A rising AI bill shows higher spending, which may reflect more usage, higher prices or a different mix of products. To judge whether the investment is worthwhile, you also need to know what work was accepted, how it compares with work done without the tool, and what changed in costs or useful delivery.
Anthropic’s 28 October announcement of smart reports, surveys and pooled group budgets for Claude Enterprise offers a useful starting point. These tools can help teams locate spending, investigate workflows and ask employees where AI helps. A defensible return on investment (ROI) calculation needs business records beyond the reporting dashboard.
Here is a practical way to connect those signals to an investment decision, using one team workflow at a time.
What Claude Enterprise reports actually measure
According to Anthropic’s smart-report documentation, checked on 4 November 2026, reports analyse sampled sessions for a chosen team, selected products and a period covering up to the previous 28 days. The analysis groups work into workstreams and describes deliverable types, task outcomes, friction and costs.
That can help an administrator identify a promising question: why does one workflow consume a large share of reported spending, or where do users repeatedly encounter problems?
Read the cost figures carefully. Smart reports use list prices. Contract terms, discounts, credits and billing arrangements need reconciling with actual invoices before those figures become a financial cost measure.
The setup previews how many sessions will be sampled. The documentation does not establish a general sampling fraction or statistical representativeness. Record the selected dates, products and session count before comparing reports. One session also need not equal one completed task, and a model’s classification of a deliverable does not establish that someone approved or used it.
Enterprise surveys add employee responses about value and time. Results include response counts, response rates and the number of surveys shown. Keep those denominators and the question wording alongside any headline finding. People who answer may differ from those who dismiss the survey, and recalled time savings need checking against observed work.
Pooled group budgets control access to shared spending allowances. Requests count against applicable member limits as well as the pool. On seat plans, pools cover usage credits; included seat usage remains available after a pool is exhausted. A configured budget therefore needs its billing context before it can be interpreted as a spending limit.
Check eligibility and privacy separately for each feature
The three features have different conditions in the documentation checked on 4 November:
- Smart reports are Enterprise-only and off by default. They are unavailable with customer-managed encryption keys (CMEK), HIPAA configuration or hybrid inference on the customer’s cloud provider. Zero-data-retention settings for Claude Code exclude Code from reports while chat and Cowork remain covered. Organisations can create up to 10 reports per calendar month at no charge.
- Surveys are unavailable with HIPAA configuration but support CMEK. For newly added CMEK, roughly the first day’s survey data remains encrypted with Anthropic’s key.
- Pooled budgets support usage-based Enterprise and Standard/Premium seat plans with usage credits enabled.
Avoid describing the collected data as anonymous. Smart reports have individual attribution off by default, but enabling it allows viewers to reveal names and session IDs, with reveals logged. Restricted-topic sessions receive aggregate analysis without summaries; personal conversations also omit sessions and names. Survey exports can contain email addresses and answers, and responses go to organisational administrators.
Anthropic explicitly prohibits using smart reports for individual performance evaluation or employment decisions. Keep the evaluation at the workflow or team level, check the organisation’s privacy requirements before collecting evidence, and restrict access to the records needed for that evaluation.
Define an accepted outcome before counting productivity
Choose one repeatable workflow and write down what success means before measuring it. For a support team, that might be a customer answer approved under a defined accuracy and completeness rule. For an analyst, it might be a brief accepted for actual use.
Count accepted outputs, rejected work, defects, review time and later rework. An increase in generated drafts could otherwise look productive while transferring effort to reviewers.
A study of coding agents across 718 firms illustrates the distinction. Code volume and review time rose, while the study found no statistically significant change in resolved issues or epics.
The FinOps Foundation’s unit-economics framework distinguishes resource units, such as tokens or seats, from business units, such as transactions or resolved cases. Applying that distinction, a useful local measure is:
Cost per accepted output = total workflow cost ÷ outputs accepted under the agreed quality rule.
Use the same workflow boundary and period for both sides of the calculation. Include the cost of failed attempts in the numerator. Compare quality and total spending alongside unit cost: a cheaper average output can still accompany more defects or a larger overall bill.
This is a proposed evaluation measure, rather than a standard field in Anthropic’s reports. Acceptance comes from the team’s work records and review process.
Compare similar work and include review time
A baseline describes what happened under the previous way of working. A comparison asks what would probably have happened during the evaluation period without the AI intervention.
