A Quickstart for Measuring the ROI of Your Claude Deployment

Key Takeaways

• Most Claude ROI numbers fail the first follow-up question. A survey of hours saved, multiplied by headcount and salary, gives a number nobody in finance believes.
• Count the full cost before you count any value. Seats and usage are the visible part. Configuration, connectors, maintenance and people's time are the part that gets left out.
• Three named workflows beat a company-wide estimate. Pick work you can count, measure what Claude changes in it, and stop there for the first pass.
• Hours saved are not dollars saved. Apply a realization rate for the share of saved time that turns into more or better work. That one input moves the answer more than any other.
• Measurement is a deployment setting. Telemetry, tagging and reporting get configured at rollout, or you reconstruct them from memory a year later.

Most Claude ROI numbers fall apart at the first follow-up question.

The usual version goes like this. Someone sends a survey asking people how many hours a week Claude saves them. The average comes back at four or five. That gets multiplied by headcount, then by average salary, and the deck says the deployment is worth several million dollars a year. Then the CFO asks the obvious question: if we saved that much, where did it go? Nobody has an answer, and the whole exercise loses credibility.

Meanwhile, the broader data gives finance every reason to be skeptical. MIT NANDA's GenAI Divide report found that 95% of the organizations it studied were getting zero return from their generative AI efforts. (That study is preliminary and built on interviews and public announcements, so treat it as a mood reading, not a census.) Still, if your CFO has read the headline, your ROI number starts with a credibility deficit.

A bigger model or a fancier dashboard won't fix that. A smaller, more honest calculation will. One worksheet. Three workflows. The full cost on one side and a discounted value on the other. You can build the first version in an afternoon, and it will survive a follow-up question.

Here's the quickstart.

ROI Starts With the Cost Nobody Counts

People spend most of their energy on the value side of ROI. The cost side gets a single line: the Claude invoice.

That undercounts a deployment badly. A real Claude rollout has costs that never show up on Anthropic's bill. Someone configures SSO and SCIM, sets up role-based groups, and decides which connectors each team gets. Someone writes the data handling rules and the acceptable use policy. Someone tests each new release against your setup. And every person you train spends hours learning the tool instead of doing their job.

If you leave those out, your ROI looks great on paper and falls apart the moment someone in finance adds them back. Better to add them yourself.

List every cost line before you count a single hour

Split the cost into two buckets.

• Monthly run cost. Seat fees, usage charges, and ongoing maintenance and admin time. This repeats every month.
• One-time setup. Workspace and identity configuration, connector setup, policy work, and the hours people spend in training. This happens once, at rollout.

Put people's time in at their loaded hourly cost, same as you will on the value side. If an internal admin spends a week on configuration, that week is a cost. The math only works if both sides use the same currency.

Three Workflows Beat a Company-Wide Estimate

The instinct is to measure everything. Every team, every use case, every prompt. That's how you end up with the survey-times-headcount number.

For a first pass, pick three workflows. A workflow here means a named, repeated piece of work with a clear finished output. "First draft of the monthly variance commentary." "First draft of an RFP response." "Summary on every escalated support ticket." Things that happen often enough to count and finish in a way you can see.

Three is enough to show a pattern and few enough to measure properly. It also forces a useful conversation about which work Claude is supposed to change. If you can't name three workflows, the problem isn't your ROI model.

Pick workflows with a volume you can count

For each workflow, you need two numbers you can defend: how many times it runs a month, and how much time Claude takes off each run.

Volume usually comes from a system you already have. Tickets closed, RFPs submitted, reports filed. Time saved per run is harder. Time a few runs with Claude and compare them with how long the work took before. Use what you measured, even if it's smaller than what people tell you. Smaller and true beats bigger and guessed.

Hours Saved Are Not Dollars Saved

This is the step most ROI models skip, and it's the one that matters most.

Say Claude cuts a two-hour task to thirty minutes. You saved ninety minutes. What happened to them? Maybe the analyst did another piece of billable work. Maybe they went deeper on the analysis. Maybe they answered email. Maybe they took a longer lunch. All of those are real outcomes, and only some of them show up as value to the business.

The researchers at the St. Louis Fed say this plainly. In their latest survey of generative AI adoption, workers report time savings equal to 1.6% of all U.S. work hours. The authors then caution that measured productivity rises less if people spend the saved time on lower-value tasks or on-the-job leisure.

