An AI enablement scorecard is one page carrying six rows. Each row holds one number, names the system it came out of, names the person who produces it, and states the value that turns the row red. Drop any of the four and the page gets built once, admired in a kickoff meeting, and never updated again.
Search the phrase and you mostly get something else: gated maturity diagnostics that rate a company 1.0 to 5.0 before it has bought anything, plus tools that score sales reps on call quality. Neither answers the question a COO is holding, which arrives about nine months after the licenses went out and sounds like “so did it work.”
This is the artifact for that question, row by row.
Two different documents share the name
A readiness scorecard looks forward. It rates leadership buy-in, data, infrastructure, skills and governance on a maturity scale, and exists to tell a board whether to start. Cisco surveyed 8,000 senior leaders across 30 markets and found about 13% of organizations qualifying as Pacesetters, four times as likely to move pilots into production.
An AI enablement scorecard looks backward. The licenses are bought, the sessions are run, and somebody has to show what changed. The two share almost no rows: readiness asks whether the conditions exist, enablement asks whether behavior moved.
If you have already trained people, you want the backward-looking page. That is the argument behind why buying licenses is not the same as building capability: provisioning is an event, capability is a measurement, and only one shows up in a console.
There is also no standard model here. Published frameworks run from three dimensions to eight, so stop shopping for the canonical one. Pick rows you can produce, and defend the choice.
The six rows
Six rows, in the order they move. The first changes in weeks, the last in quarters.
| Row | The number it carries | Where it comes from | Who produces it | Red at |
|---|---|---|---|---|
| Reach | Share of trained people who used an assistant on real work in the last 5 days | Tool admin console, usage export | Whoever holds the admin seat | Under 30% at month 6 |
| Habit | Share active in at least 9 of the past 12 weeks | Same export, same test each quarter | Same person | Under 10% at month 12 |
| Capability | Share who can show one pattern they reuse on their own work | A 10-minute check by their manager | Functional managers | Under 20% at month 6 |
| Named workflows | Count of recurring deliverables where an assistant is a standard step | A hand-kept list | The enablement sponsor | Zero at month 6 |
| Output | One dated before-and-after on a real deliverable | The function that owns the deliverable | That functional VP | Nothing dated at month 12 |
| Safe use | Share of sensitive work running through approved tools, and escalations raised | Approved-tool logs plus one incident line | Whoever owns the policy | Any sensitive work in personal accounts |
Three notes on the rows.
Habit already has a published definition. Microsoft’s analytics reference sets a habitual user as somebody averaging 1 to 14 assistant actions a week and active in at least 9 of the past 12 weeks. Borrow the test whatever your stack is. Nine weeks out of twelve cannot be faked by one keen person running bulk jobs, and it survives a quarter-end lull.
Capability is the row that costs you something. The others come out of systems. This one needs a human to read a human’s work for ten minutes, and it is the one row separating someone who uses an assistant from someone who got better at the job. The Kirkpatrick model puts behavior at level three: whether people perform the critical behaviors and are held to them. A console cannot see that; a manager can.
Output is one row, not five. One deliverable, timed before and after, beats six directional arrows, and it is the row a CFO reads first. The method behind picking it sits in our guide to proving capability moved; the thresholds above come from the bands in what good looks like at 6 and 12 months.
The two columns nobody fills in
Every template online has rows and targets. Almost none has a source column or an owner column, and those two decide whether the page exists next quarter.
Source. A number with no named system behind it cannot be argued with, which sounds like a strength and is fatal. Pick one: this console, this export, refreshed on this day.
The same word also means one thing in one console and another somewhere else. Recon Analytics measured how many of more than 150,000 US paid subscribers with workplace access turned into real users: 83.1% for ChatGPT against 35.8% for Copilot and 34.0% for Gemini. Same companies, same policies, a 47-point spread. “Adoption is 70%” is not a fact until you say which console said so.
Owner. A row owned by a committee is owned by nobody. At 50 to 500 people there is no Chief AI Officer, so the honest version is a name per row. The admin seat takes usage, line managers take capability, a functional VP takes output, whoever holds the policy takes safe use.
