An AI skills matrix is a grid that tells you who gets trained next and in what. For a finance, marketing, legal or operations team, the version that works has five capability rows, three levels, and a named owner — and it is scored from a short observed task rather than a survey.
The templates on the first page of search results are larger: six skill categories, five levels, a technical build tier, a four-tab workbook. They are not wrong so much as unsurvivable: six rows scored for twenty people is 120 cells, which gets filled once, shown on a slide, and never scored again. The matrix only earns its keep on the second scoring, when you can see which gaps closed.
The decision the grid exists to make
A skills matrix is a sequencing tool. It answers one question — given a fixed budget and 30 people, who goes first and on what — and a second one later, which is whether anything moved.
That narrow purpose rules a lot out. It is not a performance-review input: the moment people believe it is, the scores stop being honest. Nor is it a dashboard, since usage telemetry only tells you who opened an assistant. It is the operational artifact underneath the argument that capability has to be built rather than purchased, which our case against treating license count as capability makes at the level of strategy.
There is good evidence the artifact is missing. In Skillsoft’s Workforce Readiness Report: AI Edition, published June 2026 from a survey of 2,000 full-time employees, managers and executives, only 11% of employees reported receiving formal skills assessments of any kind, and 69% were only somewhat or not very clear on which skills matter in their role. Those findings describe the same hole from both ends: people do not know what they are meant to be good at, and nobody has checked.
Why the standard templates break on a non-technical team
Three structural problems, all fixable by deletion.
The build tier does not belong here. Nearly every tiered AI competency framework keeps a technical layer — fine-tuning, retrieval architecture, model operations. Those rows make the grid look complete and make it useless for a marketing department: nobody there will be scored above zero on them, and they survive every revision because deleting them feels like lowering the bar. Cut them. If your company needs those capabilities they belong to engineering, with their own instrument.
Tool-specific rows stale faster than the matrix cycles. A row for one assistant’s interface is out of date before the second scoring. Score the transferable behavior and the row survives a vendor switch.
Five levels cannot be told apart. The published ladders are reasonable on paper; the appliedAI Institute’s AI Skills Framework runs Unknown, Knows about it, Can use it, Can adapt it, Lives it. In practice two managers scoring the same person agree on three bands and argue about five, and every argument is an hour spent not training anyone.
What is left is small on purpose: one spreadsheet tab, one afternoon.
The five rows that earn their place
Each row describes something a person does to a piece of work, which is why each one can be observed rather than inferred.
| Row | What it covers | Why it is in |
|---|---|---|
| Task framing | Turning a vague request into an instruction with context, audience, constraints and a sample of good | Largest single determinant of output quality |
| Verification | Checking claims, numbers and citations against a source before work leaves the desk | The failure mode with real downside |
| Data handling | Knowing what may be pasted where, and where that line is written down | Cheap to teach, expensive to skip |
| Judgment about fit | Deciding which tasks are worth handing over, and which cost more to review than they save | Using an assistant where it pays, not everywhere |
| Workflow redesign | Rebuilding a recurring task around the tool rather than bolting the tool onto the old sequence | Where the hours come from; the only row that compounds |
Governance sits inside data handling, because a separate governance row collects policy awareness nobody can act on. Prompt craft is not a row either; it is the visible surface of task framing, and naming the underlying skill keeps the row from becoming a quiz about syntax. Six rows is the ceiling.
Three levels, anchored to behavior
Level descriptors have to be observable, or two scorers will not agree. Write them as things a person does, not as adjectives.
| Level | Task framing | Verification | Workflow redesign |
|---|---|---|---|
| 1 — Aware | One-line request, accepts the first output | Reads for tone, does not check facts or figures | One-off tasks only |
| 2 — Working | Supplies audience, constraints and an example; iterates two or three times | Checks every number and named claim against a source | Has rebuilt one recurring task around the tool |
| 3 — Fluent | Builds a reusable instruction for a task the whole team does | Knows which claim types this tool gets wrong, and checks those first | Has redesigned a task and taught someone else the new sequence |
Write the same descriptors for data handling and judgment about fit before scoring anyone. Without them you have a grid of opinions.
Two rules keep the scale honest. A person is at a level only if they do it unprompted on their own work, not when watched and coached. And Level 3 requires teaching someone else: a department’s Level 3 people are reinforcement capacity you would otherwise buy.
