AI Readiness Checklist

The starting point. If there's no clear problem tied to business value, and no one senior enough to own it, everything else is irrelevant. GenAI enthusiasm is not a use case.

• Is there a named, specific use case — not just 'we want AI'?

  The use case should reference a concrete problem, a measurable cost, and an owner.

  Critical

• Does the use case have a senior sponsor with budget and authority?

  Projects without exec sponsorship rarely survive prioritisation battles.

  Critical

• Is there a clear definition of success — not just 'better'?

  Think: time saved, error rate reduced, revenue influenced. Something you can measure in 3 months.

  Important

• Is AI genuinely the right solution — or would a simpler tool do the job?

  Challenge the assumption early. LLMs are expensive, complex, and often overkill.

  Important

• Are the key business stakeholders aligned on scope and expectations?

  Misaligned expectations are the most common cause of 'failure' that isn't actually a technical failure.

  Watch

Probably the dimension most underestimated by clients. AI doesn't improve data quality — it amplifies whatever quality exists. Inaccessible or untrustworthy data kills more AI projects than bad models do.

• Is the relevant data actually accessible — not trapped in legacy systems or behind red tape?

  Access issues often surface only mid-project. Map this before signing off on any timeline.

  Critical

• Is data quality understood and documented?

  Completeness, accuracy, freshness, consistency. If no one knows, assume the worst.

  Critical

• Is there a clear owner for the data that will feed the AI system?

  Ownerless data drifts. You need someone responsible for its ongoing quality.

  Important

• Are there PII, sensitivity, or compliance constraints on the data?

  GDPR, sector-specific regulation, internal classification policies. Surface these early.

  Important

• Is there enough labelled or historical data to train, fine-tune, or evaluate the model?

  For GenAI use cases this bar is lower — but not zero. Know what you have.

  Watch

• Can the data be refreshed reliably once the system is live?

  A model trained on stale data will degrade. Who refreshes it, how often, and at what cost?

  Watch

A demo that works on a laptop is not a production system. The question is whether the underlying platform can host, integrate, and scale a solution that real users will depend on.

• Is there a cloud or on-prem environment capable of hosting the AI workload at scale?

  Evaluate compute, storage, and latency requirements against what's available.

  Critical

• Can the target system integrate with existing data sources and downstream tools via APIs?

  Integration debt is where projects stall after the PoC. Check this before the PoC.

  Important

• Is the current tech stack stable enough to build on — or is there significant underlying technical debt?

  Building AI on a fragile foundation just shifts the problem downstream.

  Important

• Are there existing ML/AI platform components that can be reused (vector stores, embedding pipelines, orchestration)?

  Reuse reduces time-to-value and lowers maintenance burden.

  Watch

• Is the architecture observable — can you monitor model behaviour and system health in production?

  You can't manage what you can't see. Logging, tracing, and alerting must be designed in.

  Watch

Governance isn't compliance theatre. It's the operational structure that lets you deploy AI without it becoming a liability. Most organisations skip this until something goes wrong.

• Is there a defined owner responsible for AI outputs in production — not just IT?

  Business ownership of AI outputs is non-negotiable. IT can run the model; they can't own the decisions it informs.

  Critical

• Is there a process for reviewing AI outputs before they affect customers, staff, or decisions?

  Human-in-the-loop is not optional for high-stakes use cases. Define the review triggers now.

  Critical

• Are the risks of this specific use case understood and accepted by the right people?

  Hallucination risk, bias risk, regulatory risk. Has the relevant risk function signed off?

  Important

• Does the organisation have an AI or data ethics policy — or at minimum, working principles?

  If not, you'll need to establish guardrails for this project specifically.

  Important

• Is there a process for handling failures or unexpected model behaviour once live?

  What happens when it gets something wrong? Who decides, who escalates, who fixes it?

  Watch

• Are there audit or explainability requirements for this use case?

  Regulated industries and internal audit functions often require traceable decisions. Know this upfront.

  Watch

Can the organisation build it, and — critically — sustain it after go-live? Many AI projects get delivered and then orphaned. Delivery capability includes the people and change management, not just the technical skills.

• Does the team have the technical skills to build and operate the solution — or is there a realistic plan to acquire them?

  ML engineering, prompt engineering, platform ops. Know the gap before you scope.

  Critical

• Is there a realistic plan for who owns and maintains the solution after go-live?

  The most common failure mode: project team disbands, no one retrains the model, quality degrades silently.

  Critical

• Has the end-user change management been planned — not just the technical rollout?

  Adoption fails when users aren't brought along. Training, communication, and feedback loops must be designed.

  Important

• Is the project scope realistic given available time, budget, and team capacity?

  AI projects routinely underestimate data prep and integration work. Build in contingency.

  Important

• Is there a plan for retraining or updating the model as data and requirements evolve?

  Models degrade as the world changes. Drift detection and retraining cadence should be agreed upfront.

  Watch

• Are there dependencies on third parties (model providers, data vendors) that introduce delivery or continuity risk?

  API pricing changes, vendor lock-in, terms of service restrictions. Surface early.

  Watch