What to Measure in the First 90 Days of a Claude Rollout

Key Takeaways

• Seats activated is the wrong headline. It tells you what you bought and who got an invite. It says nothing about whether anyone's work changed.
• Weekly active users is the first honest number. Gallup's latest data shows a wide gap between people who touch AI now and then and people who use it every week. Your rollout will have the same gap, team by team.
• Workflows connected show whether Claude reaches real work. A connector that is switched on and never used is a line in a config file. Count the workflows people actually run.
• Hours returned needs sampling, not a survey. People are bad at estimating how much time AI saves them, so time a few real tasks.
• Cost per active user is the number your CFO will ask for. Claude Enterprise bills seats and usage separately, so cost per seat no longer tells you what Claude costs.

Here's the number most AI rollout updates lead with: seats activated. It's easy to pull. It climbs fast. And it hits 100% the week IT finishes provisioning, before most people have typed a single prompt.

That's the problem. Seats activated measures what you bought. Whether anyone works differently is a separate question, and it's the one that matters.

Gallup's tracking of AI use at work shows the gap at national scale. As of May 2026, 52% of U.S. employees say they use AI in their role a few times a year or more. Only 30% use it a few times a week, and 15% use it daily. (Gallup is measuring all AI tools across the whole workforce, not one company's Claude rollout. Read it as the shape of the problem rather than a benchmark.) Having access and having a habit are two very different things, and the space between them is where most rollouts quietly stall.

Here's how I break it down. A Claude rollout has to clear four bars, in order. People show up and come back. Claude touches real work. That work gets faster. And the cost makes sense for what you get. Each bar has one number that tells you whether you cleared it: weekly active users, workflows connected, hours returned, and cost per active user.

Seats activated doesn't clear any of them.

The first 90 days matter because that's the window where a new tool turns into a habit or a forgotten tab. If you measure the right things in that window, you can still change the outcome. If you wait for the annual renewal to find out, you can't.

Seats Activated Measures Procurement

Seat activation is an IT metric. It goes green when provisioning is done. With SSO and SCIM in place, everyone lands in Claude on day one, so the number is at or near 100% before the first training session starts.

That makes it a fine checklist item and a terrible headline. A number that can only go up, and reaches its max on day one, can't tell you anything is wrong.

Anthropic's own usage analytics for Team and Enterprise plans shows active members next to assigned seats. That ratio is the number to watch.

Make weekly active users the headline

Define "active" before launch: a person who did something with Claude in the last seven days. A chat, a project, a Claude Code session, a Cowork session. Any of them counts.

I prefer weekly over daily because plenty of roles don't need Claude every day. A deal team member who uses it on Tuesday and Thursday to draft memos is a real user. Monthly is too forgiving: someone who tried it once in week three still counts.

Then report it by team. A company-wide 70% can hide a finance team at 20%, and the finance team is the one you need to hear about.

Showing Up Once Is Easy. Coming Back Is the Signal

Every rollout gets a week-one spike. There's a launch email, a training session, some curiosity. The real reading comes in weeks four through eight, when the novelty is gone and only the habits are left.

BCG's AI at Work 2025 survey of more than 10,600 workers found 72% of respondents use AI regularly, but frontline regular use has stalled at 51%. The same research found that people who get five or more hours of training, especially in person with coaching, are far more likely to become regular users. So the training you put in shows up in the retention curve. You just have to be watching it.

Watch which teams stall after week two

Plot weekly active users by team, every week. A team that starts low and stays flat usually needs a workflow. Another reminder email won't help. Nobody builds a habit around a tool they don't have a reason to open.

Look at depth, too. Projects created, skills used, Claude Code sessions per person. A team where everyone chats and nobody has set up a project is using Claude as a search box. That's fine for week two. It's a warning sign at week eight.

Workflows Connected Show Whether Claude Reaches Real Work

Claude without context answers generic questions. Claude connected to your documents, your CRM and your data does actual work. That's why connectors matter so much in a rollout.

But "connectors live" is just another activation number. It tells you IT switched something on. The analytics dashboard shows each connector with the number of users and the count of read and write actions. That's closer to the truth.

