How to Calculate AI ROI for a Service Business

Questions this post answers

• How do I calculate ROI on an AI agent?
• How do I get a baseline before I switch an AI agent on?
• What costs belong in an AI ROI calculation besides the build?
• What should I measure to know an AI agent is paying off?
• An HVAC owner asks: "I run six trucks. What would an AI agent return?"

Most ROI guides are written for enterprises with a finance team. You have a shop, a phone that rings, and a job board. The method below works with the numbers you already have, one workflow at a time, and it ends in a single figure: how many months until the agent has paid for itself.

How do I calculate ROI on an AI agent?

Pick one workflow, record how it performs today, estimate how it performs with the agent, and turn the difference into gross profit per month. Take off the monthly plan, divide the one-time cost by what is left and you have payback in months. That is the whole calculation; the work is in getting honest inputs.

Gartner surveyed 227 chief sales officers and found that "31% cited difficulty proving ROI of AI-driven tools as a top challenge to sales objectives in 2026." If large sales teams struggle with it, a plan to "see how it goes" will not give you an answer either. Decide the numbers before the agent goes live.

The four steps

1. Baseline. Write down what the workflow does today: volume, conversion, hours, average job value.
2. Cost. Add the one-time costs: the build and your team's setup time. Treat the monthly plan as a running cost that comes off the monthly gain.
3. Gain. Take the expected change, convert it to revenue, then to gross profit using your margin. Revenue overstates the return because you still pay for labor and materials on every job.
4. Payback. One-time costs divided by monthly net gain (the gain minus the monthly plan).

One workflow at a time keeps the math honest. An agent that answers calls, one that chases quotes and one that schedules service visits each get their own sheet, because each moves a different number.

For a first rough pass on office hours alone, the automation ROI calculator takes hours per person and the number of people on the task.

How do I get a baseline before I switch an AI agent on?

Pull two to four weeks of your own records for the one workflow, before anything changes. The baseline is the only thing that lets you say later what the agent did, as opposed to what the season did.

Take a quote follow-up agent as an example. Count the quotes sent last quarter, how many turned into jobs, what a typical job was worth, and how many estimates never got an answer. Our post on quote follow-up walks through that arithmetic. For an answering or booking agent, the baseline is different:

• Inbound calls per week, split by business hours and after hours.
• Calls that reached a person and calls that went to voicemail.
• Of those, how many were new-job requests rather than existing customers.
• Estimates booked, jobs won, average ticket.
• Hours someone spends on the task each week.

Your phone system, your scheduling software and your job-management tool hold most of this. Where a number does not exist, estimate it and label it an estimate. A rough baseline you wrote down is worth more than a precise one you reconstruct after the fact.

Watch for seasonality

If you run HVAC, July and October are different businesses. Compare the same weeks year over year, or run a control: one branch, crew or phone line stays on the old process for a month.

What costs belong in an AI ROI calculation besides the build?

Count five things: the build, any discovery before it, the monthly plan that keeps the agent working, your team's time on setup and review, and changes after launch. Leave any of them out and the payback figure looks better on paper than it will in the books.

• The build. For ballparks, see our pricing: a single agent or proof of concept runs $5k to $25k over 4 to 8 weeks, and an integration $10k to $50k over 6 to 16 weeks. These are ranges, not quotes. The post on what custom AI agent development costs covers what moves a build up or down the range.
• Discovery. A paid two-week discovery is $5,000 and scopes the work before you commit to a build.
• The ongoing plan. Every project carries a managed services plan scoped to it. Put it in as a monthly line, because it recurs.
• Your team's time. Someone answers the setup questions, reviews the first weeks of conversations and corrects the rules. Give those hours a rate.
• Changes after launch. Prices move, services get added, a crew changes shifts. Budget some hours a quarter.

Gartner makes a related caution about time savings: "translating time saved directly into cost savings is not always meaningful." If the agent saves your office manager five hours a week and nothing else changes, you have five hours of capacity, not five hours of payroll. Count it as money only if the hours go to something that earns, or if you would otherwise have hired.

