74% of companies see no tangible value from their AI investments. The most common reason is not that AI doesn't work. It's that nobody set up a way to prove it does.
The pattern is almost identical every time. A team buys Claude Teams or Enterprise. They run a workshop. Three months later, leadership asks "what are we getting for this?" and the person who championed the purchase reaches for whatever number they can find. Seats provisioned. Attendance at the kickoff. A vague "people seem to like it." None of that is measurement. That's narration.
There is a defensible way to do this. It uses three axes, two instruments, and one sequencing constraint that breaks most rollouts that try to bolt measurement on at the end. Here's how it works.
The three axes
The first mistake teams make is collapsing every question about an AI rollout into one bucket they call "adoption." Adoption is one axis. There are three.
- Adoption measures behavior. Are people using the system, how often, on what.
- Enablement measures competence. Is the team able to use Claude well, and is that ability spreading or staying with one or two power users.
- ROI measures value. Dollars and hours recovered, plus the rate at which the team catches AI output that isn't good enough to ship.
Each axis answers a different question. A team can have high adoption (everyone logs in daily) and zero competence (they all use Claude like a fancier Google Search). A team can have moderate adoption and real competence (a handful of people running serious workflows) but no ROI math to defend the budget. You need all three because no single number tells the whole story.
This is the difference between a board readout and a hand-wave. "47 people used Claude this month" doesn't justify spend. "47 people used Claude this month, the IC memo Skill ran 132 times, average maturity rating moved from 2.1 to 3.4 across the team, and self-reported hours saved per analyst per week is 6.8 at a loaded rate of $185/hr" — that justifies spend.
Two instruments. Neither one alone is enough.
Each axis pulls from a different source. There are exactly two instruments. They do different jobs and they do not substitute for each other.
Platform telemetry sees behavior. It cannot see hours saved, dollar value, or competence — by design, it is content-free. The Claude Enterprise Analytics dashboard tells you who used what and how often. It is objective, it is automated, and it does not require anyone to fill out a form. That makes it the spine of the Adoption axis.
Forms see value and competence. The pre/post maturity-scale delta per person, the before/after of one target workflow per participant (time, frequency, what it's worth), the self-reported hours recovered — those numbers live in forms. They are self-reported, which is why they need a real baseline to be credible, but they are also the only way to surface dollar value and skill growth.
The "numbers, not anecdotes" claim is the two instruments together. Telemetry without forms is behavior with no value attached. Forms without telemetry is value with no behavioral evidence underneath. Pair them or the dashboard falls apart under scrutiny.
The Skill-counting trick that makes adoption countable
Most adoption dashboards stay vague because the team never built the thing that would make them concrete. There is one design decision that fixes this and it has to happen at scoping, not at the end.
Build each named workflow as a Skill.
Skill-invocation count = workflow-execution count. "The IC memo generator ran 47 times this month" is a real number off the Claude Enterprise Skill-usage endpoint. Without this, the team has to interview people and ask "how often did you use Claude to draft a memo this month?" and the answers are guesses.
This is the single highest-leverage build decision in any rollout. Three or four named workflows turned into Skills means three or four lines on the dashboard that the platform populates for free. Without it, adoption is qualitative. With it, the system measures itself.
The 90-day window that breaks late starts
Here is the constraint that catches most teams off guard, and it is the single most important sequencing rule for anyone running this on Claude Enterprise.
The Claude Enterprise Analytics data window only goes back to January 1, 2026, and only runs 90 days. That means the baseline has to be captured before training starts. Not after. Not "we'll grab it once everyone is licensed." Before Session 01, or the floor doesn't exist and the delta is unprovable.
This single sequencing rule kills more rollout measurement than any other failure mode. Teams launch, run training, build workflows, then go back six weeks later to pull baseline numbers — and find the window has moved past the start date. Now there is no clean before. There is only a current state and a hope.
If you're already mid-rollout and didn't capture the baseline, the salvage move is to anchor on the earliest available platform data and pair it aggressively with self-reported maturity baselines from the team. Better than nothing. Worse than doing it right on Day 0.
The three access patterns most proposals get wrong
Measurement breaks at the access layer more often than at the analytical layer. There are three different kinds of access required to run a real three-axis instrument, and they don't substitute for each other.
