How to actually measure the ROI of Claude at your company
"People seem to like it" is not an ROI measurement. Here is the framework for measuring what AI is actually delivering — without needing a data science team.
Most AI rollouts get evaluated on vibes. People seem more productive. The team is positive about it. Nobody is complaining. These are not bad signals, but they are not measurements — and when someone asks you to justify the spend or expand the rollout, you need something more concrete.
Here is how to measure AI ROI without needing a data science team or a dedicated analytics stack.
The three things you are actually measuring
AI delivers value in three ways:
Time savings. The same work gets done faster. A task that took 90 minutes takes 20. A document that took 3 drafts takes 1. These are real hours recovered, and hours recovered are either reinvested in more valuable work or reflect a capacity increase without a headcount increase.
Quality improvements. The output is better. More consistent tone. Fewer errors. More thorough research. Less variation between what your best and your average employee produce. Quality improvements are harder to measure but often more valuable.
New capabilities. Things that were not feasible before are now feasible. A five-person marketing team can now produce content at the volume of a ten-person team. A CS team of eight can handle the ticket volume of twelve without the customer experience degrading. New capabilities do not show up as cost savings — they show up as growth.
Most ROI measurement focuses on time savings because it is the easiest to quantify. But for most organisations, quality improvements and new capabilities are where the real value is.
How to measure time savings
Before and after task timing. Pick five representative tasks (drafting a customer email, producing a weekly report, responding to a common support ticket category). Time how long they take before Claude. Time how long they take after. The delta times the frequency times the number of people doing the task is your time savings.
You do not need to time everyone. Sample five people per team, three tasks each. Extrapolate. It will be directionally accurate.
Rough calculation: If a CS team of 8 saves 45 minutes per day per person through faster ticket drafting, that is 6 hours per day, 30 hours per week, 1,500 hours per year. At a fully-loaded cost of £40/hour, that is £60,000 in time recovered annually. Against a Team plan cost of £[cost per year], the ROI calculation becomes straightforward.
How to measure quality
Quality measurement requires you to define what quality means before you measure it. For each use case, ask: what does a good output look like, and how would I know?
For customer communications: customer satisfaction scores, response time, first-contact resolution rate. These are existing metrics — does Claude move them?
For content: time from brief to approved, number of revision rounds, stakeholder satisfaction. Again, existing metrics that Claude should move.
For internal documents: time in review, number of comments or change requests. Fewer cycles means higher quality first drafts.
If you do not have existing metrics for a use case, the simplest approach is a structured human review: randomly sample 20 outputs per month, rate them against a simple rubric (1-3 scale on accuracy, tone, completeness), and track whether the score changes over time.
How to measure new capabilities
New capabilities are the hardest to quantify because you are measuring something that did not exist before. The approach: document the constraint that was lifted. "Before Claude, we could not personalise outreach at scale — now we send personalised emails to 500 prospects per week. Here is what that pipeline looks like."
This is a before/after narrative rather than a calculation. That is fine — for board presentations and budget justifications, concrete narratives are often more persuasive than estimates anyway.
The 30-day check-in
At 30 days into a rollout, run a structured check-in with each team using Claude:
- What are you using Claude for that you weren't doing before?
- What are you doing faster?
- What have you stopped doing because Claude does it?
- What is not working?
These four questions, answered by five people per team, give you more useful information than any usage dashboard. Usage numbers tell you adoption; these questions tell you value.
What to report upward
When you report to a founder or executive, structure it as: "Here is what we are saving, here is what we are doing more of, here is what we are spending." Avoid percentages without baselines ("50% faster" means nothing without "compared to what"). Lead with the most concrete metric you have. Acknowledge what you cannot yet measure.
The organisations that sustain AI investment are the ones where someone is measuring it clearly enough to tell a credible story about value. That is your job.