AI enablement KPIs: the ones that matter and the ones that mislead
Eight AI enablement measures from job postings and adoption tools, sorted into keep, keep with care and drop. What each answers, what it misses, who names it.
Key finding
Start with three: time per task against a baseline, quality and rework on the same sample, and the share of a team still using AI after 30 days. Logins and trained-champion counts show effort, not a change in the work. Of 55 postings for in-house AI enablement leads, 39 name a way to measure, and one states a number to hit.
On this page
- Measures sorted
- 8, from 71 postings and 8 tools
- Verdicts
- 2 keep, 4 keep with care, 2 drop
- Postings that state a number to hit
- 2 of 62
Key takeaways
- Of the eight measures the postings and tools name, two show the work changed (time per task and quality), and they are the two that need a baseline and a sample.
- The easiest measures to collect are the furthest from the work: logins, active users and trained-champion counts say who showed up, not what happened.
- Start with time per task, quality and rework, and the share of a team still using AI after 30 days, all on one workflow with a baseline taken first.
Most AI enablement scorecards are built from whatever the tools already count. That gives you logins, active users and licenses issued, because those are cheap. The measures that show the work changed are time per task, quality and cost per task, and they need a baseline and a sample. So most dashboards lack them.
We sorted the eight measures that postings and tools name and gave each a verdict: keep, keep with care, or drop. Then we picked three to start with. The method comes first, because the verdicts are only as good as what was read.
How we sorted them
AI enablement KPI
A number an AI enablement lead reports to show whether AI is being adopted and whether the work changed. We call the thing counted a measure, and a KPI is a measure someone has attached a target to. That distinction matters below: most postings name measures, and almost none names a target.
postings for an employer’s own AI enablement lead that name a way to measure the work: KPIs, ROI, a business case, a dashboard or value realization. Read on October 16, 2026, from a sample of 62 public postings.
postings that state a number to hit. One asks for weekly active usage to rise from 20% to 70% within 90 days, in a customer-facing role at Liminal. One asks to triple the company’s ability to build products in 12 months, at Capital on Tap.
In one software company of about 200 people, license data showed more than 95% had accessed AI tools, and session data showed more than 30% of users barely used them. The vendor’s own write-up of one customer.
Time to make a PowerPoint with Copilot against without, in a UK government pilot. The Copilot version scored 1.5 on accuracy against 5 without it. Six Copilot users and five non-users did the observed tasks in all.
What each measure can tell you
| Who uses it | Use repeats | Work is faster | Result is right | Costs less | Where to act | |
|---|---|---|---|---|---|---|
| Counts of use | fully | none | none | none | partly | partly |
| Depth of use | fully | fully | none | none | none | partly |
| Adoption by team | fully | partly | none | none | none | fully |
| Time per task, before and after | none | none | fully | none | partly | partly |
| Quality and rework | none | none | partly | fully | partly | fully |
| Cost per task | none | none | partly | partly | fully | partly |
| Use cases through each stage | none | none | none | none | partly | fully |
| Coverage of trained champions | partly | none | none | none | none | partly |
none partly fully · Our reading, not a measurement. Full: the measure answers it directly. Half: it gives a partial or indirect answer, for example unused licenses hint at cost, and time per task is one input to cost. Empty: it cannot answer it alone. Cells follow the definitions in the posts and pages cited below.
Read the rows. No measure answers more than two of the six questions in full, and none answers both “work is faster” and “result is right”. That is the argument for a small set rather than a favorite.
Measures of use: counts, depth, teams
Counts of use
Every tool in our adoption analytics read counts use, and Microsoft’s Copilot Dashboard counts active users over the past 28 days. Employers ask for it too. Flutter’s posting, in the first ten we read, asks candidates for a record of hitting adoption targets and cites over 80% user adoption. Of the 62 postings, the one use target is Liminal’s 20% to 70%.
A count answers who touched the tool and misses what came after. See the 95% and 30% above. Second Talent’s explainer on the role says the same about rollouts: a license produces a spike in logins and no change in how work is done. Drop it as a result. Keep it as a footnote, because every other measure needs it for context.
