AI Capability Maturity Assessment · Part I. Understanding the model
The five levels
In this chapter
The five levels, Initial, Repeatable, Defined, Managed and Optimising: what each looks like, the evidence that shows it, what moves you up, and how to choose a realistic target.
Every criterion is rated on the same five levels. A level describes how the organisation works, not what it intends: each is shown by what can be produced and observed, and the pass tests on each criterion decide whether a level is met.
- 5OptimisingContinuously improving; governance feeds strategy.
- 4ManagedMeasured and enforced; evidence generated, not compiled.
- 3DefinedEnterprise-wide standards defined, communicated and followed.
- 2RepeatablePockets of good practice; documentation partial; governance reactive.
- 1InitialAd hoc, individual-led use; no accountable owner.
1 Initial
Ad hoc, individual-led use; no accountable owner.
AI is used by individuals and teams on their own initiative. Nobody can list the AI in use or say who is accountable for it. Outcomes depend on who happens to be involved, and obligations are met, if at all, by chance.
Typical evidence
- Little or no documentation; interview answers differ from person to person
- Staff using personal or public AI accounts for work
- No register, approved tool list or acceptable-use rules
- Incidents and near misses handled as general IT tickets, if at all
What moves you to level 2
- Name an accountable executive and an owner for each known use
- Open a register and run a first discovery sweep
- Publish an approved tool list and short acceptable-use rules
- Write down the first boundaries: what AI will and will not be used for
2 Repeatable
Pockets of good practice; documentation partial; governance reactive.
Some teams do things well and could do them again, but practice depends on people rather than standards. Documents exist in draft or for the largest projects only. Governance reacts to problems and deadlines, and obligations are met where someone remembers them.
Typical evidence
- Draft strategy or policy documents that have not been endorsed
- A working group without decision rights; a partial list of AI projects
- Templates that vary between teams; assessments completed after go-live
- Training available but optional, with low or untracked completion
What moves you to level 3
- Endorse the standards: strategy, risk appetite, lifecycle and acceptable use
- Make the processes mandatory rather than available
- Close coverage gaps: embedded vendor AI and staff use of general-purpose tools
- Give every in-scope use case a single accountable owner
3 Defined
Enterprise-wide standards defined, communicated and followed.
There is one way of doing things across the organisation: an endorsed strategy and risk appetite, a committee with decision rights, a lifecycle with gates, a complete register and published rules. Records show the practice is followed, not just written down. Obligations are discharged by design.
Typical evidence
- Endorsed documents plus records of them operating: minutes, decision logs, gate records, completed assessments
- A register covering deliberate, embedded and staff-used AI, refreshed by discovery within 12 months
- Published approved tool list with sensitivity ceilings; acknowledged acceptable-use policy
- Training completion of at least 80 percent, tracked by business area
What moves you to level 4
- Measure: dashboards generated from systems of record, not compiled by hand
- Enforce controls in platforms rather than by convention
- Test that controls operate (rehearsals, sampling, internal audit)
- Generate evidence as a by-product of operating
4 Managed
Measured and enforced; evidence generated, not compiled.
Practice is measured and controls are enforced by systems, not memory. Evidence is produced as a by-product of operating: approvals are bound to what was reviewed, decisions carry records, monitoring raises alerts and the executive governs from a dashboard. Supervised autonomy is safe because the organisation can see, stop and reverse what its AI does.
Typical evidence
- System configuration showing enforcement (deployment gates, approval settings, data loss prevention)
- Metrics with trends over at least two quarters, and the decisions they triggered
- Rehearsal records with measured times (halt, rollback, incident response)
- Internal audit reports on AI with findings tracked to closure
What moves you to level 5
- Close the loop: use operating evidence to change strategy, appetite and controls
- Automate assurance where evidence is already generated
- Grant and withdraw autonomy on measured performance
- Share practice beyond the organisation and invite independent review
5 Optimising
Continuously improving; governance feeds strategy.
