Every quarter, enterprises pour billions into large language models, and every quarter the same question goes unanswered: what did the money actually buy? Anthropic decided to answer it. On Wednesday the company announced smart reports, in-app surveys, and pooled group budgets – a new suite of tools designed to show companies what their AI spending is buying, team by team and workflow by workflow. Paired with fresh financials showing 900% net revenue retention, the message to finance chiefs is blunt: stop flying blind on AI spending.
This is the first time a frontier lab has shipped value-visualization tooling this deep inside its own enterprise product. The reports lean on internal analytics and employee surveys to reconstruct what a team actually produced with Claude, how much it cost, and where the workflows broke down. Anthropic says the data is not used for model training by default. For customers who have never seen a line-item explanation of their AI spending, that is a genuine shift.
In this article we break down what each tool does, the market data that made this launch urgent, what it means for vendor lock-in, and a practical playbook for getting your AI spending under measurement before your next renewal conversation.
Table of contents
- What Anthropic’s AI Spending Reports Actually Show
- Smart Reports: Seeing the Work Behind the Bill
- Surveys and Pooled Budgets: Rounding Out the Picture
- Why Tracking AI Spending Matters Now
- The Fortune 500 Problem: Spending Without Proof
- Lock-In, Switching Costs, and FDE-Lite
- What CFOs Should Do About AI Spending
- Where AI Spending Goes Next
- References
What Anthropic's AI Spending Reports Actually Show
A Quarterly Profit Changes the Conversation
The launch arrived alongside a striking set of numbers. Anthropic shared financials with investors showing the company recorded an operating profit in the second quarter, and 900% net revenue retention – meaning spending from the same customers in the third quarter of 2026 grew nine times over the same customers a year earlier.
Those figures explain the timing of the launch. A vendor growing this fast wants its customers to see why the invoices keep climbing, because customers who can see value keep paying. Making AI spending legible is now part of Anthropic’s own retention strategy, not just a chore for procurement teams.
Smart Reports, In-App Surveys, Pooled Budgets
The release has three parts. Smart reports surface what a team gets done with Claude and what that work costs. In-app surveys capture employees’ own feedback about how Claude is changing their work. Pooled group budgets let admins set one shared budget per team instead of a pile of individual caps. All three are generally available for Claude Enterprise today.
Your Data Stays Your Data
Anthropic states it will not use this telemetry for model training by default, and in-app survey responses are treated as customer data, accessed only as necessary to operate the service. For regulated organizations, availability varies by configuration, including HIPAA setups and customer-managed encryption keys – a reminder that AI spending decisions increasingly travel with data-governance reviews.
Smart Reports: Seeing the Work Behind the Bill
What a Smart Report Shows
Pick a team, a date range, and the products to include. Claude then analyzes a sample of the team’s transcripts and groups the work into workstreams. For each one, the report shows the spend, the deliverables produced, and the most common points of friction – a missing connector here, a failed tool call there.
The framing is deliberately work-first. Instead of a bill broken down by model and month, you get a picture of what the team is actually attempting with Claude: research sweeps, code migrations, document drafting, incident triage. Anthropic calls the ambition level explicit in its product copy – the reports are meant to surface the ambitious work your teams are doing and show how to turn that work into repeatable workflows the rest of the organization can run.
That last detail matters more than it sounds. Most failed AI sessions never surface; the employee just does the work manually and nobody counts the loss. Surfacing friction turns invisible failure into something a platform owner can fix, which is the cheapest way to improve the return on AI spending without buying anything new.
Recommendations You Can Act On
Each report ends with one consolidated list of recommendations, including recurring workflows that should become skills and rules worth adding to a code repository. Any finding can be shared as an artifact with the people who can act on it, and admins can ask custom questions about model usage and cost, then save those questions and insights as templates.
If a spend limit is holding someone back, admins with billing permission can raise it directly from the report. Owners can also let team leads who are not admins run reports for their own groups. Anthropic is explicit that the reports are not meant for evaluating individual performance or making employment decisions – a line worth remembering when someone asks for a productivity leaderboard of your AI spending.
The Reporting Stack at a Glance
Here is how the three new capabilities compare, so you can tell which admin owns which lever.
| Tool | What it answers | Who runs it | Guardrails |
|---|---|---|---|
| Smart reports | Which workstreams exist, what they cost, where they break | Admins plus delegated team leads | Not for individual performance reviews |
| In-app surveys | How employees value Claude and what would break first | Admins; questions, timing, placement customizable | Treated as customer data; deletable on retention schedule |
| Pooled group budgets | Whether capacity matches where the work actually is | Admins with billing permission | Per-member limits still apply as a ceiling |
Surveys and Pooled Budgets: Rounding Out the Picture
The Questions Anthropic Asks Employees
The survey template is more candid than most internal pulse checks. Default questions focus on value: the most valuable thing someone has done with Claude recently, how Claude changed the time a task took, which use of Claude a teammate should copy, and – the telling one – if the team lost access tomorrow, what would break or slow down first.
