AI Readiness Assessment: A 12-Point Enterprise Checklist 

An AI readiness assessment determines if one proposed workload has the business case, data rights, controls, infrastructure, ownership, and adoption conditions that will earn its next investment gate. It should conclude with a decision (proceed, remediate, rescope, or stop) and not a maturity label. Each point is scored 0-2, so you score out of 12. This checklist tells you how to read the score.

A funded pilot may still stall: the model performs well in a test, but no one has access to production data, knows who the live owner is, or can describe what happens when an output is wrong.

Key takeaways

• Readiness is linked to a workload, not the entire organization. When the evidence fits the next gate, when there are no critical blockers and every material gap has an owner, then a use case is ready.
• Four points (data rights, permissions, impact level, and stop rights) can be a bottleneck for a workload, regardless of the overall workload.
• Assessment should include residency and compute capacity from the start for regulated or Canadian data.

Use the checklist below before a pilot is funded, or before a promising pilot receives production investment. It connects to the enterprise AI lifecycle and the enterprise AI roadmap, and it assumes formal governance, procurement, and security review.

AI readiness assessment at a glance

Use the table as the first-pass scorecard. The evidence to collect for each point is set out in the sections below.

How to score the 12 points and read the result

Rate each point 0, 1, or 2. A 0 means no evidence or a known blocker. A 1 means partial evidence, with the gap owned and dated. A 2 means the evidence meets the standard for the next gate. The total runs from 0 to 24.

This is Cylix’s suggested method, not an industry standard, and the bands have not been validated against outcomes. Treat the total as a prompt for discussion. The gates decide.

Result	Suggested decision
20 to 24, with no 0 on points 5, 7, 8, or 9	Proceed, with the 1s tracked as conditions.
13 to 19, or any 0 on points 5, 7, 8, or 9	Remediate before the next gate.
7 to 12	Rescope to a smaller user group, data set, or level of authority.
0 to 6	Stop, unless rescoping changes the picture.

Points 5, 7, 8, and 9 carry the override because each can expose restricted data, remove the ability to intervene, or understate the harm of a failure. A workload that totals 22 and has a 0 on permissions does not proceed.

What an AI readiness assessment should decide

There are two levels of readiness. Organizational readiness includes reusable capabilities such as procurement, identity, data governance, security review, and production support. Use-case readiness asks whether the proposed system can deliver a specified result within its real-world constraints. Don’t roll them into a single comforting maturity score: a mature enterprise can have a poorly framed use case, and a less mature one can succeed with a narrow workload. First score the use case, and see what gaps are in need of broader investment.

The NIST AI Risk Management Framework integrates governance, context, measurement, and risk treatment. NIST published version 1.0 in January 2023 and a Generative AI Profile in July 2024. Map it to the laws, standards, and internal policies that govern the workload. A further reference is ISO/IEC 42001:2023, which specifies requirements for an AI management system. This checklist does not require either.

The outcome of the assessment is one of four decisions. “Proceed” means the evidence supports the next gate, the known conditions, and the owners. Remediate means named blockers must be cleared first. Re-scope reduces the user group, data set, authority, or workflow until the workload is manageable. “Stop” means the expected value does not justify the cost, exposure, or operational burden.

The 12-point AI readiness assessment

1. Is the business problem specific?

Direct answer: A use case is ready for assessment when the team can name the workflow, user, decision or task, current pain, and the consequence of leaving it unchanged.

Start with the operating problem, not the technology. “Improve productivity” is too broad. “Reduce the time claims analysts spend locating approved policy passages during a live case review” is testable because the user, task, and expected change are visible.

If nobody can point to where the AI output enters the current workflow, the proposal is still an idea rather than a fundable business case.

2. Is there a measurable baseline?

Direct answer: The team needs a credible starting measure before it can claim that AI changed the outcome.

Record the current cycle time, error rate, volume, cost, rework, service level, or another measure tied to the problem. Document how it was calculated and the period it covers. An imperfect baseline is acceptable if it is robust enough to compare against pilot results.

3. Is there an accountable business owner?

Direct answer: Sponsorship is not ownership. The workload needs a business owner with authority over the process and a technical owner accountable for the system.

An executive may fund the initiative, but a process owner decides how the workflow changes, which outcome counts as success and which residual risk is acceptable. The technical owner stays responsible for architecture, release evidence, and post-launch changes.