Depth of use
Depth asks whether use repeats. Microsoft gives the most exact definition we found: a power user takes 15 or more actions a week in at least 9 of the last 12 weeks. ActivTrak reads time in an application at session level and places teams on a scale from Stage 0, non-users, to Stage 5. The Johnson Controls draft KPIs, in our earlier read of that posting, include active use after 30 days by role.
It misses benefit. Fifteen actions a week can be fifteen quick prompts. It also needs logs, because people are poor reporters of their own use. In the short-task experiments of Yu and colleagues, 498 people used AI on 47% of tasks against 33% they predicted. At Okta, engineers’ self-reported Copilot use correlated with logged use at r = 0.48, which leaves a lot of room for error. Keep with care. Write down what “active” means and set it by role before the first report.
Adoption by team
Microsoft, Worklytics, Larridin and ActivTrak all split use by team or role. It answers where use is thin, which is the first thing a lead needs. It cannot say why. Worklytics compares high-adoption with low-adoption teams, and the pages we read describe no baseline or staggered rollout, so a gap may be teams that differed before AI. At Okta the survey group used Copilot more than license holders as a whole, so who answers shapes the picture too.
There is a privacy edge. Microsoft hides groups under a minimum size, and Torii opens a person’s own usage view on a click. Keep with care. Use it to choose where to look and to order a rollout in waves with a staggered rollout. Do not read a gap between teams as an effect of AI.
Measures of change in the work
Time per task, before and after
This asks whether the same task now takes less time. Spectris lists hours saved among its metrics. Second Talent says that without a measurement of how long a task took beforehand, every claimed gain is an anecdote. Among the 62 postings, the only target that touches time is Capital on Tap’s: baseline how fast the company builds in the first 30 days, then triple it.
Estimates are the trap. Microsoft’s assisted hours are actions multiplied by six minutes, so doubling the summaries doubles the hours. Measured, the numbers move the other way as often as not. In METR’s trial, 16 experienced developers took 19% longer per issue with AI (246 issues) after expecting 24% less time. In Yu and colleagues, 600 people predicted AI would save 55.7 seconds a task, and across 1,001 others it saved 7.5. In the UK pilot, report summaries took Copilot users 12:37 against 41:34, and an Excel analysis took 25:01 against 20:33.
It misses whether the result was right. Keep it, as a timed sample of one workflow against a baseline. Not as an estimate.
Quality and rework
Quality asks whether the result is as good, and rework asks how much of it people redo. Deepgram asks for productivity and quality, not activity. Jellyfish says it links AI use to delivery quality from engineering systems. The posting facts we recorded do not mention rework at all.
Faster work can score worse, as the PowerPoint above did. Who scores also changes the answer. At The Hague University of Applied Sciences, staff given a GenAI platform rated their own quality no higher, while automated evaluators rated it higher (we read an abstract page only). At Okta, the one output count that rose significantly was reviews performed, which fits time moving into review. It does not prove it. This measure costs reviewer time and says little about speed alone. Keep it. Score a sample with the team’s own check, applied by someone other than the author.
Cost per task
Our guide to measuring AI impact calls cost per task the closest thing to one metric, because it combines people’s time, model use, review and rework. Few sources print one. Torii’s dollar figures are an estimated token value at standard API rates, not your invoice. Microsoft’s assisted value multiplies assisted hours by a default $72 an hour. Jellyfish shows token spend by tool, team or initiative, which is spend, not cost per task.
None of the eight studies in our productivity read measured cost per task with review and rework included. We did not search the 62 postings for the term, so we cannot say how many name it. It needs a labor rate, a baseline and a quality check first. Keep with care, and add it after the two measures above are running.
Measures of the pipeline and the people
Use cases through each stage
The use case pipeline is the most common duty in the postings: 42 of the 55 internal ones (76%). Spectris asks its lead to keep one list with a business owner, expected value, risk and status. Johnson Controls counts functions that convert to a production use case. The measure answers whether work moves from idea to production, and where it stalls.
It counts the pipeline, not the effect. A use case in production can sit unused, and the stage names are yours to define. Keep with care. Count a use case as done only if it is still in use after 30 days.
Coverage of trained champions
Training, champions and community appear in 33 of 55 internal postings (60%). Spectris and J.D. Power put champions in the business units. The nearest numbers are HPE’s training completion, Synchrony’s participation and Flutter’s time-to-competency. In what we recorded, no posting names champion coverage as a target.
