An AI enablement maturity model is a grading instrument: it says where your company’s AI capability sits today, which is a different question from what to do next. Most published models grade the wrong thing. They score what a company has built — a data platform, a governance council, a written AI strategy — and a company of 200 people has none of those, should not build most of them, and comes out indistinguishable from a company that has done nothing at all.
The four levels below grade something a mid-market company can verify in an afternoon: what your capability survives. Not whether people are enthusiastic or how many licenses are active, but whether the way work gets done would hold if the facilitator left, if your best user resigned, or if the quarter got busy.
The reason to grade honestly is that almost nobody does. ActivTrak’s analysis of 120,620 employees across 1,009 organizations between Q4 2025 and Q2 2026 found 43 percent using AI at all, 27 percent using it as a research assistant, 14 percent using it to execute tasks, and 2 percent with it integrated into daily workflows (Fortune, August 2026). Observed behavior, not self-report.
Why published AI enablement maturity models misgrade a 200-person company
The serious models are built for the enterprise. MITRE’s AI Maturity Model runs six subject areas across 20 dimensions and five levels, from Initial through Adopted, Defined, Managed, and Optimized (MITRE). Gartner’s version scores seven areas; Credo AI’s governance model has six levels. Each is defensible for a company with 8,000 employees and a chief data officer.
Run a 200-person professional-services firm through 20 dimensions and most answers are “not applicable.” There is no engineering function to score, no data platform, no AI center of excellence, and building one would be a mistake rather than a step up. The score lands at Level 1 whether the firm trained four teams last quarter or has done nothing at all.
MITRE is also the one source honest enough to say that not all organizations need or want level 5 in every area. Every vendor model is implicitly a ladder to the top rung, because the top rung is what the vendor sells.
The second problem is that these models grade declared capacity. A written AI strategy scores points; whether anyone’s Tuesday changed does not. That is the same error as reading a license-usage dashboard as a capability report.
The four levels at a glance
Each level is defined by what the capability is stored in, and therefore what removes it.
| Level | Capability lives in | Survives a busy quarter | Survives the power user leaving | Survives a reorganization |
|---|---|---|---|---|
| 1 Licensed | The invoice | No | n/a | No |
| 2 Individual | Named people | Usually | No | No |
| 3 Team | A written method and a manager’s routine | Yes | Yes | No |
| 4 Institutional | Hiring, onboarding, workflow definitions | Yes | Yes | Yes |
Three independent estimates put the top of this ladder in the same band. MIT CISR’s survey of 721 companies found 7 percent at its fourth and final stage (MIT Sloan, February 2025). ActivTrak observed 2 percent at workflow integration. Deloitte’s 2026 Global Human Capital Trends found 6 percent of leaders claiming progress on designing human-AI interaction at all (Deloitte, March 2026). Three methods, one order of magnitude, all far below how companies grade themselves.
Level 1, Licensed: the capability lives in the invoice
You bought seats. Some get used. Nothing about how work is produced has changed, and if the contract lapsed the only loss is a line item.
Level 1 is not a failure state. It is where every company starts, and most sit here longer than their dashboards suggest, because activity numbers look adequate. ActivTrak’s platform data includes a case where 38 of 142 licensed Microsoft Copilot seats were used more than twice a week (ActivTrak, July 2026). That is a procurement outcome, not a capability.
Evidence to open: the seat report next to the renewal date. If you cannot name one recurring deliverable whose written steps changed since the licenses arrived, you are at Level 1. Why that gap is so common is the case against treating licenses as capability.
Level 2, Individual: the capability lives in four people
Something real is happening. Three or four people have worked out how to use these tools on their own work, they are faster than they were, and colleagues notice. This is where most companies that have made genuine progress sit, and the most misread rung on the ladder, because it produces visible wins.
It is also the fragile one. The method exists only in those people’s heads. Nothing is written and nobody else has been shown the steps, so the capability leaves when they do. Enthusiasm itself is durable — 82 percent of employees who adopted AI in ActivTrak’s sample kept using it — but individual durability is not organizational capability.
Level 2 has a second signature: time gets saved and nothing is done with it. BCG’s 2026 AI at Work research found 42 percent of frontline regular users saving a full day or more per week, and 66 percent of them getting little or no direction on how to use it (BCG, June 2026).