4. Are the users and decision points known?

Direct answer: Readiness depends on knowing who interacts with the system, who relies on its output, who may be affected, and what authority the AI has at the decision point.

A system that drafts an answer for review needs different controls from an agent that sends it or changes a customer record. State whether the system drafts, recommends, ranks, approves, or acts; that choice drives human review, logging, and release of evidence.

Data access, quality, and rights

5. Can the team lawfully and practically access the required data?

Direct answer: Data is ready only when the team has a lawful permitted use and a production access path that works at the required speed and scale.

List each source system, owner, access method, jurisdiction, sensitivity class, and permitted use. Include prompts, conversation logs, embeddings, evaluation sets, and user feedback, which can contain sensitive or derived information. A repository that needs manual exports or months of integration is not “available” in any sense that matters to delivery; treat that work as a dependency with an owner and a date.

The enterprise question is where the data will be processed and stored. The Privacy Commissioner of Canada’s guidance on cross-border processing says the organization remains accountable for personal information a processor handles abroad and that it should assess the risk of foreign access, since a contract cannot override another country’s laws. In Québec, a privacy impact assessment is required before personal information is disclosed outside the province, and the disclosure may proceed only if the information would receive adequate protection, per the Commission d’accès à l’information (in French). Score point 5 at 2 only when residency is decided and documented. Cylix describes its AI Fabric as keeping data in its home jurisdiction, with sovereign, hybrid, and cloud-resident deployment models.

6. Is the data fit for the task?

Direct answer: Data quality should be tested against the workload rather than against a generic quality score.

Review completeness, timeliness, consistency, representation, and known error patterns. For structured data, inspect schemas, missing fields, and identifiers. For documents, check sample formats, duplicates, version conflicts, and scan quality. For event data, verify timestamps, ordering, and loss.

7. Can data lineage and permissions survive production?

Direct answer: A production AI system must preserve source permissions, traceability, and revocation after data enters retrieval, indexing, caching, or other derived stores.

Retrieval-augmented generation needs source-level access controls, current indexes, and traceable citations. A user should not receive a passage simply because the model can retrieve it. Test what happens when a document is revoked, corrected, or deleted, and how long the change takes to reach the index.

Risk, governance, and human accountability

8. Is the impact level understood?

Direct answer: Control depth should increase with the severity, scale, autonomy, and irreversibility of a credible failure.

Record the worst credible outcomes for customers, employees, operations, finances, safety, rights and reputation, and whether an affected person can detect and challenge an error. Hard-to-reverse outcomes and higher autonomy call for stronger evaluation and approval.

Canadian federal departments can use the Algorithmic Impact Assessment, a questionnaire that places a system at one of four impact levels, for systems that fall within the scope of the Directive on Automated Decision-Making. Other organizations should map the requirements that apply to their sector and jurisdictions.

9. Are approval, escalation, and stop rights assigned?

Direct answer: A control is weak if nobody has both the authority and the evidence required to use it.

Name the person who can approve the next stage, impose conditions, pause the system, and authorize a restart. Define where human review is mandatory and what the reviewer sees, and test the intervention itself. These responsibilities belong in an AI governance framework that defines decision rights, control requirements, and accountability across models, agents, data, and vendors.

Integration, infrastructure, and operating ownership

10. Can the system integrate and perform under real conditions?

Direct answer: A workload is technically ready when the production design can connect to required systems, preserve identity and permissions, meet service needs, and fail safely.

Identify interfaces, identity flows, network boundaries, latency targets, availability needs, and transaction volumes. Verify which actions are read-only and which can change records or call tools.

Estimate compute and storage from a stated workload, then test the estimate against peak demand, concurrency, context length, retrieval calls, and fallback behaviour. Infrastructure should follow the workload requirement, not define it.

For a workload on private or sovereign infrastructure, capacity is a readiness question in its own right. Ask who owns the GPUs, how jobs are scheduled between teams, and what happens when demand outgrows the cluster. Cylix’s managed HPC service lists Slurm scheduling, capacity forecasting, and power, cooling, and rack-density planning, and its Orbit layer tracks accelerator saturation and queue wait. Score 10 points at 2 only when the capacity plan covers peak load.

11. Is there an owner for production operations?

Direct answer: Every production layer needs a named operator, a support expectation, and an escalation path before release.