It shows effort. A trained person may use AI daily or never. Drop it as a target and keep it as a diagnostic, for example to explain a team where depth of use is low. A champion network is a good thing to run. That is a different claim from it being a good thing to score.
The eight side by side
| Measure | Answers | Misses | Named by | Verdict |
|---|---|---|---|---|
| Counts of use | Who touched the tool | Whether use repeated or helped | Microsoft, Worklytics, Larridin, Torii, Nudge; Flutter; Liminal | Drop as a result |
| Depth of use | Whether use repeats | Whether it helped | Microsoft, ActivTrak; Johnson Controls draft | Keep with care |
| Adoption by team | Where use is thin | Why, and cause | Microsoft, Worklytics, Larridin, ActivTrak | Keep with care |
| Time per task, before and after | Whether the work is faster | Whether it is right or cheaper | Spectris, Second Talent; Capital on Tap; Nolli | Keep |
| Quality and rework | Whether the result is as good | Speed alone | Deepgram, Jellyfish, Nolli | Keep |
| Cost per task | Whether it costs less per correct result | Nothing alone, but needs the others first | Nolli; Microsoft and Torii give estimates | Keep with care |
| Use cases through each stage | Where ideas stall | Whether production use lasts | Spectris, Johnson Controls; 42 of 55 postings | Keep with care |
| Coverage of trained champions | How much effort went in | Any effect on the work | HPE, Synchrony, Flutter (nearest measures) | Drop as a target |
Our reading, not a measurement. Time per task is placed as a timed sample, not as an estimate from action counts, which is how the Copilot Dashboard computes assisted hours. The two dots marked as warnings are the drops, and the two marked as good are the keeps. The order of the dots matters more than the exact spot.
The pattern is plain once drawn. The measures that sit in the bottom right are the ones with a dashboard. The two that sit at the top are the ones that need someone to time a sample and score it.
Which three to start with
Time per task, on one workflow
Pick a workflow that recurs and has a clear end. Time a sample of real cases before anyone changes how they work, then time the same kind of case after. Take the median of each.
Do this todayOur guide’s example times 40 cases by hand. Ten per workflow is a start.
Quality and rework, on the same cases
Score the same sample with the team’s own check, with someone other than the author scoring. Count what was redone. Faster only counts if the result is as good.
Do this todayReport time and quality side by side, never one without the other.
Depth of use: who is still working the new way at 30 days
This is the cheapest of the three, because most tools already log it. It tells you whether the other two numbers will last, and it needs a written definition of “active” by role.
Do this todayRead it next to the first two, by team, and treat a gap as a place to look.
Why these three. The first two are the only measures in the grid that show the work changed, and both come from a sample you take yourself. The third is the cheapest way to learn whether the change sticks. The first and third are the pair our earlier post on the job picked. Together the three cover speed, quality and whether the change sticks, which are three of the five measures in the guide. Cost per task and people’s time come later, and cost per task is built from the first two. Counts, team splits, pipeline and champion numbers stay on the page as context. Whoever owns this job can report all eight, and should lead with the three.
What this cannot say
- It cannot say which measure is most accurate. We found no public test that compares them, and the verdicts are our reading.
- It cannot say what employers measure in practice. Postings are wishes, the 62 are what one search found, and the people inside those companies may have targets we cannot see.
- It cannot say the three picks work for your workflow. Most of the studies behind the time and quality numbers are about software developers, short online tasks, university staff and civil servants. Finance, support and operations work is barely there.
- It cannot give you benchmarks. Several vendors publish averages, and we left them out because none gives a method or a sample.
- A measure with a target attached can change behavior. We read no study of that effect in AI enablement, so we say so and leave it there.
Where Nolli fits
Where Nolli fits, and where it does not
Nolli measures adoption and efficiency by team and workflow from the work itself: who uses which AI tools, skills and agents, and how the same workflow performs done by people, by people with AI and by agents, on time, quality and cost per task. In this post’s terms, that covers depth of use, adoption by team, time per task, quality and cost per task. It does not set your definition of “active”, run a champion network, or supply a comparison group such as a staggered rollout. It learns from work on company devices, the mail and calendar your company connects, and the logs of AI agents, and it does not see personal accounts on personal devices. Its before-and-after comparison works per workflow. Nolli has no published customer results, and this post contains none. No public price. Request a demo. See Nolli for AI enablement teams.
