An AI maturity model is a staged scale that places an organization’s use of AI somewhere between scattered experiments and AI built into everyday operations, so its leaders can see which capabilities hold back the next stage. The published models compared below score strategy, people and technology, and none gives data more than one pillar. In our view that pillar deserves more weight: when Google researchers interviewed 53 AI practitioners working on high-stakes AI for a 2021 study, 92% of them reported at least one “data cascade”, a run of problems further down the line that started with the data.
Key takeaways
- MITRE’s AI Maturity Model (guide dated November 2023) scores six pillars and 20 dimensions on five levels. Its model chapter names them Initial, Engaged, Defined, Managed and Optimized; its executive summary and website call level 2 “Adopted”.
- Gartner’s AI Maturity Assessment page (Internet Archive copy of October 9, 2026) groups organizations into five stages, from Foundational to Transformational, and assesses seven pillars; the in-depth report is for Gartner clients.
- Microsoft’s agentic AI adoption maturity model (Microsoft Learn, last updated May 20, 2026) rates five capability pillars from Level 100 to Level 500 and tells users not to collapse the result into a single overall score.
- MITRE’s guide says that when several business units are involved in a dimension, the assessment uses the lowest level among them, unless one specialized unit does that work for the whole organization, in which case that unit’s level counts; it puts the interval between assessments at six to 12 months, which it calls likely sufficient in most cases.
- In the CHI 2021 paper “Everyone wants to do the model work, not the data work”, 92% of the 53 AI practitioners interviewed, all working on high-stakes AI such as health and conservation, reported one or more data cascades, and 45.3% reported two or more in a project.
What is an AI maturity model?
It has two parts: a short ladder of levels, and the areas you score on that ladder. Gartner’s own FAQ describes the job as finding current capabilities and gaps, then feeding the roadmap. MITRE adds a point worth keeping in front of the sponsor: the right target depends on the organization’s mission, and not every organization needs level 5 in every pillar.
An AI maturity model describes an organization that already uses AI, across all of its projects. That is the difference from a readiness check on a single project, covered further down. It also sits on top of a data maturity model: the AI stages describe what the organization does with AI, and the data levels describe whether the data under it will hold.
Four published models, side by side
We read each model on its owner’s page in November 2026. Gartner’s is summarized from its public page only; its in-depth report is for clients.
| Model | Who publishes it | Levels | What it scores | Where data sits | Access |
|---|---|---|---|---|---|
| MITRE AI Maturity Model | MITRE; guide dated November 2023 | Five, from Initial to Optimized (level 2 is “Engaged” in the model chapter and “Adopted” elsewhere) | Six pillars, 20 dimensions, one multiple-choice question per dimension | A Data pillar with four dimensions: architecture, security and privacy, AI data governance, accessibility | Model free to download; the assessment tool on request from MITRE |
| Gartner AI Maturity Assessment | Gartner | Five stages, from Foundational to Transformational | Seven pillars; client reports show a current and a target level for each capability | Data is one of the areas it scores | A demo after a contact form; the in-depth report is for clients |
| Microsoft agentic AI adoption maturity model | Microsoft Learn; updated May 20, 2026 | Level 100 Initial to Level 500 Efficient | Five capability pillars for AI agents | Shared with technology in the “Technology and data” pillar | Free to read |
| Our five-level model | Level Up Your Data, 2026 | Initial, Developing, Defined, Managed, Optimized | Eight areas, from AI strategy and value to skills, culture and adoption | Its own area, Data Readiness for AI, which includes the documents generative AI answers from | Free Quick and Detailed assessments |
One caution on the Gartner row. Vendor summaries that rank for “gartner ai maturity model”, BMC’s among them, still list a different set of stage names that starts with Awareness, so check the date of whatever you are reading, and go back to Gartner’s own page before you put its stage names in a board paper.
The five stages in plain words
In our reading, the ladders climb in a similar order under their different names. At the bottom, individuals try AI tools on their own and results depend on who did the work. Next come pilots in a few teams, run their own way, and few of them reach daily use. The middle stage is the first one where something is written down and shared: an AI strategy, roles, approved platforms and governance, applied across most of the organization, with solutions that reach production.
Above that, AI is run as a portfolio, and value, risk, quality and adoption are measured against targets. At the top, AI is part of how the business competes, and the organization can show a regulator or a customer that its AI is trustworthy. Those are the descriptions behind our own five levels. MITRE’s site adds a useful test for any step up: moving to the next level depends on having met the benchmarks of the level below.