Assign ownership for prompts, models, retrieval indexes, data pipelines, integrations, access policies, incidents, and vendor changes. If internal teams cannot provide the required coverage, compare staffing, platform, and managed AI services before launch. Delegating technical operations does not transfer business accountability or risk acceptance.

Procurement belongs here too. Before signing, ask each vendor where inference runs, who can access prompts and logs, how model changes are announced, what exit looks like, and which obligations pass to subcontractors. A 2 means those answers are in the contract, not a sales deck.

12. Will the workflow, people, and measures change together?

Direct answer: Adoption is ready when the new workflow, training, management reinforcement, feedback path, and outcome measures have been designed together.

Confirm how users will be trained and how feedback reaches the product team. Identify incentives that could cause overuse, avoidance, or superficial compliance. Set acceptance thresholds before the pilot.

Worked example: scoring a claims assistant

Take an internal claims assistant that retrieves approved policy passages for human analysts. It answers well in a sandbox, but production permissions and index revocation are untested.

Points 1 to 5 score 2 each: workflow, baseline, owners, users, and approved data sources are documented. Point 6 scores 1 because sampling is partial. Point 7 scores 0, since document-level authorization and revocation are untested. Points 8 to 11 score 1 each: impact assessed, stop rights named but not exercised, load testing planned, operations owner named but the runbook incomplete. Point 12 scores 2.

The total is 17 out of 24, in the remediate band, and the 0 on point 7 would force that outcome at any total. The decision is “remediate,” not “17 out of 24, ready.” The permission gap blocks progression because it can expose restricted information. The operating gap can move forward as a named condition if the next gate does not depend on production support.

Turn the result into a sequenced roadmap.

Convert the findings into three work queues, not a coloured scorecard:

• Blockers: missing rights, critical controls, undefined ownership, or anything else that prevents the next gate.
• Near-term remediation: work inside the current initiative, such as a data sample, permission test, or integration proof.
• Scale foundations: reusable capabilities such as identity patterns, observability, model inventory, and vendor review.

Each action needs an owner, an evidence requirement, and the next decision gate. These outputs feed the enterprise AI roadmap. Readiness is an input to planning. It does not replace the pre-production AI risk assessment, which tests the deployed control evidence before release.

Cylix in practice

Cylix places assessment at the start of its enterprise AI lifecycle, the first of five stages. In its Managed AI service, the assessment step covers readiness, use-case prioritization, architecture review, and a review of data governance, security, and sovereignty. Its data engineering approach begins with business and AI readiness, so data investment is tied to business outcomes rather than treated as an isolated technical project.

Frequently asked questions

What are the stages of AI readiness?

There is no single standard five-stage model; vendors publish their own maturity scales, and they differ. This checklist avoids stages. It scores a workload against 12 points and ends in one of four decisions: proceed, remediate, rescope, or stop.

Who should conduct an assessment of AI readiness for the enterprise?

The business owner and technical owner should lead it, with input from security, privacy, legal, data owners, and procurement. An outside party can facilitate, but the business owner still accepts the residual risk.

How is an enterprise assessment different from a small-team checklist?

Enterprise assessments add formal decision gates, residency and cross-border review, compute capacity planning, vendor and procurement terms, and named stop rights. A smaller team can cover the same ground with lighter evidence and should not copy governance it cannot staff.

Does data residency change an AI readiness assessment?

Yes, this is the case whenever the workload uses personal or regulated information. In Canada, an organization remains accountable for data a processor handles abroad, and Québec requires a privacy impact assessment before personal information leaves the province. Decide where prompts, logs, and embeddings are processed before choosing architecture.

What is a good AI readiness score?

No standard score exists. With the 0 to 2 method in this article, 20 or more out of 24, with no 0 on points 5, 7, 8, or 9, supports proceeding. Lower totals, or a 0 on those points, mean remediation or re-scoping first.

Do we need NIST AI RMF or ISO/IEC 42001 to run the assessment?

No. The NIST AI Risk Management Framework is a voluntary reference for managing AI risk, and ISO/IEC 42001 specifies requirements for an AI management system. Either framework can help you structure your controls, but you can still score the 12 points without them.

Fund evidence, not enthusiasm

An AI readiness assessment should show what outcome matters, what evidence exists, where the blockers sit, and who owns the work. Low scores do not automatically end an initiative. They change its sequence, scope, or funding conditions.

Apply the 12 points to one real workload before approving a broad program.

Next step: Request a consultation to assess one workload against its current business case, data, controls, architecture, and operating constraints.