Answers stream onto a survey page as they arrive and export to a spreadsheet. Members see the survey as a short card after a reply in the tool they are already working in, which should lift response rates above the usual email graveyard. For an AI spending review, these answers are the qualitative half of the evidence: the reports show the work, the surveys show whether people would miss it.
Pooled Budgets Move Money to the Work
A pooled group budget gives a team one shared monthly allowance. Heavy users no longer hit an individual cap early in the month while lighter users leave budget unspent. Member limits still exist as a guardrail against unusually high individual use; every request counts against both the member limit and the pool, and a member stops at whichever runs out first.
When someone does hit a limit, they request more from inside Claude and the request lands with admins. The setup suggests an amount based on the group’s recent usage, and billing admins get email alerts at 50%, 75%, 95%, and 100% of the budget – the first honest budget alarm many of these organizations will ever have received for AI spending.
Why Budget Design Matters Here
Per-seat caps were inherited from the SaaS era, where usage is roughly uniform. Agentic coding is the opposite: usage concentrates in a handful of power users and busy sprints. Pools that follow the work keep your AI spending flowing to where it produces code, instead of rationing it evenly and capping your best engineers by accident.
Why Tracking AI Spending Matters Now
Jellyfish Data: 10x Growth, 2% of Engineering Cost
The market context is why this launch is not a footnote. Research firm Jellyfish, which measures software development organizations, released a report the day before Anthropic’s announcement covering its customers’ use of agentic coding tools across 276,000 developers. Claude Code and other tools built on Anthropic models are the most widely used, responsible for nearly a third of all proposed code changes submitted for review.
More important for the budget line: spending on those tools has grown 10x since the beginning of the year, and the cost now equals about 2% of total engineering cost. “Even if you’re saying 2% is still small, that 10x growth is growth that most CFOs didn’t plan for,” said Nicholas Arcolano, chief scientist at Jellyfish. “We’re certainly getting to the point where people are realizing, okay, it’s not enough to just spend a bunch. You now have to optimize it, make it work.”
Three Numbers Behind the Launch
Take a moment with the magnitudes below – they explain why AI spending tooling went from optional to shipped-in-the-same-quarter as a record profit.
Codex, Astra, and Quiet Switching Costs
Anthropic leads among enterprise customers, and the competitive threat has not dented that. OpenAI’s Astra model, released this summer, looked like a genuine challenge to Claude Code’s position, yet Anthropic’s adoption numbers kept climbing. “It took a long time for us to see Codex really pop in our data at all, and it has come up with Astra,” said Nikolas Albarran, AI product lead at Jellyfish. Looking at X, he noted, you would think otherwise: “Everyone in the Valley is like, ‘Oh, Codex is better.'”
The Fortune 500 Problem: Spending Without Proof
Who Drives Token Spend – and Who Does Not
Businesses have been hunting for this kind of visibility for years, and the results are lopsided. Usage among venture-backed firms is driving token spend: startups wire AI into engineering from day one, and their usage curves look like Anthropic’s retention charts. Most of the Fortune 500, meanwhile, are not seeing major gains from deploying AI at all.
That gap is the real target of the new tools. A venture-backed company already measures everything in dashboards; a legacy enterprise often cannot say which department spent what, on which models, producing which artifacts. Visibility tools from the vendor arrive precisely where the measurement muscle is weakest – which is also where AI spending churns fastest after a failed pilot. The stakes are real: at frontier firms, AI spending is now measured in hundreds of millions of dollars per quarter.
The volatility argument cuts the other way too. Coverage of AI spend per employee at top firms showed it slumping during August 2026, prompting the question of whether the drop was summer doldrums or a warning sign. You cannot answer that question at all without continuous, per-team measurement of AI spending – one more reason the reporting layer matters far more than any single quarter’s numbers.
ROI Becomes the Procurement Gate
The pattern repeats across large organizations: AI budgets approved in an optimistic wave, then audited in a skeptical one. When the audit comes, teams without measurement lose their budget. We covered the mechanics of that audit in Five Keys to Controlling AI Token Costs, and the discipline applies whether or not your vendor ships reporting tooling – see also our cost optimization services practice, where the first deliverable is almost always a usage map, not a discount.
