We did not invent the method broken down here. It was published by the agency AgencyBoxx: in April 2026 it posted on dev.to a year-long log of its own agents across 62 clients and a team of 10 people, and used it to count the time freed up. The number from their case — 32 hours a week — belongs to them, not to us. What is reproducible is the sequence of steps itself: it works on any process, and below it is walked through twice — on their data and on our own implementation.
Why "the agent does something" is not ROI
The typical scenario: the agent is launched, it does something, everyone is happy. Then the budget review arrives, the question "is it paying off?" is asked — and silence falls, because nobody measured. That is exactly why agents are cut first: not because they do not work, but because there is no data to defend the decision with.
An agent's ROI is a formula, not a feeling: saving = baseline × volume − review time, and that saving is compared with the cost of the agent. Everything counted without a baseline is faith.
The five steps
Step 1. Discrete tasks. Not "let's automate support", but a concrete list: email triage, deadline monitoring, picking an answer from the knowledge base, reminders about overdue payments. One process — one row. If a task does not fit into a row, it is not discrete yet: keep breaking it down.
Step 2. Baseline. How many hours the task takes today — real minutes, not "roughly". The most boring step, and the one skipped most often. A week with a timer gives you a number you will not be embarrassed to put on the table in front of the CFO.
Step 3. Volume. Multiply the time of one run by the number of repetitions per week. Fifteen minutes on email triage × 20 times a week = 5 hours.
Step 4. Minus review time. The agent is not autonomous: part of its outputs goes to a human for review. This share is measured in outputs — how many answers out of a hundred someone reads before sending — and the time spent on it is subtracted from the saving. If more than 20% of outputs have to be reviewed, the agent is not ready yet and it is too early to count ROI.
Step 5. Total. The sum of savings across all tasks, multiplied by the number of people on the team the tasks concern.
Example one: someone else's, from an open case study
AgencyBoxx broke down its 32 hours like this:
Task | Hours freed per week |
|---|---|
Email triage | 9.2 |
Deadline monitoring | 7.5 |
Time tracking | 5.8 |
Finding new clients | 4.2 |
Knowledge base queries | 3.8 |
Other | 1.5 |
Total | 32 |
The case study then converts hours into money. The authors quote $183,000–319,000 of "operational capacity" a year — at their own blended agency rates. Your rate is different. 32 hours a week is 1,664 hours a year; at a rate of $25–40 an hour that comes to $42,000–67,000. The wording matters too: the agent did not "save money", it freed up hours that can be moved to clients who pay more.
A freed-up hour turns into money only if there is somewhere for it to go.
Example two: ours, with a network of 500 dealers
A manufacturing company with a network of 500 dealers (the name is not disclosed by agreement) kept two managers as a "human interface" to its accounting system: a dealer asked about stock or a shipment status, the manager opened the ERP and answered.
The same method, step by step. The task is discrete: answering a routine dealer query from ERP data. Baseline and volume were rolled into an annual figure for the routine — ≈UAH 700,000 (≈$17K). The review stayed in place: the agent has read-only access, holds 95%+ accuracy on routine queries, and hands the remaining cases to a human. Costs: ≈UAH 110,000 (≈$2.7K) to build plus ≈UAH 18,000 a year (≈$440) of maintenance. The net effect in the first year — ≈UAH 572,000 (≈$14K), payback 2–3 months. People laid off: zero.
The full breakdown of this case is a separate article, with a layout of four layers of protection.
The unit that makes customer flow easy to measure
When an agent works not with internal routine but with customer requests, hours are awkward to count: conversations vary in length. The unit then becomes the conversation. In our implementation for a large Ukrainian online clothing store with a network of 30+ offline shops (the name is not disclosed by agreement) 6,448 conversations went through the agent between April and July 2026, 1,782 of them in June, with a median response of 13 seconds. Model costs — ≈€0.10 (≈$0.11) per conversation.
In this unit the comparison with a human reads without a calculator: what a conversation costs with the agent, and what the same conversation costs with a manager. We covered this separately in the article AI agent vs manager.
Where the ROI does not add up
The task is rare. A process that takes an hour once a month is 12 hours a year, which will not even cover the setup. Volume is the main multiplier in the formula: no volume, no ROI. The empirical threshold: an agent starts to pay off from roughly 200 requests a month; at a flow of 50–200 it is cheaper to put the process in order first.
Review eats the saving. If the agent makes frequent mistakes and a human effectively redoes the work after it, the saving exists only in the presentation. That is why corrections are logged from day one — every one of them.
Automating a function instead of a process. "Let's replace all of support" is expensive, slow and rarely works the first time. The right sequence is one discrete process at a time, from simple to complex.
When we at Grow2.ai walk a client through this calculation and the numbers do not add up, "it will not pay off" is a result too. It is cheaper to hear that before the implementation.
What the agent itself costs
The saving has to be compared with the full cost, not just with tokens. Starter — €1,800 for a 14-day pilot, then from €49 a month. Practice — €3,600 for 21 days, then from €99. Operator — from €6,000 for 30 days, then from €149. Model costs per conversation are on top. The pilot is paid upfront; if the agreed KPI has not moved at the day-30 review, we refund it in full. Typical payback is 2–6 months; in the dealer case it came to 2–3.
The budget lines missing from the first estimate are collected separately: what an AI agent really costs. Grow2.ai is the AI division of Auspex; setting up the processes and the CRM the agent will live in is handled by Auspex.
A checklist for the first quarter
- Pick one process, not a function — incoming email triage, for example.
- Measure the baseline: how many hours it takes today, without rounding.
- Run the agent for two weeks.
- Log every correction and calculate the saving minus review time.
Want to calculate the ROI for your own process? The free 2-minute AI audit does it on your numbers. General questions about prices and timelines are collected in the AI agents for business hub.