AI agent ROI without illusions: formula, scenarios, and red flags for operators

You calculate AI agent ROI by comparing the monthly gain the agent produces against what it costs to run, then dividing the one-time setup cost by that net figure to get a payback period. Total ROI percentage is secondary. For a small service business, the question that matters is how many months until the project has paid for itself.

Use this sequence and write each number down before anyone starts building:

1. Count the monthly event volume. This is the number of missed calls, unanswered web forms, quotes with no follow-up, or manual CRM updates that happen every month today.
2. Assign a value to each event. Use gross profit per job, not revenue. For a labor task, use hours saved multiplied by the loaded hourly cost of the person doing it.
3. Estimate how many events the agent actually handles. Not every missed call is a real buyer, and not every real buyer books.
4. Add the running costs. Platform fee, telephony or per-minute usage, and the time a human spends reviewing transcripts and fixing mistakes.
5. Add the one-time cost. Setup, integration, script writing and testing.
6. Divide. Payback in months equals one-time cost divided by (monthly gain minus monthly running cost).

The most common error is step two. Owners multiply recovered jobs by the full job price and ignore materials, labor and overhead. That makes almost any project look profitable. If you want a faster way to size step one for phone leads, the Missed Call Revenue Calculator estimates the monthly and yearly revenue that goes to voicemail.

A worked example turns the formula into something you can check against your own numbers. Every figure below is an assumption chosen to show the arithmetic. It is not research and it is not a SkoreHood client result.

Example: assume 40 missed calls per week at a home-service business. That is about 160 per month. Assume half are real prospects (80), the voice agent books 30% of those (24 appointments), and half of the appointments turn into paid jobs (12 jobs). Assume an average job of $400 with 45% gross margin, so each job is worth $180 in gross profit.

Two things are worth noticing. First, the project survives even when the agent performs at half the assumed rate, which is the test a decision should pass. Second, the result is driven almost entirely by the booking rate and the margin per job, not by the platform fee. If your own margin is thin or most of your missed calls are spam or wrong numbers, replace these inputs and the answer changes quickly.

The costs people forget are the ones that are not on the vendor's pricing page: human review time, usage charges that scale with volume, integration work, and the cost of mistakes. A flat monthly fee is only one line of the real running cost.

Review and maintenance time

Someone has to read transcripts, correct the script when it mishandles a question, and update pricing or service-area rules. Budget a few hours a month at the start, and treat it as a labor cost even if the owner does it personally. Our guide to why AI voice agents fail covers the failure patterns that review is meant to catch.

Usage and per-minute charges

Voice and messaging agents often bill by the minute or by the message on top of a base fee. Check what happens in a busy month. A model that looks cheap at 100 calls can look different at 500. For a comparison of flat and metered models in the answering-service context, see AI receptionist cost.

Integration and data cleanup

An agent that books a job but cannot write to your calendar or CRM creates a manual task instead of removing one. If your CRM has duplicate contacts or stale owners, fix that first or price the cleanup in.

The cost of a wrong answer

If the agent quotes a wrong price or promises an arrival window you cannot meet, someone pays for it. Define what the agent may not say, and define when it hands the call to a person. Build that into the setup estimate rather than discovering it in week three.

The workflows that pay back fastest have a clear dollar value per event, a high event volume, and an existing baseline you can compare against. Missed-call recovery, quote follow-up and lead routing usually meet all three. Open-ended tasks such as "improve customer experience" meet none.

Use these filters when choosing what to automate first:

• The event already happens every week. If you only see the problem twice a quarter, the savings will be too small to measure.
• The cost of the event is visible. A call that goes to voicemail or a quote that gets no reply has a direct revenue link. A vague "better internal communication" does not.
• The workflow has a narrow definition of done. Booked appointment, sent follow-up, updated CRM field. Narrow steps are easier to test and easier to fix.
• A human can take over. If the agent fails, there should be an obvious fallback so the downside is capped.

If you are unsure where the biggest leak is, the Automation Opportunity Scanner scores manual work, slow replies and handoffs so you can pick one workflow to test first. For phone leads specifically, the Missed Calls service is built around exactly this kind of narrow, measurable job.

The clearest red flag is a proposal that cannot name the first metric it will improve and the baseline it will be compared against. Everything else on this list is a variation of the same problem: the value is promised, not defined.

• No baseline. Nobody knows the current answer rate, response time or follow-up rate, so nobody can say whether it moved.
• Replace-the-whole-process scope. A project that touches sales, support and scheduling at once will take months and has no single point where you can check the result.
• Training time with no output. If the first two months produce no live workflow, you are paying for research, not automation.
• ROI quoted as a percentage with no inputs. A figure like "300% ROI" means nothing without volume, value per event and cost.
• No owner for exceptions. When the agent cannot handle a case, nobody is assigned to take it.
• Revenue counted instead of profit. Always ask which number the claim uses.

Anthropic's own guidance on agents makes a related point: start with the simplest solution and add complexity only when it demonstrably improves results, as described in Building effective agents. For a small business, that usually means one workflow, one metric and one human fallback.

You check by comparing the same metric before and after, over at least 30 days, using the numbers you wrote down before launch. If you did not record a baseline, record one now from the last 30 days of history before you switch the agent on.

A simple review cadence works for most small teams:

1. Week 1: read every transcript or message thread. Fix script gaps and wrong answers immediately.
2. Day 30: compare appointments booked, response time and answer rate against the baseline. Calculate actual gross profit from the jobs that came through the agent.
3. Day 60: recompute payback with real inputs. If the half-performance case from your estimate is closer to reality, decide whether to tune the workflow or stop.
4. Quarterly: re-check volume and margin. A seasonal business can swing the result a lot.

Keep a log of calls or leads the agent handled badly, even if the total still looks positive. Those cases show where to tighten rules or add a human handoff. Our note on human handoff for voice agents explains how to set those rules.

If the numbers do not support continuing, a clean stop after 60 days is a fine outcome. The point of ROI work is to make that decision cheap and early, not to defend a project that has stopped earning.