Lock-In, Switching Costs, and FDE-Lite
Forward-Deployed Engineering, Light
There is a strategic read of the launch, too. All the major AI labs are investing in forward-deployed engineering (FDE) organizations – teams that embed AI consultants inside customer organizations to get them productive. Anthropic’s new reports can be seen as FDE-lite: a lighter version of that hands-on help, delivered through the product instead of a services invoice.
The recommendation engine is the tell. When the vendor’s report tells you which workflows to turn into skills and which rules to add to your repository, your internal playbooks get written in the vendor’s grammar. That makes the product better for you – and makes AI spending reviews harder to imagine without that vendor’s tooling.
The agent.md Quirk Problem
Jellyfish’s Albarran expects the moat to widen on its own. “We’re at a point now where it’s mature enough, where once you’re in, it’s going to be harder and harder to switch because you’ve spent the effort,” he said. Even if the frontier models end up mostly equivalent, each has “its own slight quirks of how they handle agent.md files” and workflow conventions, “and so that cost to switch is not zero.”
That is the honest summary of enterprise AI right now: great enterprise motion plus incumbent weight. “This stuff just worked and worked well, and a lot of people just put a lot of effort into standardizing workflows around it,” Albarran said of Anthropic. “And there’s not a big enough reason to switch.”
What This Means for Your Leverage
Anthropic’s public ambitions reinforce the point: its S-1 filing laid bare how much of the story rides on keeping enterprise customers hooked, and the rival Microsoft-OpenAI revenue machine is running quarters in the tens of billions. Expect every vendor to start selling measurement alongside intelligence. The negotiation lesson is simple: if the vendor’s reports become your only source of truth about your own AI spending, you are renting your insight. Build an independent usage map as well.
What CFOs Should Do About AI Spending
Baseline, Pool, and Map the Workstreams
Start with a 90-day baseline: total AI spending across vendors, by team, by product, by model tier. Turn on the vendor reports where they exist – they cost nothing – but reconcile them against invoices, because report totals and billing totals rarely agree. Then adopt pooled budgets deliberately: pool by team only where work genuinely flows between teammates, keep member limits on, and wire the 75% alert to whoever owns the budget.
Measure Outputs, Not Logins
The failure mode of AI measurement is vanity metrics: seats, logins, hours. Anthropic’s workstream framing is a better template – what was produced, what it cost, where the workflow broke. Track deliverables per dollar, friction counts, and rework rates. And hold the line Anthropic itself drew: these reports are not for evaluating individual performance, and using AI spending data for employee surveillance will poison the very surveys that make the numbers meaningful.
Negotiate With Your Own Data
Finally, use the visibility while you still have negotiating room. A workstream-level picture of AI spending lets you ask precise questions: why does team A pay 3x per merged pull request versus team B, which model tier actually serves the long tail, and what would committed volume be worth. The table below contrasts the old posture with the measured one.
| Decision | Unmeasured posture | Measured posture |
|---|---|---|
| Budgeting | Equal per-seat caps set by gut feel | Pools sized on recent usage, alerts at 50-75-95% |
| Success metric | Seats activated and logins | Deliverables and cost per workstream |
| Failure signal | Invisible – users quietly revert to manual work | Friction list: missing connectors, failed tool calls |
| Renewal | Annual conversation on vendor-supplied numbers | Quarterly review on your own usage map |
Where AI Spending Goes Next
The Reporting Wars Will Spread
Anthropic turned usage transparency into a retention feature; competitors will feel the pressure to match it, and enterprise procurement teams should encourage the trend loudly. Expect reporting suites, survey telemetry, and budget pools to become standard equipment across AI vendors within a year – the same way observability became standard in cloud infrastructure once everyone realized nobody would trust an unmeasured bill.
The Optimized Era of AI Spending
The deeper shift is cultural. The 10x growth Jellyfish measured happened under old-era financial controls: credit cards, per-seat caps, and a shrug. Once AI spending has workstreams, deliverables, and friction reports attached, it becomes a normal line item with normal management – which is exactly what most CFOs wanted in the first place.
The tools to find out what your AI spending is buying now exist. The only remaining question is whether your organization reads the reports, or keeps paying without opening the envelope Baseline first, pool second, and let the friction reports decide which workstreams deserve the next quarter’s budget.
References
Anthropic just launched tools to show companies what their AI spending is buying (TechCrunch)
Smart reports, in-app surveys, and pooled group budgets now generally available (Anthropic)
AI Engineering Trends report (Jellyfish)
Anthropic’s IPO filing reveals its strategy for keeping enterprise customers hooked (TechCrunch)
AI spend per employee slumped at top firms in August (TechCrunch)
Get started with smart reports (Claude Help Center)
Create surveys for your organization (Claude Help Center)
Manage pooled group budgets on enterprise plans (Claude Help Center)