The distinction matters. A team might finish more work after introducing AI because demand changed, experienced staff joined or the tasks became easier.
The UK Treasury’s guidance on evaluating AI interventions recommends defining objectives, outcomes, assumptions and a baseline early. It discusses randomising access or rollout timing, alongside other evaluation approaches. This is methodological guidance that businesses can adapt, not evidence that a particular product produces a return.
For a proportionate team evaluation:
- Choose a period that covers a normal work cycle and enough relevant tasks to inform the decision. There is no universal minimum duration for every workflow.
- Compare similar tasks using the same acceptance criteria. Where practical, allocate AI access randomly or use a planned phased rollout.
- Measure time through review and rework, including unsuccessful attempts.
- Record task mix, staff experience, training, model/version and workflow changes.
If only a before-and-after comparison is feasible, describe the other changes that could explain the result. It can still inform a local decision, provided the uncertainty remains visible.
Reconcile the bill and account for the whole workflow
Start with actual invoices and contracts for seats, subscriptions, APIs and usage credits. Then identify implementation, integration, training, oversight, infrastructure and review costs relevant to the workflow.
The FinOps allocation guidance supports direct mapping where possible and a documented approach to shared costs. For example, a shared integration bill could be allocated using a stated usage share, or retained centrally with that choice made explicit.
Use one currency and evaluation period. Separate one-off setup costs from recurring costs, and explain how much setup cost the period carries. Show incremental costs relevant to expanding or stopping the pilot separately from a fully allocated view of operating it.
Reconcile each cost once. Adding list-price consumption to an invoice that already covers that consumption would double count spending. Allocating part of a shared bill moves an existing cost into the workflow calculation; it does not create another bill.
Establish what happened to the time or money saved
A survey response about saved time is a lead for measurement. Observed time freed can create capacity, but an unchanged salary still leaves the payroll bill unchanged.
Multiplying hours by an assumed loaded hourly rate gives a valuation of that capacity. Label the rate and the assumption. To support a claim about useful additional capacity, record how the freed time was redeployed and what accepted work resulted.
Financial evidence can include reduced paid overtime, an avoided contractor invoice or another documented spending change, assessed against a credible comparison. Extra output also needs demand and value evidence. If revenue rises, consider how much can reasonably be attributed to the intervention and the costs of delivering that revenue.
Keep benefit categories separate. Counting the same freed hours as both payroll savings and the full value of extra output inflates the result. Report meaningful improvements in quality or risk separately when a defensible monetary value is unavailable.
FinOps expresses financial ROI as (financial benefits − costs) ÷ costs × 100. Before using it, define the period, cost scope and benefit categories. A return built from assumed hourly values should clearly expose those assumptions; documented cash savings and estimated capacity value answer different investment questions.
A compact worksheet for the next evaluation
Use this worksheet for one workflow and one defined period. It is a measurement plan, not a completed assessment or an estimate of any organisation’s return.
| Question | Evidence to collect | Assumption or missing data | Next measurement |
|---|---|---|---|
| Where is AI being used? | Report scope, sampled-session count, survey shown/responded counts | Samples and respondents may not represent the team | Record dates, products, populations and survey wording |
| Is the output useful? | Approval records, rejected outputs, defects and rework | Transcript labels do not establish acceptance | Audit outputs against the same quality rule in both conditions |
| Did AI change productivity? | Comparable task records and timings through review | Task mix, experience or other changes may explain the difference | Compare like tasks; randomise access or phase rollout where practical |
| What did the workflow cost? | Invoices, contracts, staff effort and shared-cost records | Discounts, allocations and setup-cost treatment need agreement | Reconcile costs once and document inclusions |
| What benefit was realised? | Verified spending changes or additional accepted work | Freed hours need evidence of redeployment; output needs demand/value evidence | Have the relevant finance or workflow owner verify the change |
| Is the evidence enough to decide? | Scoped costs, supported benefits and remaining uncertainty | Attribution or valuation may still be unresolved | Collect the evidence most likely to change the expand, revise or stop decision |
If invoice reconciliation is the main gap, resolve that before commissioning another usage survey. If outputs are plentiful but acceptance is unknown, audit a proportionate set of completed tasks. If reported savings depend on time nobody has tracked, measure the full workflow through review and rework in the next normal work cycle.
