Claude Enterprise administrators can now pick a team and a date range and get back a report on the work, not only the bill. Claude reads a sample of the team's sessions, sorts them into kinds of work, and prices each kind. The report lists the output types, such as analysis, documentation, drafts and code, with the average cost per session for each and where the most expensive sessions cluster. It is off by default, and each organization gets ten reports a month at no charge. Pooled budgets arrive with it: a group shares one monthly allowance, so a heavy user is not stopped by a personal cap while a colleague leaves money unused.
A total tells you who spent, not what it bought
Admins could already see spend by person and by group, and set caps. That is the electricity bill for a whole building: it shows consumption, and it says nothing about which machine was left running. We wrote earlier that AI bills grow with use the way electricity does, and the next step is to be able to read the meter by machine.
For example, a finance director at a hospital group has 40 people on Claude and a monthly bill of $9,000. By person, the bill only shows who uses it most. By kind of work, it might show supplier-invoice matching, drafted emails and month-end analysis, and that analysis sessions cost ten times as much as the rest because many of them stall on a broken link to the ledger export. One afternoon of fixing that link would save more than any cap on any person.
More spend does not buy proportionally more
The numbers from software teams show why this matters. A developer-week with ten times the median number of tokens produces only about 2.2 times as many merged code changes. Over the past year, the all-in cost of a merged change, salary included, fell by a third at the median company, from $1,202 to $820, and AI's direct share of the $799 average was about $18.
Those two facts fit together. AI is cheap next to the salaries it works beside, so shaving the bill is rarely the point, because $18 is about 2% of the cost. The waste is inside that 2%: long sessions that loop, repeat or fail. A manager who can see them can pay happily for the sessions that save an hour and stop paying for the ones that save nothing. And since the report also covers documents and analysis, the same question reaches teams that never write code.
The cost lands on the team lead
The cloud lesson is that a bill gets managed by the person who can change the work. That is why owners can let team leads who are not admins run reports for their own groups, and why an admin can raise someone's spending limit from inside the report. The AI cost conversation moves from IT and finance to whoever runs the team. Within a year or two, I expect finance departments to ask each team for AI cost per finished piece of work, the way they now ask for cloud cost per product.
A team lead needs both halves of that answer. The report gives the cost. The in-app surveys give the value, because employees are asked which use helped most and how much time it saved. Neither measures the finished output, so the judgment stays with the manager. One thing to do before switching reports on is to tell the team that their sessions will be read. Anthropic's own guidance is that the reports should not be used to judge individual performance or make employment decisions, and a team that hears that from its manager first will trust the numbers more.
A team that knows what a finished analysis costs can decide where more AI is worth paying for. A team that knows only the total has one choice, which is to spend less.