Fill both columns before you write a single target. A scorecard is a list of standing requests to named people, and the targets are the easy part.
The baseline you cannot recover later
Most of this page can be rebuilt late. One row cannot.
The before-time on the deliverable you plan to claim is gone the moment training starts. Ask a team in March how long the January close pack took and you get a guess shaped by the answer they think you want. Before the first session, pick one recurring deliverable and time it honestly: hours of effort, calendar days end to end, how often it gets reworked. Twenty minutes of work decides whether the output row can carry a number at all.
The second trap runs the other way. Your usage baseline was never zero. The Census Bureau puts AI use at 32% of US firms with 100 to 249 employees and 37% above 250, and far more workers use these tools than their employers report adopting. Measure reach before the sessions or you book other people’s habits as your result.
Started already and took no baseline? Pick a deliverable the training has not touched, time it this week, and train that team next. What a capability audit covers and what it costs prices the structured version, but the output baseline is free and you get one shot at it.
What stays off the page
Six rows means cutting things that feel like progress. Each exclusion has a sponsor somewhere in the building.
| Keep off | Why it looks good | What it measures |
|---|---|---|
| Course completion rate | Hits 90% and feels like coverage | Attendance. Nobody got better by finishing a video |
| Session satisfaction score | Always high, cheap to collect | How the room felt that afternoon |
| Seats purchased or assigned | Big number, finance already has it | Procurement, not use |
| Prompts or messages per user | Looks like depth | Volume. One person pasting spreadsheets beats ten doing real work |
| Self-reported daily use | Cheap survey, flattering number | Self-image. Reported use runs well above console figures |
| A peer-benchmark percentile | Sounds like context | Someone else’s sample, usually firms ten times your size |
The last two are the common failure. Published benchmarks mostly come from companies above 1,000 employees with an AI owner and a change function you do not have. Collect self-report if you like, and put the console number in front of the board.
Scrap learning, the share of training that never reaches the job, is a long-running problem in corporate learning rather than anything new about AI, and completion rates are the metric that cannot see it.
Do not roll it up into one score
The maturity genre loves a composite. One framework doing the rounds weights two of its eight dimensions at 1.5 times the rest and divides by nine, producing a tidy 2.7 out of 5.
Skip it. The weights are invented, and nobody defends 1.5 over 1.3 when a VP pushes back. Worse, a composite buries the row that should stop the meeting: strong reach, strong habit and no moved output scores respectably and is a real problem.
Six numbers, six colors, no total. If procurement or a client wants the page mapped onto a recognized framework, the safe-use row fits the govern and measure functions of the NIST AI Risk Management Framework. Do that as an appendix, not as the page’s organizing idea.
What the page costs to run
The question that decides whether the scorecard survives, and one the corporate-training literature never answers. What we budget with clients, per quarter, after the first build:
- Usage export and the two console rows: 30 minutes for the admin seat
- Capability checks: 10 minutes per manager, so under two hours across a dozen
- Workflow list: 20 minutes with the sponsor, mostly deletions
- Output row: 30 minutes once the first timing exists
- Safe-use row and the escalation line: 20 minutes
Call it half a day of spread-out effort, most of it not yours. The first build runs closer to a day, because each row’s source has to be found and the baseline timing has to happen. Anything costing much more than that gets quietly abandoned, which is why 40-metric templates never see a second update.
Reading a red row
Targets without failure conditions produce a status report. Each red threshold implies its own action, and two of the three are not training problems at all.
Reach or habit red, capability fine. People can do the work and are not doing it. Usually friction or permission: the tool sits outside the apps they live in, or their manager never asked. More training will not move this. A 45-minute clinic on the three tasks people are stuck on will.
Capability red, reach fine. People are using an assistant and have not got better. The honest reading is that the sessions taught the tool instead of the work. Our walkthrough of how to diagnose a push that did not work splits these two cases, which look identical from a distance and need opposite fixes.