Evidence from operation continuously improves strategy, controls and design. Autonomy is granted and withdrawn on measured performance within limits the systems cannot exceed. Assurance is continuous, and the organisation shares its practice with its sector or with government.
Typical evidence
- Trend data across at least two review cycles showing improvement
- Documented changes to strategy, appetite or controls made because of evidence
- Automated containment and continuous assurance in operation
- External contributions: shared patterns, published lessons, participation in standards or sector forums
Staying here
- Level 5 is sustained rather than reached: protect it with re-assessment, independent review and early attention to new classes of AI capability
Choosing a target level
- Level 3 is the realistic 12 to 24 month target for most organisations and is the level at which most current obligations are met by design.
- Level 4 is warranted where AI acts with supervised autonomy, makes or shapes decisions about people at scale, or handles sensitive information. The step from 3 to 4 is where assurance moves from reviewing outputs to governing systems at runtime.
- Level 5 is rarely the right target for every domain. Targets should follow the risk and ambition of the AI the organisation actually runs.
Scores between levels are reported honestly: a domain at 2.8 is Repeatable, approaching Defined, never rounded up. See bands in the scoring method.
Each level, domain by domain
What each level looks like in each of the eight domains. Open a level to compare the domains side by side.
1Initial across the eight domains
| Domain | Level 1: Initial |
|---|---|
| 1. Strategy and leadership | No AI strategy or risk appetite. AI use emerges from individual teams and vendors. No executive owns AI outcomes, and leaders learn about AI use when something goes wrong. |
| 2. Governance and operating model | No accountable owners, register or governance forum. Nobody can list the AI in use or say who approved it. Decisions are made in projects or by email and cannot be traced. |
| 3. Risk, assurance and transparency | No AI impact assessments, tiers or AI-specific gates; the risks are unknown. People affected are not told AI is involved and have no specific path to challenge a decision. AI incidents are not recognised as such. |
| 4. Data, knowledge and semantics | The quality of data feeding AI is unknown. There are no shared definitions, no rules on what information may enter AI tools, no record of what AI did, and AI content is not treated as a record. |
| 5. Platform, architecture and security | AI runs wherever teams set it up, including personal accounts and unapproved services. There are no AI-specific security controls. Vendor AI features are enabled without review. Nobody monitors AI behaviour, and halting a system means turning off the whole application. |
| 6. Design and engineering practice | AI is built or configured from demos and prompts with no specification. Autonomy is whatever the tool does by default. Testing stops at a demonstration. AI-generated code and content get no extra review, and safety is considered after problems appear. |
| 7. Use-case delivery and value | Use cases start wherever someone has budget or enthusiasm. There are no stages, no graduation test and no benefit measures. Each team builds from scratch. |
| 8. People, capability and culture | No AI training or acceptable-use rules. AI capability depends on a few enthusiasts or a vendor. Staff either distrust AI or rely on it without checking, and problems go unreported. |
2Repeatable across the eight domains
| Domain | Level 2: Repeatable |
|---|---|
| 1. Strategy and leadership | An ambition has been stated in a draft or a speech but is not endorsed or tied to priorities, boundaries or risk appetite. An AI lead is named but holds no delegated authority or budget. AI reaches executive agendas occasionally, as project updates. |
| 2. Governance and operating model | Some owners are named and an informal list of AI projects exists. A working group meets but discusses rather than decides. Embedded vendor AI and staff use of general-purpose tools are missing from the list. Assurance happens when someone asks for it. |
| 3. Risk, assurance and transparency | Some prominent use cases have been assessed, using varying templates and often after deployment. Tiering is informal and inconsistent. Disclosure is partial. AI incidents go through general IT or privacy processes, and near misses are not recorded. |
| 4. Data, knowledge and semantics | Quality and definitions are managed in some domains but not linked to AI use. Policy forbids some information in AI tools but nothing enforces it. Logs exist but are technical and short-lived. Retention depends on each tool’s defaults. |
