A Quickstart for Measuring the ROI of Custom Claude Agents

Key Takeaways

• Most agent ROI cases are written once and never checked. The estimate gets made before the build, and nobody goes back to it after launch.
• Six inputs are enough for a first number. Task volume, hands-off rate, minutes returned, loaded rate, run cost and build cost give you payback in months.
• Hands-off rate moves ROI more than anything else. The share of runs the agent finishes without a person redoing them decides whether the agent pays back in months or years.
• Measure the manual baseline before you build. People are bad at estimating their own time savings, so time real tasks instead of asking.
• Set the decision date and the kill line up front. Read the numbers at day 30, decide at day 90, and agree in advance what "stop" looks like.

Most custom agent business cases have a short life. Someone writes one to get the build approved, it says the agent will save a few hundred hours a quarter, and then nobody opens it again. The agent ships. People use it, or they don't. Six months later a budget review asks what it returned, and the honest answer is a shrug with a usage chart attached.

That pattern shows up in the survey data, too. In IBM's 2025 CEO study of 2,000 chief executives, only 25% of AI initiatives had delivered their expected ROI over the past few years. In the same survey, 68% of CEOs said their organization has clear metrics to measure innovation ROI. Both things can be true at once. Having metrics on paper and actually measuring a specific agent against them are different jobs.

I think ROI on a custom Claude agent is simpler to measure than most teams assume. It needs a handful of inputs, a few weeks of logging, and the discipline to decide in advance what result would make you stop. That's what this quickstart covers: what to measure before you build, the six numbers that go into the worksheet, and how to read them at day 30 and day 90.

Measure the Workflow Before You Build the Agent

The baseline is the step everyone skips, and it's the one you can't go back and redo. Once the agent is live, nobody does the task the old way anymore, so the old number is gone.

You might think you can just ask people how long the task takes. You can't, at least not reliably. METR's 2025 randomized study of experienced open-source developers found they took 19% longer to finish issues when AI tools were allowed, while believing afterward that AI had sped them up by 20%. METR now says those results are out of date for current tools, and I'd agree the tools have moved a lot. The point that holds is about perception. People's sense of their own time is a poor measuring instrument, in either direction.

Model-based estimates have the same problem. Anthropic's study estimating productivity gains from Claude conversations put typical task speedups at about 80%, and then said plainly that the figure leaves out the time people spend checking and refining output outside the chat. The same write-up notes that past controlled trials found much smaller savings, roughly 14% to 56%. That is an honest paper. It is also a good reason to measure your own workflow instead of borrowing someone else's percentage.

Time twenty real tasks by hand

Pick the workflow the agent will take on. Have the people who do it today time 20 real instances, start to finish, including the boring parts: finding the inputs, fixing mistakes, waiting on someone. Twenty is not a statistical study. It's enough to replace a guess with a range, and it takes a week, not a quarter. While you're at it, count how many of these tasks actually happen per month. That number matters more than people expect.

Six Inputs Are Enough for a First ROI Number

Here's the whole worksheet. Six inputs, and each one has a place it comes from.

For R, if you don't have a number from finance yet, the Bureau of Labor Statistics' June 2026 compensation data puts average employer cost for private industry workers at $46.89 per hour worked. Treat that as a floor. Analysts, paralegals and associates cost more.

Write the formula down before launch

Monthly gross value is V × H × (M ÷ 60) × R. Subtract C to get monthly net value. Divide B by monthly net value and you have payback in months. For a 12-month ROI, take twelve months of gross value, subtract the build cost and twelve months of run cost, and divide by that same total cost.

Here's an illustrative example, not client data. An agent sees 400 tasks a month. It finishes and keeps 70% of them. Each kept task returns 25 minutes of a $60-an-hour person's time. It costs $1,500 a month to run and $40,000 to build. That's $7,000 a month in gross value, $5,500 net, payback in about 7.3 months, and a 12-month ROI of about 45%. Not a moonshot. A real, defensible number.

Hands-Off Rate Moves ROI More Than Token Price

Run the same example with one change. Drop the hands-off rate from 70% to 40%, the kind of number you get when the agent handles the easy cases and kicks everything else back to a person. Gross value falls to $4,000 a month. Net value falls to $2,500. Payback stretches from about 7 months to 16.

Now push it the other way, to 85%. Net value rises to $7,000 a month and payback drops under six months.