Everything green, output red. The hardest to accept, because the dashboard looks healthy. Capacity was created and nobody redirected it. The fix is a management conversation about which backed-up work the freed hours land on. McKinsey keeps finding that workflow redesign separates the small group of AI high performers from everyone else, and redesign is nobody’s job by default.
Safe use red. Fix it before anything else, whatever the other five rows say.
Two reds at one review is the signal to decide rather than iterate, and the rules for evaluating a pilot before you scale it cover that call. Write the thresholds down before the program starts, or they get set to whatever the numbers happened to be.
The page at month nine
A filled-in scorecard for a 180-person company, nine months after hands-on sessions with 60 people. The values are illustrative; the shape is the point.
| Row | Number | Verdict |
|---|---|---|
| Reach | 52% used an assistant on real work in the last 5 days | Green, inside the on-track band |
| Habit | 21% active in 9 of the past 12 weeks | Amber, early for month 9, wants 25% |
| Capability | 38% can show a pattern they reuse weekly | Green |
| Named workflows | 3 recurring deliverables | Green |
| Output | Monthly reporting pack, 11 hours to 6, timed in February and September | Green, the row the CFO reads twice |
| Safe use | Approved tools for all client material, 2 escalations closed | Green |
One amber, no reds, one dated output, on a single screen. Note what it does not claim: no composite score, no peer percentile, no revenue attribution. Six numbers somebody can be asked about by name.
Frequently asked questions
What is an AI enablement scorecard?
A one-page document tracking whether an AI enablement program changed behavior and output, across six rows: reach, habit, capability, named workflows, one output change, and safe use. Each row carries a number, the system it came from, a named owner, and a threshold that turns it red.
How is it different from an AI readiness scorecard?
A readiness scorecard rates conditions before you invest, scoring leadership, data, infrastructure and governance on a maturity scale. An enablement scorecard measures results after you invested. If licenses are out and sessions are run, the readiness version answers a question nobody is asking.
How many metrics should be on it?
Six, and no more than eight. Templates with 40 candidate metrics fail on maintenance rather than logic: every row costs somebody time each quarter, and a page that takes much more than half a day to refresh stops getting refreshed.
Where do the numbers come from if we have no AI owner?
Three places, none needing a dedicated role. The tool admin console gives reach and habit. Line managers give capability with a ten-minute look at one person’s work. The function owning a deliverable gives the output row. Assign each row to whoever is already closest to that system.
Do we need a baseline, and can we get one after training?
You need one, and all of it is recoverable late apart from the before-time on the deliverable you plan to claim. Memory of how long something took is unreliable once people know what answer is wanted. If you have started, pick a deliverable the training has not reached, time it this week, and train that team next.
Should completion rates and satisfaction scores be on the scorecard?
No. Both run near-universally high, neither predicts anything, and both crowd out rows that do. Track them if your learning system produces them free, and keep them off the page that goes to finance.
How often should it be updated?
Quarterly for the whole page, with the two usage rows worth a monthly glance in the first six months while habits are forming. The output row moves on its own clock and often changes once a year, which is normal and not a reason to swap it for something livelier.
How do we turn the scorecard into an ROI number finance will accept?
Pair the output row with employer hourly cost, using a published figure such as the Bureau of Labor Statistics data on employer costs for employee compensation rather than salary alone. Put the assumptions on separate lines so each can be argued with. A single confident percentage invites somebody to hunt for the input you buried.
Key takeaways
- Six rows, each with a number, a source system, a named owner and a red threshold. A row missing one of the four will not get filled in next quarter.
- Most scorecards published under this name rate readiness before purchase, which answers a different question.
- Source and owner are the columns templates omit, and they decide whether the page survives a quarter.
- Time one recurring deliverable before the first session. It is the one measurement you cannot reconstruct later.
- Keep completion rates, satisfaction scores, seat counts and self-reported use off the page. They are high, flattering and blind to transfer.
- Do not roll the rows into a composite. A respectable average is how a healthy-usage, zero-output program hides.
If you want this page drafted against your own tooling, managers and the deliverable you need to move, book a free AI readiness call. We will tell you which rows you can fill in this week, and which one to go and baseline before you train anybody else.
Arthur Wandzel
