The AI Readiness Checklist

Strategy and mandate

• There is a written statement of what AI is for in this organisation, and what it is not for.
• A named executive owns AI outcomes, not just AI tooling.
• AI objectives are expressed as business outcomes (cost, cycle time, service quality), not adoption counts.
• There is a decision on which workflows will not be automated.

Workflow and process

• The target workflow is documented as it actually runs, including exceptions and workarounds.
• We know the current baseline: volume, cycle time, error rate, cost per case.
• Handoffs between people and systems are written down, with the owner of each.
• There is an agreed definition of "better" for this workflow before any tool changes.

Knowledge and data

• The documents an AI would need are findable, current, and version-controlled.
• We know which sources are authoritative and which are superseded.
• Data quality problems (duplicates, gaps, inconsistent formats) are known and bounded.
• Someone is accountable for knowledge staying correct after the pilot ends.

Technology and access

• We know where the data for this workflow actually lives — including spreadsheets and inboxes.
• There is a route to integrate AI with existing systems that does not require a rebuild.
• We have a sandbox or test environment separate from production.
• Someone can support the workflow operationally after handover.

Governance and decision rights

• For each AI-assisted decision, we can name who approves, who can override, and who is accountable.
• Prompts, outputs, overrides, and approvals are logged and reviewable.
• There is a written fallback to a manual process when the system is wrong or unavailable.
• Procurement rules cover AI vendors, data handling, and model changes.

People and capability

• The people who do the work were involved in designing the change.
• Staff can explain what the AI does and where it is unreliable.
• Training covers exceptions and escalation, not only the happy path.
• There is a plan for the roles this changes — including roles that shrink.

Risk and trust

• Sensitive personal or commercial data is identified and minimised before it reaches a model.
• We have tested accuracy on our own hard cases, not just vendor demos.
• Bias and disparate impact have been examined where decisions affect people.
• We know what happens when the model is confidently wrong — and who catches it.

Measurement and learning

• Success criteria and a failure threshold were agreed before the pilot started.
• We will measure the workflow after rollout against the pre-pilot baseline.
• There is a review date at which the pilot is continued, changed, or stopped.
• What we learn is captured somewhere other people can find it.