This seems to be the central question for any company aggressively rolling out AI.
First everyone burns tokens: hooking up the most powerful models, giving employees access, letting agents run with almost no restrictions.
Then the bill arrives. Turns out AI gets used a lot, it costs a decent amount, and proving the payoff is surprisingly hard.
After that comes the usual sulking, the limits, and the ROI conversations.
What changed in Claude Enterprise
Anthropic just shipped an update to Claude Enterprise that actually tackles this problem.
There's now Smart Reports — reports that analyze employees' work sessions and show what tasks they're solving with Claude, what comes out the other end, what it costs, and where things go wrong.
On top of that, there are employee surveys about time saved and pooled budgets for teams. So Anthropic isn't just asking you to track AI consumption — it wants you to look at what work it's actually helping get done.
No revolution here. But I like the logic.
Token counts tell you nothing about efficiency until you understand what you got for the money.
How to measure the payoff
If you've got a project or a team with dozens of tasks, you can start with a simple spreadsheet — we covered a similar approach in our guide on what to delegate to AI and what not to. For each task, log:
- what you did without AI and how long it took
- what you do with AI now and how long it takes
- what the task costs once you factor in review and fixes
- how good the output actually is
- whether the process can be repeated without the "magic" author and their endless prompting
And this is where it gets interesting.
Smart Reports can spot recurring processes and recommend turning them into Skills — folders with instructions, scripts, and templates that help Claude handle specific tasks along a defined playbook.
You could turn report prep, brief analysis, code review, or deck assembly into a Skill — we've covered similar automation scenarios here. Then refine it, drop it in a repo, and have the whole team use it.
Anthropic has an open repo with example Skills and a published spec for building them. There's no universal tool that'll automatically calculate your company's AI ROI — but you don't need Claude Enterprise to build your own measurement system.
The logic runs roughly like this:
1. find tasks where AI actually helps
2. measure the output and the cost
3. identify repeatable processes
4. package them as Skills
5. check quality
6. scale
What to keep in mind
There's a catch, though. If an employee used to spend four hours on a deck and now spends one, the company isn't automatically three hours of salary richer.
So it's worth distinguishing time saved from actual money saved or extra revenue generated.
Still, even this fairly basic tracking will help you see where AI genuinely changes the economics of the work — and where it's just creating the feeling of insane productivity.