The data maturity each stage needs
This is the part the published models leave thin, so here is our mapping. Each row names the data management practice an AI stage can’t do without, and the DAMA-DMBOK knowledge areas to score for it.
| AI stage | What AI work looks like | The data maturity it depends on | Areas to score |
|---|---|---|---|
| 1 Initial | Staff use public AI tools on their own | A classification that tells people which data may not go into an outside tool, and an acceptable-use rule they have been shown | Data Security |
| 2 Developing | Pilots in a few teams, each built on a one-off extract | The fields each pilot uses are profiled and have an owner who can say what they mean | Data Quality; Metadata Management |
| 3 Defined | Solutions reach production under a documented process | Pipelines that rerun, lineage from source to training set, owners who approve each new use of their data, and versioned documents behind any generative AI assistant | Data Integration and Interoperability; Metadata Management; Data Governance; Document and Content Management |
| 4 Managed | AI is run as a portfolio with targets for value and risk | Quality thresholds on the data each live model depends on, checked in production, with an owner who acts when a threshold is missed | Data Quality; Data Governance |
| 5 Optimized | AI is part of how the business operates, and its trustworthiness can be shown | An evidence trail for every model: which data trained it, the right to use that data, and the quality history, ready for an auditor | Metadata Management; Data Governance |
Read it as a dependency, not a second scale. For illustration: a retailer scores its AI at stage 3 because two forecasting models run in production, but its last data maturity assessment put Metadata Management at level 1. Nobody can trace a forecast back to the sales feed it was trained on, so the stage 3 row isn’t met in full, and the organization is closer to stage 2 than its project count suggests. MITRE makes the same point from its own side: at level 3, its Data pillar asks for an approved enterprise-wide data architecture and for AI governance with accountable executives and audits in place.
If you work with DCAM rather than the DMBOK, its analytics management component is the closest match for the governance of models; our post on AI data governance walks through it.
How to run an AI maturity assessment
- Set the scope. Name the unit you are assessing, the whole organization or one business unit, and record the date and who took part, as MITRE’s guide asks.
- Pick people who see different parts. Business, technology, data, risk and legal each know something the others don’t. MITRE recommends a cross-functional team, though one knowledgeable person can assess a small entity such as a single project.
- Answer separately, with evidence. Each person picks the statement that matches what they can show, not what is planned. Allow a “Don’t know” answer, and exclude it from the score instead of guessing a level.
- Read the area scores before the overall one. Microsoft warns that a single score can mask critical gaps that slow adoption later. Where several units are involved, MITRE uses the lowest level among them, unless one unit does that work for everyone.
- Put the data scores next to them. Run, or reuse, a data maturity assessment for the knowledge areas in the table above, and check each AI stage against its row.
- Plan one level up in one or two areas. Microsoft’s guidance describes a common approach: pick one or two priority pillars and move each up one level.
- Repeat it on a fixed cycle. MITRE warns that fully reaching the target levels could take years, especially in larger organizations, so each round measures progress rather than a finish line.
Data Maturity Pro’s free AI Readiness and Maturity Assessment uses the five levels above across eight areas, with 16 questions in the Quick version (about 10 minutes) and 56 in the Detailed one (about 45 minutes). It averages each person’s answers before combining them, and a team average of 2.5 to just under 3.5 counts as level 3. It reports an overall level as well as the area scores, so read step 4 before you quote the overall.
AI maturity or AI readiness?
Readiness asks whether you can start one AI project now, with the data and rules it needs; our guide to the AI readiness assessment scores a single use case. Maturity asks how well the organization runs AI across all its projects, and it only means something once there are several. The words blur in practice: Gartner’s FAQ says its maturity model informs the actions that raise AI readiness and value realization, and our own assessment carries both words in its name because a first run doubles as a readiness baseline.
Governance has its own maturity, which our AI governance maturity model scores with a rubric. What goes inside the governance itself, and which standards describe it, is in our post on the AI governance framework. Analytics without AI has a ladder of its own too, set out in the analytics maturity model.
DAMA-DMBOK is a trademark of DAMA International, and DCAM of the EDM Council (now the EDM Association). Level Up Your Data is not affiliated with or endorsed by either body, and the stage-to-data mapping in this post is ours.
Frequently asked questions
What are the levels of AI maturity?
Each model names its own. MITRE uses Initial, Engaged (or Adopted), Defined, Managed and Optimized; Microsoft’s agentic model runs from Level 100 Initial to Level 500 Efficient. Our five are Initial, Developing, Defined, Managed and Optimized, where level 3 means documented strategy and governance applied across most of the organization, with solutions that reach production.
What is the Gartner AI maturity model?
It is the model behind Gartner’s AI Maturity Assessment toolkit. Gartner’s public page groups organizations into five stages, from Foundational to Transformational, and scores seven pillars, data and governance among them. The heat map and the analyst recommendations come with the in-depth report for Gartner clients.
How often should you reassess AI maturity?
MITRE’s guide puts the interval at six to 12 months, which it calls likely sufficient in most cases. Reassess sooner when something large changes, such as a first model in production or a new regulation that applies to you, since a move in one area can make another the new limit.
Can you be mature in AI with low data maturity?
In our view, not for long. Models can reach production on hand-built extracts, but the problems arrive later: the Google study traced compounding downstream failures to data issues and described them as opaque and delayed, with poor indicators. Score the data areas a model uses before you claim the stage it sits in.
Summary
The published models compared here all use five levels; they differ on the names, on the pillars, and on how much of the model you can see without paying. What they leave thin is the data under each stage. Score your AI areas and the data areas from the mapping table side by side, then pick the one area whose low data score blocks your next stage.