Nothing else moved. Same volume, same minutes, same run cost. This is the part I'd want every team building a custom Claude agent to internalize: the number that decides ROI is mostly a quality number. You could cut token spend in half in that example and save $750 a month at most, since tokens are only part of the $1,500 run cost. Moving hands-off rate by 15 points is worth twice that.

Engineer for completion, then for cost

That changes what you spend build time on. Evals with real examples of good output. Approval gates on the steps that need a person, so the agent can finish everything around them. Connectors that give the agent the context it needs, so it stops handing tasks back for missing information. Once hands-off rate is where you want it, then tune model choice and prompts for cost. Doing it in the other order optimizes the smaller lever first.

Instrument the Agent on Day One

The worksheet only works if the inputs come from logs. And logs only exist if someone set them up before the first user touched the agent.

You don't need much. For every run, record which workflow it belongs to, how it ended (finished, handed to a person, or failed), and what the reviewer did with the output (kept as is, kept with edits, or redid it). Tag usage and cost by workflow so C isn't a guess. Then, once a week, time a small sample of reviews so M reflects reality and not the baseline alone.

Log outcomes, not just tokens

Token dashboards tell you how much the agent worked. They don't tell you whether the work counted. The single field that makes the worksheet possible is the reviewer's action on each output. Without it, you can't calculate hands-off rate, and without hands-off rate, every ROI number is a guess dressed up as a calculation. (This is also your earliest warning when a model update or a process change starts to degrade quality.)

Read the Numbers at Day 30. Decide at Day 90.

The calendar matters as much as the formula. Here's the rhythm I'd suggest for a first custom agent.

• Before build: baseline. Time 20 tasks, count monthly volume, and write down the kill line.
• Weeks 1 to 4: instrumented pilot. A small group of real users, every run logged with its outcome and the reviewer's action.
• Day 30: first read. Fill in the worksheet with real H, M and C. Expect hands-off rate to be lower than you hoped. That's normal, and it's the input you'll spend the next two months improving.
• Day 90: decision. Compare payback against the kill line and choose: scale it, fix it, or stop it.

Set the kill line in advance

Agree on what failure looks like before you have any data. For example: if payback is longer than 18 months at day 90, or hands-off rate is under 50%, we fix it or stop it. Pick your own thresholds. The point is that they get written down while nobody is attached to the agent yet. After 90 days of work, every team finds a reason to keep going.

Some of the Value Won't Fit in the Worksheet

The worksheet counts time returned. Some agents earn their keep in other ways, and I don't want to pretend otherwise.

An agent that turns a three-day review into a same-day review might matter more for a deal timeline than for labor cost. An agent that applies the same checklist every time might catch errors a tired person misses. An agent that handles intake overnight might mean customers get answers before anyone logs on. None of those show up cleanly in V × H × M × R.

I don't have a clean formula for them. I've seen teams try to convert every one of them into dollars, and the result usually reads as padding, which makes finance trust the solid numbers less.

Keep a second column for what you can't price

List those benefits next to the worksheet, with whatever evidence you have: cycle time before and after, error counts, response times. Don't add them into the payback number. If the agent pays back on time returned alone, the second column is upside. If it doesn't, the second column is where the real conversation happens, and it's a more honest one for being kept separate.

A Rough Number You Trust Beats a Precise One You Never Check

None of this is sophisticated. That's the point. A custom Claude agent should be able to answer one question 90 days after launch: did it pay for itself, or is it on track to? Six inputs and a calendar get you there.

This is how we build at PressW. Every custom Claude agent we ship comes with evals, grading criteria and telemetry, and we start with interviews of the people who do the work, which is where the baseline comes from. The goal is that the ROI worksheet fills itself in from data that already exists, instead of being reconstructed for a budget review.

I'll be honest about what's still open. I don't think anyone has a settled way to measure agents that run on their own across several workflows at once, where one run touches three teams' work. Attribution gets messy fast. For a first agent on one workflow, though, the math above holds up.

If you have an agent in pilot, or one about to start, try this: time 20 tasks by hand this week, and add a "kept, edited, or redone" field to every output. You'll have a real payback number within a month. Our Custom Claude Agent Development team builds that measurement in from day one, and if you want to walk through your workflow with us, Book a Discovery Call.

FAQ