Required levels by role
Not every role needs every level, and pretending otherwise is how a budget gets spent on the wrong people. Set required levels before scoring anybody, so the gap arithmetic is not reverse-engineered from the results.
| Role | Framing | Verification | Data handling | Judgment | Redesign |
|---|---|---|---|---|---|
| High-volume producer | 3 | 3 | 2 | 2 | 2 |
| Specialist contributor | 2 | 3 | 3 | 2 | 1 |
| Team lead | 2 | 2 | 3 | 3 | 3 |
| Department head | 1 | 2 | 3 | 3 | 2 |
| Coordinator or analyst | 3 | 2 | 2 | 2 | 3 |
The pattern worth noticing: the people who need the highest hands-on levels are rarely the most senior. Frameworks that ladder capability to seniority get this backwards, and a department following them trains its leadership while leaving its highest-volume producers at Level 1.
Why a self-rating will not score this
The default method is a self-assessment form, sometimes with a manager column beside it. It is fast, free, and measures the wrong thing.
The clearest evidence is the gap inside a single survey. Skillsoft found that 86% of employees use AI tools at work while only 24% feel fully equipped to use them effectively — and that 77% of leaders believe their organizations have set employees up for success. That 53-point spread is not a measurement; it is two acts of self-report, and a matrix scored the same way inherits it.
Self-rating also fails in both directions at once. Confident people rate themselves at Level 3 on verification because nobody has shown them what a missed citation looks like; careful people rate themselves at Level 1 because they know what they do not know. The errors do not cancel, so the grid tells you to train the cautious and leave the overconfident alone.
The alternative that measurement vendors prescribe is behavioral telemetry — prompt length, session depth, share of work done with assistance. It is honest about the problem, and it assumes an analytics platform ingesting mail, calendar and tool data that a 150-person company does not have and should not buy for this. Our survey of fluency assessment tooling covers the market. There is a method in between.
The twenty-minute work sample
Score the grid from a short observed task on the team’s own work: one person, twenty minutes, one scorer watching.
Pick a task the person does every week — a client update, a variance commentary, a case summary — and hand them a real input from their own queue with anything confidential removed. Ask for the deliverable they normally produce, with the tools they normally use. Then stay quiet, because the moment you prompt, you are scoring your own prompt.
What to write down: whether the instruction carried audience, constraints and an example; what they did with the first output; whether any number or named claim got checked and against what; what they pasted and whether they hesitated; and whether the sequence was the old one with a tool inserted or a rebuilt one.
Three constraints make it hold. Use two scorers on the first three people, because inter-rater agreement is what makes later scores comparable. Score the artifact alongside the behavior: a well-framed task that produced an unchecked error is a Level 2 on framing and a Level 1 on verification. And promise before the first session that this does not reach a performance review — one breach ends the method permanently.
Twenty minutes per person across 30 people is ten hours over two weeks — the whole cost of knowing what your team can do. What to capture before training is in our guide to the pre-training measurement window, and the instrument itself in our framework for scoring a team before and after.
Reading the filled grid
Subtract current from required per cell. The shape of the gaps matters, not the total.
A row short across almost everyone is a curriculum item. If 24 of 30 people sit at Level 1 on verification, that is one session for the whole department — and the highest-return session you will run, because it removes the risk that can stop the program.
A row short for a few people is coaching. Three people below requirement on framing need an hour each with someone at Level 2.
A cell above requirement is not a surplus. Some templates total the gaps and treat a negative number as a win, which quietly rewards over-skilling; a person at Level 3 where Level 1 is required is an unassigned teaching resource.
A row short only at team-lead level is the most expensive pattern to ignore, because everyone downstream then picks their own targets.
Resist the single composite score. A department average of 1.8 is unactionable and invites the between-team comparison that destroys honest scoring. Report by row and by role; which figures belong on a monthly page is covered in our rules for setting enablement KPIs.