Evidence to open: write down the names of the people the method would leave with. If the list fits on a sticky note, that is your level.
Level 3, Team: the capability lives in a written method
At Level 3 a team’s work is produced differently and the difference is recorded. Someone wrote the steps down. There is a context document, a set of reviewed prompts, and a rule about what gets checked before anything reaches a client. A manager asks to see the method in a regular meeting, which is what stops it drifting. A new joiner learns the method because it exists as a thing to learn, not because they sit near the right person.
This is the first level where capability survives turnover, and the one that corresponds to real financial movement. MIT CISR found enterprises in its first two stages performing below their industry average and those in stages three and four above it, and a later update from the same researchers put the largest financial impact in the move from stage two to stage three.
What Level 3 does not survive is a reorganization. The method belongs to a team; split that team across two functions and the written layer stops matching anyone’s work. The parts that have to exist to hold this level are set out in the anatomy of a company where fluency is real.
Evidence to open: the process document for one recurring deliverable, edited in the last six months, and that team’s weekly meeting agenda, with a line on it where the method gets shown.
Level 4, Institutional: the capability lives in how you hire
At Level 4 capability has moved out of teams and into the company’s defaults. Job descriptions assume the tools. Week one of onboarding names the sanctioned tools and the review rule. Workflow definitions were rewritten rather than annotated, and nobody tracks AI usage separately, because there is no separate thing to track.
Level 4 is rare and does not require a governance apparatus. It does require the written layer to be maintained by whoever owns the workflow rather than by a central function. For the governance vocabulary, the four functions of the NIST AI Risk Management Framework — Govern, Map, Measure, Manage — are meant to be applied iteratively rather than as a gate you pass before starting (NIST).
Evidence to open: last quarter’s onboarding pack and the most recent job description you posted. If neither mentions the tools your teams use daily, you are at Level 3 describing yourself as Level 4.
The three losses that set your level
Grading against 20 dimensions produces a number nobody trusts. Grading against three losses produces an answer you can defend in a meeting. Pick one recurring deliverable — a client report, a proposal, a monthly close — and ask three questions about it.
- If this quarter got unusually busy, would the method survive? Under pressure people revert to the way they know. If no, you are at Level 1 regardless of how the licenses look.
- If the person best at this left next month, would the method survive? If someone would have to work it out again from scratch, you are at Level 2.
- If that team were split across two functions tomorrow, would the method survive? If the written steps would stop matching anyone’s job, you are at Level 3. If they would travel with the work, you are at Level 4.
Answer in that order and stop at the first no. The three questions take an afternoon, cost nothing, and are harder to flatter than a questionnaire, because each one names a specific person or document.
Levels go backward, and most models pretend they do not
Every published maturity model is monotonic. Real programs regress, almost always on the same event: the person carrying the method leaves, and the company drops from Level 3 to Level 2 without anyone recording it. The dashboard does not move, because the remaining users still have active seats.
ActivTrak’s data shows a related trap. Healthy utilization peaks at the task-execution stage, then falls roughly five points once AI is genuinely embedded in workflows, to a level statistically indistinguishable from employees who barely use the tools. A falling activity number can mean regression or real integration, and the number alone cannot tell you which.
A level is a claim with an expiry date. Re-grade once a quarter against the same three questions and the same deliverable.
Which level you should stop at
For a company of 50 to 500 people: Level 3 on the three or four workflows carrying the most document-heavy repetitive work, and Level 1 everywhere else. That is the recommendation, stated plainly, because no vendor model will state it.
Level 4 across a whole company is a reasonable target for an organization with thousands of employees and a function whose job is maintaining it. At 200 people the attempt usually produces a center of excellence with no delivery mandate, a policy document, and a committee, none of which changes how a deliverable gets produced.
The corollary: an even score across every function is the wrong goal. A product team at Level 3 and a facilities team at Level 1 is a correctly allocated program, not an uneven one.