| 5. Platform, architecture and security | An enterprise assistant is approved, but other AI is built ad hoc. Vendor controls are relied on without mapping AI threats. The approved tool list is not published. Uptime is monitored but behaviour is not, and halt procedures are undocumented. |
| 6. Design and engineering practice | Some requirements exist but describe features, not behaviour. A person is expected to check outputs, without criteria or records. Builders write a few tests and run them manually. Developers are told to review AI output, but nothing enforces it. |
| 7. Use-case delivery and value | An ideas list exists but is not the required route. Pilots report results without criteria set in advance and rarely scale. Benefits are estimated in business cases but not measured. Sharing is informal. |
| 8. People, capability and culture | Training is available but optional and completion is low. Rules sit inside a long ICT policy. Champions help informally. Adoption is measured by licence counts. Some staff report problems, but nothing visible happens. |
3Defined across the eight domains
| Domain | Level 3: Defined |
|---|---|
| 1. Strategy and leadership | An endorsed strategy states purpose, priorities and boundaries, supported by an endorsed risk appetite with defined tiers. A Chief AI Officer or executive sponsor holds written authority, including to pause AI use. Investment is managed as a portfolio with gates, and the executive reviews a defined AI dashboard at least quarterly. |
| 2. Governance and operating model | Every in-scope use case has a named accountable owner. A committee with terms of reference and decision rights meets at least monthly and keeps a decision log. The register is complete after a discovery sweep, including embedded AI. Three lines of assurance are assigned, and decision rights and escalation are written down. |
| 3. Risk, assurance and transparency | A standard impact assessment is completed before deployment for every in-scope use case. Tiers are applied by written criteria. A lifecycle with go or no-go gates scaled by tier is followed and recorded. A transparency statement is current, decisions about people are inventoried with a human review path, and an AI incident runbook with halt authority has been rehearsed. |
| 4. Data, knowledge and semantics | Priority data domains feeding AI have owners, quality thresholds and a measured baseline. Knowledge sources are curated with owners. Each AI tool has an approved sensitivity ceiling. A decision-record standard is applied to in-scope uses, and AI content is mapped to retention rules. |
| 5. Platform, architecture and security | In-scope AI runs on an assessed reference architecture with documented residency. AI threats are mapped to implemented controls, and AI applications are security tested before release. A published approved tool and model list states uses and ceilings. Vendor AI is assessed and contracted with AI clauses. Each system is centrally logged, monitored for behaviour and has a documented halt and rollback with named authority. |
| 6. Design and engineering practice | Every in-scope component has a signed-off specification including constraints, failure modes and fallback. Autonomy and approval rules are documented per system and approvals are recorded. Evaluation suites with agreed pass criteria cover typical, edge and adversarial cases, with shadow runs for higher tiers. Merge gates and scanning apply to AI-generated code, and secure-and-safe design reviews are signed off before release. |
| 7. Use-case delivery and value | All use cases enter through one intake and a prioritised portfolio with recorded refusals. Every use case shows its lifecycle stage with evidence. Pilots have success criteria and decision dates, benefits are baselined and reported, and a maintained pattern catalogue is checked before new builds. |
| 8. People, capability and culture | Foundational training is mandatory and at least 80 percent complete. A short acceptable-use policy is published and known by most staff. Capability needs are mapped with a sourcing plan and role-based learning. A resourced champion network covers every major area, and a visible channel exists for raising AI concerns, with every report answered. |
4Managed across the eight domains
| Domain | Level 4: Managed |
|---|---|
| 1. Strategy and leadership | Strategy drives a measured portfolio: every funded initiative maps to a priority, funding is released at gates on evidence, and pilots are stopped when the evidence says so. The executive governs from a dashboard generated from systems of record, and minutes show decisions taken because of it. |
| 2. Governance and operating model | The committee decides at gates on evidence, including refusals and conditions tracked to closure. The register is fed by intake and procurement, reconciled regularly and audited. Internal audit has completed an AI audit. Owners present their own submissions and attest annually. |