Count workflows people run, not connectors you turned on

Define a workflow as a named, repeated task with an owner. "First-pass CIM summary." "Weekly pipeline recap." "Month-end variance notes." Then track two things: how many named workflows ran this week, and how many people ran each one.

My bar for day 90 is simple. Every team in the rollout should have at least one workflow that runs every week without anyone reminding them. If a team doesn't, that's the next thing to build.

Hours Returned Is the Metric Everyone Wants and Nobody Measures Well

This is the one the CFO actually cares about. It's also the hardest to measure honestly.

The easy way is a survey: "How much time does Claude save you each week?" The problem is that people are bad at answering that. METR's study of experienced open-source developers is the clearest example I know. Sixteen developers worked 246 real tasks. With AI tools allowed, they took 19% longer. Afterward, they still believed AI had sped them up by 20%.

I want to be fair to that study. It used early-2025 tools, METR itself says the results are out of date, and the setting was narrow: experts working in code they knew very well. I wouldn't use it to argue AI slows people down. I'd use it to argue that how fast people feel is a different number from how fast they are.

Claude's dashboard includes an estimated time saved figure. It's a useful direction-finder. It's still an estimate.

Sample the time on a few tasks

Pick three to five workflows. Time a handful of runs before Claude and a handful after. Multiply the difference by how often each workflow runs, which you already have from your usage data.

You'll get a smaller number than the survey gives you. You'll also get one you can defend in a board meeting. (Turning that into a full ROI case is its own exercise. For the first 90 days, a defensible hours number is enough.)

Cost Per Active User Replaces Cost Per Seat

Claude Enterprise pricing changed how this math works. Anthropic's Enterprise billing page says the seat fee covers platform access and usage is charged separately, at standard API rates. On sales-assisted plans, usage is billed monthly in arrears based on what your organization actually consumed.

So "what does Claude cost per person?" no longer has a fixed answer. It's the seat fee plus usage, and it moves every month.

The fix is to change the denominator. Divide total Claude spend (seats plus usage) by the people who were actually active, not by the seats you bought.

Here's an illustrative example. Say a company has 200 seats and 120 weekly active users. Dividing total spend by 120 gives a cost per head about two thirds higher than dividing by 200. The first number is the real one. The second is the one that ends up in the budget deck.

Read the two numbers together

Cost per active user means very little on its own. Read it next to weekly active users.

• Both rising. Your active people are going deeper, running agents and longer workflows. Probably good. Check that the workflows driving it are the ones you'd fund on purpose.
• Cost rising, users flat. A few people, or one runaway workflow, are driving the bill. Look at spend concentration before you do anything else.
• Cost falling, users rising. The rollout is working. More people are sharing the fixed seat cost, and nobody is burning tokens without a reason.

A 90-Day Measurement Plan

Here's how I'd sequence it. One question per phase, one headline number, and one decision the number should drive. If a metric doesn't change a decision, drop it from the report.

Set the baselines before anyone logs in

Agree on the definitions with finance before launch. What counts as active. Which workflows you'll track. How you'll time them. Then time those workflows the old way, before anyone has Claude.

This takes an afternoon up front. Reconstructing it at day 90 is close to impossible, because nobody remembers how long the memo took in July.

What I'd Want on the Day-90 Slide

Four numbers, by team, with the trend line: weekly active users against seats, workflows running every week, sampled hours returned, and cost per active user. Next to each team, one sentence on what you're going to do about it.

That's it. No seat activation. No "prompts sent."

I'll be honest about what's still unsolved. I don't think anyone has a good benchmark yet for what weekly active use should look like by function. Finance teams and engineering teams adopt at very different speeds, and the analytics Anthropic ships keep getting better every quarter, so the right metrics a year from now may look different. Standing Claude up is the straightforward piece. Getting each team to look at its own numbers and change how it works is the real project.

What I'm sure of is that the companies that agree on what to measure before launch will know at day 90 whether the rollout worked. Everyone else finds out at renewal.

If you're planning a rollout, start with the definitions. That's how we run Claude deployment and Token & Usage Analytics for the teams we work with, and if you want to compare notes on what you're seeing, Book a Discovery Call.

FAQ