How to calculate AI agent ROI: $3-4 a day that buys back 32 hours a week

An agency with 62 clients, a team of 10, one Mac in the back room, and $3-4 a day on AI tokens. After 12 months of logging: 32 freed-up hours a week — four full working days. That's the OpenClaw case: the author behind the agencyboxx handle spent a year recording what his agents actually do and published the method on dev.to. The method is worth stealing — which is what I did, testing it on my own 17 agents. Below are all five steps, plus the honest part: where the ROI doesn't add up.

Why "the agent is doing something" isn't ROI

The typical SMB scenario: an agent gets launched, it does something, everyone is pleased. Then budget review comes around, someone asks "is it paying off?" — and silence, because nobody measured. That's exactly why agents get cut first: not because they don't work, but because there's no data to defend the decision with.

AI agent ROI isn't a feeling, it's a formula: savings = baseline × volume − review time, and you compare those savings against the cost of the system. Anything calculated without a baseline isn't ROI — it's faith.

The 5-step method

Step 1. Discrete tasks. Not "automate support" but a concrete list: inbox triage, SLA monitoring, prospecting for new clients, answering routine queries from the knowledge base. One process — one line. If a task doesn't fit on one line, it isn't discrete yet — keep breaking it down.

Step 2. Baseline. How many hours the task takes right now — actual minutes, not "roughly." This is the most tedious step, and it's the one everyone skips. A week with a time tracker produces a number you won't be embarrassed to put on the table in front of a CFO.

Step 3. Volume. Time per run × repetitions per week. 15 minutes of inbox triage × 20 times a week = 5 hours.

Step 4. Subtract review time. The agent isn't autonomous: a human reviews 10-20% of its output. That's a real cost line, and it comes out of the savings. If review takes more than 20%, the agent isn't ready — calculating ROI is premature.

Step 5. The total. Sum the savings across all tasks, multiplied across everyone on the team those tasks touch.

What the OpenClaw case showed

Task	Hours freed per week
Inbox triage	9.2
SLA monitoring	7.5
Time tracking	5.8
Prospecting for new clients	4.2
Knowledge base queries	3.8
Other	1.5
Total	32

At a market rate of $25-40 an hour, that's $183-319K of "operational capacity" a year. Note the wording: not "AI saved money" but "AI freed up hours that were shifted to higher-paying clients." A freed-up hour becomes money only if it has somewhere to go.

Testing it on my own 17 agents

The same method on my Paperclip pipeline: 17 agents on a Mac Mini, $2-4 a day on API tokens, 6 months of data. The content pipeline used to take 20-25 hours a week; now it's 8-10. A delta of ~15 hours × 52 weeks ≈ 780 hours a year, against a system cost of $900-1500. These aren't lab benchmarks, they're working numbers: five years ago I would have hired a separate person for that workload.

Where the ROI doesn't add up

The honest part — without it, the method turns into marketing.

The task is rare. A one-hour process once a month is 12 hours a year, which won't cover even the agent's setup. Volume is the formula's main multiplier: no volume, no ROI.

Review eats the savings. If the agent errs often and a human is effectively redoing its work, the savings exist only in the slide deck. That's why you log corrections from day one — every single one.

Automating a function instead of a process. "Let's replace all of support" is expensive, slow, and rarely works on the first try. The right sequence is one discrete process at a time, from simple to complex.

When we walk through this calculation with a client at Grow2.ai and the numbers don't add up, "it won't pay off" is a result too. It's cheaper to hear that before the rollout than after.

A checklist for your first quarter

Pick one process (not a function) — inbox triage, for example
Measure the baseline: how many hours it takes now, no rounding
Run the agent for 2 weeks
Log every correction and calculate the savings minus review time

Start small and scale up. Not the other way around.

Want to calculate ROI for your own process? The Grow2.ai AI audit does it in a fixed-scope assessment. Broader context: what an AI agent really costs and the AI agents for SMB guide. The processes and CRM the agent will live in are the domain of our sister brand Auspex; for strategy and thinking on AI — Andrew Maryasov.

Frequently asked questions

What is a baseline, and why can't you calculate ROI without one?

A baseline is the measured time a task takes today, before the agent: actual minutes logged over a week or two of observation, not an eyeball estimate. Without it, there's nothing to compare against after launch: the agent is doing something, but the savings can't be proven — and at the first budget review it gets cut. In practice: pick one process, log its time in a time tracker for a week, write down the number. That's your reference point.

How much does it cost to run AI agents?

In the OpenClaw case — $3-4 a day on AI tokens for agents serving an agency with 62 clients and a team of 10. My own example: 17 agents on a Mac Mini burn $2-4 a day, and the whole system comes to $900-1500 a year. But that's the cost of tokens and hardware, not full TCO: in more complex deployments, integration, monitoring, and maintenance usually cost more than the tokens themselves.

How do you convert freed-up hours into money?

Multiply the hours by what an hour of work costs. In the OpenClaw case, 32 hours a week at $25-40/hour came to $183-319K a year. But it's more honest to call that operational capacity, not profit: an hour turns into money only when there's somewhere to redeploy it — clients, sales, product. If the freed-up time has nowhere to go, the savings stay on paper.

When does an AI agent not pay off?

Three typical cases. The task is rare: a one-hour process once a month is 12 hours a year — not enough to cover even the setup. Review eats the savings: if a human is checking more than 20% of the agent's output, it isn't ready to work on its own. And automating a whole function at once: "let's replace all of support" is almost always more expensive and riskier than one discrete process at a time.

Which process should you measure first?

One that repeats daily and has a clear input and output — the classic example is inbox triage. Measure the baseline without rounding, run the agent for 2 weeks, log every correction, then calculate the savings minus review time. One process, two weeks, real numbers — that's enough to decide on the next step.