| 3. Risk, assurance and transparency | Gates are enforced so nothing reaches production without a recorded decision. Assessments are refreshed on change and tiering consistency is tested. Review requests and incidents are tracked and analysed, with corrective actions closed. Audit evidence can be produced on demand. |
| 4. Data, knowledge and semantics | Quality is monitored for production AI with defined responses to breaches. Ceilings are enforced technically. Decision records are tamper-evident, complete and used to answer reviews and audits. Legal holds and disposal apply to AI content and have been tested. |
| 5. Platform, architecture and security | Guardrails are enforced as code. AI controls are monitored and adversarially tested for higher-tier systems. Unapproved AI use is detected and followed up. Vendors notify model changes under contract. Halt and rollback are rehearsed with time to halt measured, and agents have their own identities with least privilege. |
| 6. Design and engineering practice | Specifications are versioned and traced to tests. The platform enforces autonomy settings and binds approvals to what was reviewed, and approval and override rates are monitored. Evaluations run automatically on every change and block failing releases. Provenance of AI-assisted work is recorded and its quality measured. |
| 7. Use-case delivery and value | Intake and graduation decisions are made within target times on measured results. Benefits are measured from operational data and reported quarterly, shortfalls included. Scaled services have operating owners before go-live. Reuse is measured, and patterns carry assurance evidence that shortens gates. |
| 8. People, capability and culture | Training completion is above 95 percent and its effectiveness is measured. Acceptable use is monitored and awareness is above 85 percent. Capability is measured against the map and approvers are trained before they act. Adoption is measured by meaningful use. Near-miss reporting is healthy and reports lead to visible changes. |
5Optimising across the eight domains
| Domain | Level 5: Optimising |
|---|---|
| 1. Strategy and leadership | AI is part of corporate strategy rather than a separate plan. Evidence from operation (benefits, incidents, overrides, re-assessment) drives annual refreshes of strategy and appetite. Leaders can answer any external question about the organisation’s AI from evidence already held. |
| 2. Governance and operating model | Governance is built into how work runs: decision rights are encoded in tooling, the register is a live system of record, standard decisions are delegated by rule, and the committee spends its time on exceptions and trends. Governance evidence feeds strategy and appetite. |
| 3. Risk, assurance and transparency | Assurance is continuous and draws on generated evidence such as test, shadow and monitoring results. Lessons from assessments and incidents improve controls across the portfolio. Transparency measures are reported and improving, and affected people are consulted on high-risk designs. |
| 4. Data, knowledge and semantics | A shared semantic layer is used across systems and agents. Data issues found by AI monitoring are fixed at source. Decision records drive explanations at the point of decision and are shared across the supply chain where needed. |
| 5. Platform, architecture and security | The platform is a governed shared service with self-service onboarding for approved patterns. Routing selects approved models by task and sensitivity. Containment is partly automated, resilience and provider exit are tested, and AI security findings are shared with partners. |
| 6. Design and engineering practice | Autonomy is earned and withdrawn on measured performance within limits the system cannot exceed. Specifications are machine-readable and enforced at runtime. Evaluation suites grow from production incidents, and safe defaults are inherited from shared patterns. |
| 7. Use-case delivery and value | The portfolio is balanced across value, risk and capability building, and realised benefits are reinvested deliberately. Proven patterns scale to new areas in weeks as shared services, and lifecycle data is used to remove friction each year. |
| 8. People, capability and culture | Learning and speaking up about AI are part of the culture. Rules are embedded in tools at the point of use. Staff help redesign work around AI and take part in evaluation. The organisation develops and shares AI capability beyond its own walls. |
This assessment is not legal advice and is not a certification. Obligations and dates were checked against official sources on 15 October 2026; confirm them against the source instruments, which can change.