed.
If you need help turning the checklist into a pilot plan, start with Business Automation. The pricing page explains how assessment and pilot work are priced. For the narrative version of the assessment, read the AI readiness self-assessment step-by-step guide.
Sources
- NIST, AI Risk Management Framework, accessed 17 November 2026.
- ISO, ISO/IEC 42001: Artificial intelligence management system, accessed 17 November 2026.
- DataGuidance, Bangladesh: Data Protection Overview, accessed 17 November 2026.
- OECD, AI Principles, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.