- Telemetry access is issued by the client's Primary Owner. Either an Analytics API key with
read:analyticsscope, or the Owner runs the native CSV export on a cadence and hands it over. One-time Owner action. No Claude seat required for the consultant. - Build seat is the only access pattern that requires an actual Claude login. Building Projects and Skills against real client material and deploying them to the team's AI Portal is dramatically faster with a guest or member seat in the client's workspace. Without one, the fallback is co-building live while a power user drives the screen — it works, but factor the slower pace into scope.
- Participant emails are needed to match each person's enablement baseline form to their re-measure form. Email is the clean key. Anonymized codes work if the client is privacy-strict, but they add friction every time someone changes teams or leaves.
Name all three in the SOW. The Primary Owner action items have to be listed explicitly, with dates, so they happen before Day 1. Otherwise the engagement starts and the baseline doesn't, and the consultant is two weeks in before anyone realizes the dashboard has no data to populate.
What a real dashboard looks like at end of engagement
The output of a properly instrumented rollout is one document a board or LP could read in five minutes. Three sections, each with the delta.
Adoption (platform). Active users, week over week. Projects created and used. Skill invocations per named workflow. The shape of the curve from Day 0 baseline through end of engagement.
For a recent 25-person capital management firm engagement in Q2 2026, the adoption section was: 4 power users at baseline, 19 active weekly users at Day 60, three named Skills (IC memo, board pack, research validator) with 47, 21, and 96 invocations respectively in the final 30 days. No interviews required to produce any of those numbers. The platform did it.
Enablement (forms). Maturity-scale movement per person — usually a 1-to-5 scale across prompting, Project setup, Skill use, integration awareness, output review. Build-vs-observe ratio during training sessions. Average movement across the team, plus the distribution (because three people moving from 1 to 4 plus 15 people staying at 2 is a very different story from everyone moving from 2 to 3).
ROI (forms + the validator). Hours recovered per workflow, dollarized at a loaded rate the client provides. Plus the validator's quality-flag rate — the rate at which validation Skills catch unsourced or low-confidence AI output before it reaches a decision-maker. That last one is the only ROI signal the system can measure directly, and it is the one that lands hardest with leadership: not "we use Claude more," but "we caught 11 instances of AI output that would have shipped to the IC without a sourced citation."
Scaling the framework down
The three-axis instrument is built for engagements that justify it — series workshops, partnerships, and the 60-Day Rollout. For a Quick Start or a single workshop, the same logic scales down to the lighter version.
The light pattern is forms only. Intake baseline plus one 30-day re-measure. No platform telemetry, because a single workshop doesn't justify the Owner-side access setup. What carries the light version is the maturity delta plus one yes/no outcome per participant: "did you build the workflow we trained on, yes or no." That's enough for a 90-minute workshop. It is not enough for a $30K engagement.
The heavier pattern — full three-axis, platform telemetry plus forms plus mid-engagement pulse — is what produces a board-ready dashboard. That's the pattern in The 60-Day Claude Rollout™.
The work this replaces
Most "AI adoption measurement" before this framework collapsed adoption, enablement, and ROI into one undifferentiated bucket. Teams reported "adoption" as a single percentage. The percentage was usually a guess. The framework above retires that approach because each axis is provable on its own terms — and because the platform-side and form-side instruments are doing the jobs they're actually built to do, instead of pretending one can do both.
If you're rolling Claude out and want this installed end-to-end — three workflow builds, a board-ready dashboard, the access patterns named in the SOW, the baseline captured on Day 0 — that's exactly what The 60-Day Claude Rollout™ productizes. From $22,500 for a 5-to-15-person team.
If you want to read the related posts in this series: The Three Things Every AI Rollout Misses walks through the structural gaps that turn a working Claude license into a 90-day fade. Why Most AI Training Fails covers the 23% vs 67% adoption gap between generic and role-specific training.
Nicole Patten is the founder of Elevate Online and runs a Claude-specific training practice. She spent 7 years at Google as a Senior UX Engineer before dedicating her career to helping teams use AI responsibly and effectively. 100% of her business runs on Claude.