How do you measure the ROI of Claude in your company?

Measure the ROI of Claude, powered by Anthropic, by comparing a baseline for specific tasks with the same tasks after rollout, across four layers: adoption, task outcomes (time, throughput and quality), cost per task, and the business result the task feeds. Claude Team and Enterprise analytics cover adoption and spend. Time, quality and business impact need your own baselines, time studies and review samples, collected before people start using Claude.

Why do most AI ROI numbers fall apart?

Most AI business cases multiply an assumed minutes-saved figure by headcount and salary. Finance teams reject those numbers for three reasons. First, the minutes were never measured; they came from a survey or a vendor estimate. Second, saved time does not equal saved cost unless it is redeployed to work with value, or unless it avoids a hire. Third, nobody checked whether quality held. A faster first draft that needs twice the editing is not a saving.

A defensible measurement plan fixes all three. It names a small set of workflows, measures them before and after, checks output quality on a sample, and counts the full cost of the tool. It also separates what Claude analytics can tell you from what only your own data can.

What is a practical framework for Claude ROI?

Report all four together. Adoption without outcomes is activity. Outcomes without cost is a partial picture. Business results without the task layer cannot be attributed, because many things move a win rate.

Step 1: Which workflows should you measure?

Pick three to five workflows that are frequent, have a clear unit of work, and already leave a record somewhere. Define the unit precisely. 'Writing' is not a unit. 'First draft of a security questionnaire response, measured from assignment to submission for review' is. Good candidates usually include one drafting task, one analysis task and one lookup or question-answering task, so you learn where Claude helps and where it does not.

• Sales: first draft of an RFP section, account research brief before a discovery call.
• Support: case summary at escalation, knowledge article draft from a resolved case.
• Finance: monthly variance commentary per cost center, reconciliation exception notes.
• HR and legal: policy question answered with a citation, first-pass contract clause comparison.
• Engineering: pull requests merged, review turnaround, test coverage on changed files.

Step 2: How do you build a baseline?

Collect baselines for two to four weeks before the target group gets access, or keep a comparison group that starts later. Three methods work, in order of reliability. System timestamps are best: ticket opened to summary posted, document created to submitted for review. Observed time studies come next: a sample of tasks timed by the person doing them with a simple start-stop log, at least 20 to 30 instances per workflow so one unusual task does not swing the median. Self-reported estimates are weakest and should only fill gaps.

Baseline quality at the same time. Write a short rubric per workflow, for example accuracy, completeness, tone and policy compliance scored 1 to 4, and have a reviewer score a sample without knowing who wrote it. Record the rework rate too: the share of outputs sent back for correction. Without a quality baseline you cannot show that speed came without new errors.

Step 3: What can Claude analytics tell you about adoption?

Claude includes an analytics dashboard for Team and Enterprise organizations. On Team, Owners and Primary Owners can view it. On Enterprise, Owners, Primary Owners and Admins can, though Admins see everything except spend. As of October 2026 the dashboard groups metrics into sections that answer who is using Claude, how they are using it, what the results are and what it costs.

A spend report can be exported as CSV for month to date, last month, the last 90 days or a custom range, with fields including user email, product, model, total requests, prompt and completion tokens, and net and gross spend in US dollars. Data refreshes daily with a one-day delay. That file is the raw material for cost per task.

Enterprise organizations can also pull data programmatically through the Claude Enterprise Analytics API, which reports per-user daily activity across chat, Claude Code and other products, organization-level active-user summaries, project, Skill and connector usage, and, on usage-based Enterprise plans, cost and usage reports. Only the Primary Owner can create its keys. Data is available from January 1, 2026, engagement data typically lands with a one-day lag, and cost figures can be revised until about seven days after month end, so close a month's ROI numbers only after that window. Companies using the Claude Platform rather than Enterprise can use the separate Claude Code Analytics API with an Admin API key.

Step 4: How do you measure outcomes after rollout?

Repeat the baseline method on the same workflows after people have had four to six weeks to learn the tool. Measuring in the first week understates the gain, because people are still learning. Use the same timestamps, the same time-log form and the same rubric. Score quality blind: mix pre-rollout and post-rollout outputs and have a reviewer score them without knowing which is which. Track rework rate for at least a month, because errors in AI-assisted work often surface downstream.

Count throughput as well as time. If a support lead now writes knowledge articles for every resolved escalation instead of a third of them, minutes per article may not move, but the backlog does. That is often the larger result.

Step 5: How do you calculate cost per task?

Cost per task is the total Claude cost for the group in the period divided by the units of work completed with Claude. Seat cost comes from your plan. As of October 26, 2026, claude.com/pricing lists Team standard seats at $25 per month billed monthly ($20 billed annually) and Enterprise at $20 per seat per month billed annually with usage billed separately at API rates. On usage-based Enterprise plans, use the spend report or the Analytics API cost endpoints to allocate usage by user and product. Add your internal costs too: admin time, training hours and any partner fees, spread over the first year.

The worked example below uses hypothetical numbers to show the arithmetic. Replace each input with your own measurements.

Two cautions apply even when the math looks good. The $7,200 is capacity, not cash. Say what the team did with the 160 hours, such as clearing backlog or absorbing volume growth without a new hire. And the seats serve other work too, so allocating the full $750 to one workflow is conservative, which is the right direction for a business case.

Step 6: How should you report ROI to leadership?

• Use one page per quarter: adoption trend, three to five workflow results with before-and-after numbers, cost per task, and one business metric per workflow.
• State the method next to each number: timestamps, time log or estimate. Leaders trust a smaller measured number more than a large estimate.
• Report workflows where Claude did not help. That builds credibility and redirects training.
• Show seat utilization and reassign unused seats before asking for more.
• Re-baseline once a year, because the work itself changes after a rollout.

What are the most common measurement mistakes?

• Starting measurement after rollout, which leaves no baseline to compare against.
• Using dashboard estimates of time saved as the result instead of as a prompt to measure.
• Measuring only volunteers, who are usually the most enthusiastic users and inflate gains.
• Ignoring quality, so faster work with more errors looks like a win.
• Leaving out usage charges on Enterprise, or internal admin and training time, from the cost side.