What it costs, and who owns it in month seven
This is where published templates go quiet. They prescribe a full reassessment every six months and rarely price the cycle, which is why most matrices die after one pass.
| Activity | Hours, 30-person department |
|---|---|
| Writing descriptors, setting required levels | 4 to 6, once |
| Scoring at 20 minutes per person, plus notes | 10 to 12 |
| Calibrating two scorers on the first three people | 2 |
| Reading gaps into a training sequence | 2 to 3 |
| Re-scoring at six months | 10 to 12 |
Roughly twenty hours to stand up and twelve per cycle. Against a team workshop, which runs between $2,000 and $15,000 in the current market, that is defensible overhead — and the only thing that makes the spend reviewable.
Ownership decides whether the matrix survives. At 50 to 500 people there is usually no L&D function and no Chief AI Officer, so this lands on an operations lead or a chief of staff as a second job. Name that person in writing, set the re-score date before the first scoring, and put the grid where the department can see its own rows. An artifact nobody owns on a specific date is not a measurement system; it is a slide. How the grid feeds the rest of a program is set out in our measurement guide for proving capability moved.
One pattern from delivery: the departments that keep their matrix re-scored one row at three months rather than the whole grid at six, so the full re-score had a precedent instead of being a first attempt.
Frequently asked questions
What is an AI skills matrix?
A grid of AI capability rows against proficiency levels, scored per person, used to decide who gets trained next and to check later whether the gaps closed. For a non-technical department: five rows, three levels, a named owner, a re-score date.
How many capabilities and levels should the grid have?
Five rows and three levels, with six rows as a hard ceiling. The constraint is re-scoring rather than completeness: a grid that takes more than an afternoon to fill for 30 people gets filled once, and one scoring tells you nothing about movement.
Should a non-technical department include technical AI rows?
No. Delete fine-tuning, retrieval architecture and model operations rather than leaving them as an unreachable top tier. Nobody in finance or marketing will be scored on them, and they survive every revision because cutting them feels like lowering the bar.
Can we just use a self-assessment survey?
Not as the score. Self-ratings measure confidence and err in both directions at once: the overconfident rate themselves high on verification, the careful rate themselves low on everything. A survey is good for one thing — asking which tasks people would most like to hand over.
How do we score a capability without a formal test?
Watch a twenty-minute real task. Hand the person a genuine input from their own queue, ask for the deliverable they normally produce, stay silent, and score the five rows from what they do. Two scorers on the first three people makes later scores comparable.
How often should the matrix be re-scored?
Re-score one row at three months and the full grid at six. The single-row check keeps the habit alive because it fits in an afternoon and still produces a decision. Set both dates before the first scoring, not after.
Who should own the matrix if there is no L&D function?
One named person, usually an operations lead or chief of staff, with the re-score dates written down and visible to the department. Shared ownership across a manager group produces no ownership, and the matrix dies between the first and second scoring.
Do AI certifications belong in the grid?
No. A certificate records attendance at a curriculum, not what someone does to their own work unprompted, and the grid measures the second thing. Record certifications elsewhere if your budget needs the paper trail.
How does the matrix connect to a budget request?
It supplies the two things a finance reviewer asks for: a specific population with a named gap, and a dated before-figure the same instrument will measure again. “Twenty-four of thirty are below requirement on verification” is fundable; a completion rate is not.
What if the department refuses to be scored?
Treat it as information about the method, not the people. Refusal almost always traces to a belief that scores will reach a performance review. Score the work rather than the person, publish nothing by name, and start with the team lead so the first session is the one with the most to lose.
Key takeaways
- Five rows, three levels, one owner, two dates. A larger grid gets filled once, and one scoring measures nothing.
- The rows are task framing, verification, data handling, judgment about fit, and workflow redesign. Technical build rows come out entirely.
- Write descriptors as observable behavior, require it unprompted on the person’s own work, and make Level 3 require teaching someone else.
- Do not score from a self-rating. The 53-point gap between the 77% of leaders who believe their people are equipped and the 24% of employees who feel equipped is what a self-rated grid reproduces.
- Score from a twenty-minute observed task, two scorers calibrating on the first three people. Ten hours covers a 30-person department.
- Read gaps by row and by role, never as a composite average. A cell above requirement is an unassigned coach.
Working out which five rows describe your department, and whether the training you have been quoted would move any of them, is most of what a first conversation covers. Book a free AI-readiness call and we will go through your roles, your required levels, and the one row worth scoring first.
Dirk Jan van Veen, PhD