Request a demoQuestions people ask
What KPIs should an AI enablement team track?
Start with time per task against a baseline, quality and rework on the same sample, and the share of a team still using AI after 30 days. Add cost per task once the first two run. Treat logins, license counts and champion numbers as context.
Are logins and active users useful AI KPIs?
As context, yes; as a result, no. In one vendor case study of a software company of about 200 people, license data showed more than 95% had accessed AI tools while session data showed more than 30% of users barely used them.
How many AI enablement job postings name a target number?
Two of 62 read on October 16, 2026. One is a customer-facing role at Liminal that cites weekly active usage rising from 20% to 70% in 90 days. The other, at Capital on Tap, asks to triple the company’s ability to build products in 12 months. Thirty-nine of 55 internal postings name some way to measure.
What is the best single AI metric?
There is not a safe single one. The closest is cost per task done correctly, because it combines time, model use, review and rework, but it needs a baseline and a quality check first. Our pick for a pair is time per task next to quality on the same sample.
Should trained champions count as a KPI?
We would drop it as a target and keep it as a diagnostic. Champions appear in 33 of 55 internal postings as a duty, and we recorded no posting that scores their coverage. A trained person may use AI daily or never.
Sources
- “AI Enablement Leader (job posting).” Johnson Controls, accessed October 15, 2026.
- “AI & Automation Enablement Lead (job posting).” Flutter Group, accessed October 15, 2026.
- “AI Adoption & Enablement Lead (job posting).” Hewlett Packard Enterprise, via Workday, published August 11, 2026.
- “VP, Head of AI Enablement (job posting, Job ID 2602496).” Synchrony, posted September 25, 2026.
- “AI Enablement Lead (job posting, Particle Measuring Systems).” Spectris, via Workday, published August 24, 2026.
- “Head of AI Enablement Engineering (job posting).” Deepgram, via Ashby, accessed October 15, 2026.
- “Head of AI Enablement (job posting R-100190; expired).” J.D. Power / Autodata, via London Tech Jobs, posted June 30, 2026.
- “Applied AI Enablement Lead (job posting).” Liminal, via Greenhouse, accessed October 16, 2026.
- “Head of AI Enablement & Engineering (job posting).” Capital on Tap, via Greenhouse, accessed October 16, 2026.
- “AI Enablement Lead: Key Skills & Responsibilities in 2026.” Second Talent, August 14, 2026.
- “Connect to the Microsoft Copilot Dashboard for Microsoft 365 customers.” Microsoft Learn, accessed October 16, 2026.
- “AI Adoption Dashboard for Enterprise Teams.” Worklytics, accessed October 16, 2026.
- “AI Adoption Dashboard: Track AI Usage & Cost by Team & Role.” Larridin, accessed October 16, 2026.
- “AI Impact: Turn AI investments into measurable results.” ActivTrak, accessed October 16, 2026.
- “Measuring What Matters: The Missing Link in AI Adoption.” ActivTrak, April 29, 2026.
- “Measure AI Coding ROI and AI Impact Dashboard.” Jellyfish, accessed October 16, 2026.
- Uri Nativ. “AI dashboards.” Torii Support, June 11, 2026.
- Joel Becker, Nate Rush, Beth Barnes and David Rein. “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.” METR, July 10, 2025.
- Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan Jurafsky and Robert D. Hawkins. “The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks.” arXiv (2605.22687), May 2026.
- Fei Liu and Todd McKinnon. “Beyond the Hype: The Efficiency-Throughput Gap with GitHub Copilot.” Communications of the ACM, September 17, 2026.
- Manika Garg, Samuel Kernan Freire, Sunčica Bruck, Aslam Tanjung and Theo Bakker. “Generative AI in Higher Education Staff Work: Perceived versus Measured Effects (abstract page).” SSRN (7570368), October 6, 2026.
- Digital Data and Technology Monitoring and Evaluation team. “The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade.” UK Department for Business and Trade, August 2025.
- “The system for AI enablement teams.” Nolli, accessed October 16, 2026.
Plain-text copy: /blog/ai-enablement-kpis.md