What should I measure to know an AI agent is paying off?

Measure the business number the agent was built to move, then check that it reached profit. Hours saved is the easiest figure to get and the least convincing one on its own.

Two published surveys show the gap. ServiceTitan's report found that among AI users, "66% saving at least three hours per week. However, fewer than half report improved profitability (47%) or increased revenue (43%)." In property management, an IREM and AppFolio survey reported "Eighty-five percent of AI users report a productivity gain" while "Only 48% report a positive impact on NOI or profitability." Both are vendor-linked surveys of users, but the pattern is the point: efficiency is common, and turning it into profit takes a plan.

ServiceTitan's contractors asked for the same thing. In its report, "38% say AI trained on trade-specific data would encourage greater adoption, second only to clearer ROI at 40%."

Pick one lead measure and one profit measure per agent:

Review the lead measure weekly and the profit measure monthly. If the lead measure moves and the profit measure does not, the agent is working and something downstream is stuck: an estimator who cannot get to the extra appointments, or a crew that is already full. That is useful to know, and it points at capacity, not at the agent.

An HVAC owner asks: "I run six trucks. What would an AI agent return?"

With the example numbers below, an agent that books estimates from missed calls pays back its build in under four months, and in about eight and a half if it captures half as many. Say a six-truck shop's office misses calls while every tech is out on a job, and the agent answers those calls, books estimates and writes them to the schedule. Here is the full calculation with invented inputs.

Example inputs, not data from a real company:

• Inbound calls a week: 90.
• Share that goes unanswered after hours or while the office is on other calls: 20%.
• Share of those unanswered calls that are new-job requests: 50%.
• Share of those requests the agent captures and books as an estimate: 50%.
• Close rate on booked estimates: 40%.
• Average ticket: $1,800.
• Gross margin: 35%.
• The build: $15,000, a ballpark inside the single-agent range on our pricing page.
• The managed plan: $500 a month (example figure).
• Weeks per month: 4, to keep the working simple.

Step 1, estimates booked. 90 calls x 20% = 18 unanswered calls a week. Half are new-job requests, so 9. The agent books half of those: 4.5 extra estimates a week, or 18 a month.

Step 2, jobs won. 18 estimates x 40% close rate = 7.2 jobs a month.

Step 3, revenue and gross profit. 7.2 jobs x $1,800 = $12,960 a month in revenue. At a 35% margin, gross profit is $4,536 a month.

Step 4, net gain. $4,536 minus the $500 plan = $4,036 a month.

Step 5, payback. $15,000 / $4,036 = 3.7 months. Over the first twelve months, net gain is $4,036 x 12 = $48,432, less the $15,000 build, so $33,432 ahead.

Now stress the result. Those inputs are guesses until your baseline replaces them, so halve the capture rate to 25% and leave everything else alone. Jobs won fall to 3.6 a month, revenue to $6,480 and gross profit to $2,268. After the $500 plan, net gain is $1,768 a month, and payback is $15,000 / $1,768 = 8.5 months. Year one is $21,216 less $15,000, or $6,216 ahead.

Two things to take from the pair. First, the project survives a pessimistic case, which is what you want before signing. Second, the capture rate and the share of unanswered calls that are real jobs drive the answer more than anything else, so those are the two numbers your baseline must pin down first.

This example leaves out the office hours the agent frees up, which makes it conservative on that side. It also leaves out your own setup hours and any quarterly changes, so add those from the cost list above. If the payback only works when you count soft benefits, treat the project as a weaker bet and say so.

Sources

• Gartner Survey Shows 31% of Chief Sales Officers Cited Difficulty Proving ROI of AI-driven Tools, Gartner (May 19, 2026)
• ServiceTitan report finds contractors shifting focus from AI adoption to implementation and productivity, ServiceTitan with Thrive Analytics (September 29, 2026)
• AI Adoption in Real Estate 2026, IREM and AppFolio (September 18, 2026)