What moving up one AI enablement level costs
Each move has its own cost shape, which is why a single budget line for AI enablement tends to fund the wrong one.
| Move | What it takes | Market cost |
|---|---|---|
| 1 to 2 | Access, permission, a few people with a reason to try | Licenses you already pay for |
| 2 to 3 | Hands-on sessions on the team’s own documents, the method written in the room, a manager routine, a retest weeks later | Roughly $2,000 to $15,000 per session |
| 3 to 4 | Rewriting onboarding, hiring criteria, workflow definitions | Internal time, two to three quarters |
| Beyond 4 | Custom automation of a workflow already running at Level 3 | Roughly $25,000 to $150,000 |
The 2-to-3 move is routinely underfunded and has the clearest evidence behind it. It fails when training happens away from the work: generic prompting produces Level 2 enthusiasm, while a session on the team’s own live documents produces a written method. Automating a workflow still at Level 2 buys an expensive version of a process nobody agreed on.
Grade each team, not the company
A single company-wide level hides the only information you can act on. ActivTrak’s data from one financial-services rollout — same licenses, same program — put the share of non-users at 20 percent in risk management, 45 percent in retail banking, and 25 percent in wealth management. One rollout, one claimed level, three realities, and a company-wide average that describes none of them.
Grade four or five teams separately with the three-losses test, write the levels in a column, and the budget decision makes itself: fund the 2-to-3 move for the team whose work is most document-heavy, leave the Level 1 teams alone, re-grade in a quarter. The sequencing behind that order is in the five stages from pilot to default behavior, and the full program view in our AI enablement playbook.
If you want a second opinion on the grade before spending against it, that is what a free AI-readiness call is for. Book one here and bring one deliverable and the three questions.
Frequently asked questions
What is an AI enablement maturity model?
A staged framework for assessing how far a company has moved from owning AI licenses to having capability that holds without supervision. Published models grade areas such as strategy, data, governance, and culture. The four-level version here grades one axis a mid-market company can verify: what the capability survives.
How many levels should an AI enablement maturity model have?
Four is enough under 500 people. Five- and six-level models exist because enterprises distinguish degrees of institutionalization a 200-person company will never reach. More levels do not add precision at this size.
How do I tell which level we are at without running a survey?
Ask whether the method on one recurring deliverable would survive a busy quarter, the best user leaving, and the team being split up. Stop at the first no, then confirm with documents: the seat report, the process doc, the meeting agenda, and last quarter’s onboarding pack.
Which level should a company of 50 to 500 people aim for?
Level 3 on the three or four workflows carrying the most repetitive document work, and Level 1 everywhere else. Level 4 company-wide needs a function dedicated to maintaining it, which companies this size do not have and should not create.
Can a company go backward a level?
Yes, and it is the most common failure. A team drops from Level 3 to Level 2 when the person maintaining the written method leaves and nobody picks it up. Usage dashboards miss it, because the seats stay active.
How long does it take to move up one level?
The 2-to-3 move takes one to two quarters for one team: a session, the written method, a manager routine, and a retest weeks later to confirm it survived without the facilitator. The 3-to-4 move takes two to three quarters, because it touches hiring and onboarding cycles.
What does it cost to move from Level 2 to Level 3?
Facilitated hands-on sessions run roughly $2,000 to $15,000 each at market rates depending on format and cohort size. The written method and the manager routine cost time rather than money, and they are the parts that fail when skipped.
How is this different from an AI readiness assessment?
A readiness assessment asks whether you are prepared for one specific investment, and is answered once. A maturity model tracks capability over time against the same repeated test. Readiness is a gate; maturity is a trend line.
Do we need governance in place before we can move up?
No. Policy and capability advance together, and the NIST framework is explicitly iterative rather than a prerequisite. A one-page review rule people follow beats a thirty-page policy that arrives before anyone has a method to govern.
Key takeaways
- Published AI enablement maturity models grade what a company has built. At 50 to 500 people that returns a Level 1 score whether you trained three teams or none.
- Grade on survival instead: what the capability lives in, and so what removes it — the invoice, four people, a written method, or the way you hire.
- Three estimates agree the top rung is rare: 7 percent at MIT CISR’s final stage, 2 percent at workflow integration across 120,620 employees, 6 percent of leaders in Deloitte’s 2026 research.
- The three-losses test grades a real deliverable in an afternoon: a busy quarter, the best user leaving, the team being split. Stop at the first no.
- Level 3 on a few document-heavy workflows is the right destination at this size, and levels regress when the person carrying the method leaves, so re-grade quarterly.
Dirk Jan van Veen, PhD