The task-level numbers are even more generous, which is exactly why they need a discount. Anthropic's own analysis of Claude conversations and productivity estimates that Claude speeds up individual tasks by about 80%. Anthropic is upfront that the figure doesn't count time spent outside the conversation, and that controlled trials have found much smaller gains. I believe the direction. I wouldn't put 80% in a budget.

Apply a realization rate

Multiply your time saved by a realization rate: the share of saved time you believe turns into more or better work.

I'd start at 50% unless you have a reason to go higher. If the saved time clearly goes into more output (more tickets closed by the same team, more proposals out the door), go higher. If it disappears into a busy calendar, go lower. Write down why you chose the number. When someone challenges it, the reasoning is what they'll want to see.

The Worksheet Fits on One Page

Here's the whole quickstart. Three lines of math.

Step 1. Monthly value of each workflow = runs per month × hours saved per run × loaded hourly cost × realization rate.

Step 2. Full cost of the deployment = monthly run cost (seats, usage, maintenance) and one-time setup (configuration, connectors, training time).

Step 3. ROI and payback. First-year ROI = (12 × monthly value − first-year cost) ÷ first-year cost. Payback in months = one-time setup ÷ (monthly value − monthly run cost).

Here's how it looks with numbers in it. These are illustrative figures to show the mechanics, not benchmarks.

Now say the seats, usage and maintenance for the people in these workflows come to $1,800 a month, and setup cost $9,000 in configuration, connectors and training time. First-year cost is $30,600. First-year value is $53,160. That's a first-year ROI of about 74%, with payback in about three and a half months.

Smaller than the survey number. Every input has a source.

Report a range, not a point

Run the same worksheet at a low and a high realization rate. In the example above, 30% gives a first-year ROI of about 4% and payback in about ten and a half months. At 70%, ROI is about 143% and payback is about two months.

Same workflows. Same costs. A huge spread. That spread is the honest answer, and it tells you where to focus. If the realization rate drives the result more than anything else, the work that pays off is making sure saved time goes somewhere useful. That's a management question, and no amount of model tuning will answer it.

Measurement Is a Deployment Setting

This part gets skipped because people file it under reporting. The decisions that make it work all happen during rollout.

Claude Code, for example, can export usage, cost and tool activity through OpenTelemetry, and Anthropic's monitoring docs show how administrators can switch it on for everyone through the managed settings file. That's a deployment task. So is deciding how usage gets tagged by team, which connectors are scoped to which roles, and who owns the monthly report.

If you set those up at rollout, the worksheet fills itself in over time. If you don't, you spend month six trying to reconstruct what happened from invoices and memory.

Put the ROI inputs in the rollout plan

Before the first person logs in, decide three things. Which three workflows you'll measure. Where each input in the worksheet will come from (a system, a dashboard, or a timed sample). And who updates the worksheet each month.

That's how we approach Claude Deployment at PressW: access and data controls get configured before first login, and the deployment comes with monthly reporting so the numbers exist when someone asks for them.

Rerun It Every Quarter

The first version of the worksheet is a starting point. Workflows change. New teams come online. Usage grows as people get comfortable, and the cost side moves with it.

Every quarter, rerun the three workflows with fresh inputs. Then ask whether to add a fourth. The ones worth adding are usually obvious by then, because a team that has seen its own number will tell you what it wants measured next.

Retire workflows that don't pay

The quarterly rerun also tells you what to stop doing. If a workflow's realization rate stays low after a quarter, either the work changed or the saved time isn't going anywhere useful. Fix that or move the effort to a workflow that pays. An ROI model that never kills anything isn't doing its job.

A Rough Number You Can Defend

I'm convinced this quickstart beats the survey-times-headcount number every time. It's smaller. It counts the costs people forget. And every input has a source someone can check.

I'm less sure about the realization rate. Nobody has a clean way to measure where saved time goes yet, and I'd be suspicious of anyone who claims they do. My 50% starting point is a judgment call. I'd rather show my reasoning and adjust it each quarter than pretend the number came from a study.

If you want to try it, pick three workflows this week and fill in the worksheet with what you know today. If the inputs are hard to find, that tells you something about how your deployment was set up. We help teams configure Claude with measurement built in through Claude Deployment and Token & Usage Analytics, and if you want a second set of eyes on your numbers, Book a Discovery Call.

FAQ