95% of organizations are realizing no returns despite spending $30 to $40 billion on AI. That's an MIT figure from 2025 and it has not improved. The story is almost always the same. The license was bought, the workshop was run, three people use Claude heavily, the rest of the team barely touches it, and six months later leadership quietly drops the line item.
The interesting thing is that the people running these rollouts are not lazy or unsophisticated. They are smart operators who got the obvious things right. They picked a good model, they ran a kickoff, they hired a trainer. The reason the rollout fades is not what they did. It is what nobody told them to do.
Three structural gaps cause almost every Claude rollout fade I've seen. Each one is invisible until the budget review hits and there's nothing to defend with. Each one has a fix that's cheap if you do it on Day 0 and expensive if you try to retrofit it.
Gap 1: The missing baseline
The first gap is the one nobody notices until it's too late: no measurable before-state.
Most teams describe their pre-rollout situation in adjectives. "Claude usage was low." "Adoption was spotty." "A few people were using it." Adjectives are not a baseline. A baseline is a number, captured before training starts, that the after-number can be compared against.
Without a baseline, every "after" number is unverifiable. You can say usage doubled, but doubled from what? You can say the team is more productive, but more than which day? When the CFO asks "is this worth $30,000?" the only honest answer without a baseline is "I think so." That answer doesn't survive a budget review.
The fix is mechanical. Three things, captured before Session 01:
- Platform baseline. Snapshot current Claude Enterprise usage before any training. Active users, sessions, project usage. The Claude Enterprise Analytics data window only goes back to January 1, 2026, and only runs 90 days — which means if you wait until after training to grab the baseline, the window may have moved past your start date. Capture it on Day 0 or you lose the option forever.
- Enablement baseline. Maturity-scale self-rating per person across prompting, Project setup, Skill use, and output review. Five questions. Five minutes per person.
- Workflow baseline. Each participant picks one target workflow and records the current time, frequency, and dollar value. Not abstract. The specific workflow they want Claude to help with.
This takes about 90 minutes of work across the team and one configuration step from the Primary Owner. The cost is real but trivial. The cost of not doing it is having no defensible answer to "what changed."
Gap 2: The missing build
The second gap is the one that surprises even teams who took training seriously. They ran a great workshop. The team left enthusiastic. Six weeks later, the workshop felt good but nothing changed about how work actually happens.
The reason is that nobody built anything during the engagement. Workshops teach prompting. They don't install workflows. The decks fade in two weeks. The new habits never form because there was no Claude-native artifact for the team to use on Monday morning.
This is the part where most rollouts plant trees but never plant fruit. The training is the tree. The Skills, the Projects loaded with real company context, the validators that catch low-quality output before it reaches a decision-maker — those are the fruit. Without them, you have a team that knows how to use Claude in theory and reverts to working the old way in practice.
The fix is to make builds part of the engagement, not an aspiration for the team to pick up after. Pick three workflows that are real, painful, and high-frequency. Build each one as a Skill — not a chat template, a deployed Skill in the client's AI Portal that the rest of the team can invoke. Then build a validator Skill that runs over output and flags anything unsourced or low-confidence.
Three builds is the minimum that creates institutional muscle memory. One is a demo. Two is a coincidence. Three is a pattern that the team starts to repeat on their own. The math on this is consistent: role-specific training built around real workflows hits 67% sustained adoption. Generic training without the build piece hits 23%. The build is the difference.
This is also where the Skill-counting trick becomes load-bearing. Skill-invocation count equals workflow-execution count. "The IC memo generator ran 47 times this month" is a real number off the Claude Enterprise Skill-usage endpoint. Without Skills, adoption stays qualitative forever.
Gap 3: The missing measurement loop
The third gap is the one that decides whether the rollout survives its first leadership review. There is no measurement loop in place.
By "loop" I mean the closed system that turns the baseline plus the builds into a defensible dashboard. Adoption from platform telemetry. Enablement from a re-measure form sent to the same people who filled out the baseline. ROI from self-reported hours saved per named workflow, dollarized at a loaded rate. The validator's quality-flag rate as the one ROI signal that's machine-measured.
Most engagements skip this loop because the work to set it up looks small from the outside and turns out to be load-bearing from the inside. Access patterns have to be named in the SOW. The Primary Owner has to issue an Analytics API key with read:analytics scope, or commit to CSV exports on a cadence. Participant emails have to be collected before Day 0 so the post-engagement form can be matched to the baseline form. Build seats have to be provisioned so the consultant can deploy Skills into the actual workspace.
If any of these three access patterns is missing, the dashboard has gaps. The gaps become "we don't really know" lines in the readout. The readout doesn't survive leadership.
The measurement loop is not optional. It is the difference between an engagement that produces a board-ready dashboard and an engagement that produces a feel-good narrative.
What the three gaps look like together
Here's the failure pattern when all three gaps are present, which is the modal case:
The team runs a half-day workshop. Everyone leaves happy. Two power users start running real workflows in Claude. Eight people use it intermittently for a few weeks. By Week 6, the power users are doing 95% of the team's Claude work. By Week 10, leadership asks for an update and the rollout owner pulls together a deck with screenshots of cool things power users built. No numbers. By Week 14, the rollout is "ongoing" in a way that means "nothing structured is happening anymore."
Here's what the same engagement looks like with the three gaps closed:
Baseline captured on Day 0 — platform, enablement, workflow. Three named Skills built and deployed during the engagement. Validator running over high-stakes output. Day 60 re-measure pulls platform telemetry off the API, sends the re-measure form to the same people who filled out the baseline, dollarizes the hours saved at the loaded rate the client provided. The output is one document a board could read in five minutes. The rollout converts from "ongoing" to "month two of the partnership."
The category this belongs in
This is why most AI training fails not because the trainer was bad, but because the structure around the training was missing. Workshops and decks alone are an incomplete product. The full product is the workshop plus the builds plus the measurement loop, with the baseline captured before the first session and the dashboard built into the engagement from Day 0.
Section sells you a cohort. McKinsey sells you a deck. The category this belongs in is the operating-layer install — the rollout that ships training, builds, and a measurable adoption story together. That's The 60-Day Claude Rollout™, and it exists because the gaps above kept showing up in every engagement that started without them named.
If you're already mid-rollout
If you're three or six months into a rollout that's fading, the three gaps above are diagnostic. Walk through them in order. If the baseline doesn't exist, the salvage is to anchor on the earliest available platform data and a retroactive maturity self-rating — better than nothing. If the build is missing, pick the three highest-frequency workflows and build them as Skills now; you can still create the muscle memory before the engagement ends. If the measurement loop is missing, the access patterns can usually be retrofitted in two weeks.
The cheaper version of all of this is to start with the gaps closed. If you're rolling Claude out and want all three of these built in: The 60-Day Claude Rollout™. From $22,500 for a 5-to-15-person team.
Related reading: How to Measure AI Adoption Without Making It Up walks through the three-axis instrument in detail. From Claude Enterprise License to Measurable ROI in 60 Days is the playbook view of the same engagement.
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.