By Oliver · AI Architect, BuildAClaw · Oct 15, 2026 · 10 min read

How to Measure the ROI of an AI Worker: The Numbers to Track in Your First 90 Days

Most owners can tell you what their AI worker costs. Almost none can tell you what it returns. Here are the 9 numbers that close that gap, with the benchmarks we see at day 30, 60, and 90.

Of the small businesses we've onboarded onto local AI workers, roughly 7 in 10 could not tell us, at day 60, whether the thing was paying for itself. They felt it was helping. Their inbox was quieter. But nobody had a number. And "it feels useful" is exactly the sentence that gets an AI project cancelled the first time a budget review rolls around.

I've set up and tracked enough OpenClaw deployments on Mac Mini M4 hardware to know the pattern. The owners who keep their AI workers, and expand them, all did one thing in the first 90 days: they measured before, during, and after. This article gives you the exact framework. It's the same one we hand to every client.

Why Most AI ROI Calculations Fail Before Day One

The most common mistake is skipping the baseline. You can't claim your AI worker saved 22 hours a month if you never measured how long the work took before. I've watched owners guess "probably about 10 hours a week" and then be shocked when a two-week time log showed 4.5. Guessing in either direction wrecks your credibility with a partner, a bookkeeper, or yourself.

The second mistake is measuring only cost. In the 138 community leads we analyzed, 15 flagged cost as their main worry, and the typical question was some version of "after the initial setup, how much will I spend per month?" That's the right question, but it's half the equation. A $40/month running cost is meaningless until you know it replaced $1,100/month of labor.

The third mistake is a fuzzy scope. "Automate admin" isn't measurable. "Handle inbound lead replies within 5 minutes" is. Before you track anything, pick one workflow, name it, and write down what success looks like in a single sentence.

Step 1: Build Your Baseline (Days -14 to 0)

Before the AI worker touches anything, log the target workflow for two full weeks. Not one. Week-to-week variance in small business operations is high, and a single week will mislead you. Track these four inputs:

• Volume: How many times does the task occur? (Example: 312 inbound emails, 41 invoices, 58 appointment requests.)
• Time per task: Median minutes, not average. Outliers skew averages badly.
• Who does it and at what rate: Use a fully loaded hourly rate. A $22/hour employee costs closer to $28-$30/hour after payroll taxes, benefits, and overhead. For an owner, use what an hour of your selling time is actually worth, not what you'd pay a temp.
• Error and miss rate: How often does it get dropped, late, or wrong? Missed follow-ups and late invoices are the hidden cost most people never count.

Multiply it out. If 312 emails take a median of 4 minutes each, that's 20.8 hours per two weeks, or about 45 hours a month. At $29/hour loaded, that's $1,305/month in labor attached to one workflow. That's your denominator. If you're still deciding which workflow to start with, our guide on 5 signs your business is ready for an AI worker is a good filter.

Step 2: Calculate True Cost (The Part Cloud Vendors Hide)

On the cost side, a local AI worker is refreshingly simple, and this is where the local-first angle changes the math entirely. A cloud AI stack bills you by seat, by token, or by task. A local worker running OpenClaw on a Mac Mini M4 has a fixed profile:

Add everything up for the first 90 days: hardware, setup, three months of running cost, and the time you personally spend reviewing and tuning (budget 3-4 hours a week in month one, dropping to 1 hour by month three). Yes, your time counts. If you leave it out, you'll overstate ROI and lose trust when someone else does the math. For a deeper cost comparison, see AI workers vs SaaS tools: the real TCO comparison.

Step 3: The 9 Numbers to Track

After watching dozens of dashboards, I've cut it down to nine metrics in three buckets. You don't need a BI tool. A single spreadsheet, updated weekly, is enough.

Bucket A: Efficiency (is it saving time?)

• 1. Hours reclaimed per month: (baseline hours) minus (hours humans still spend on the workflow, including review). Target: 60%-80% reduction by day 60.
• 2. Task completion rate without human touch: The share of tasks the AI worker finishes end-to-end. Healthy range: 70% at day 30, 85%+ at day 90. Below 60% at day 60 means your instructions or integrations need work.
• 3. Cycle time: How long from trigger to done. A human might take 6 hours to respond to a lead. A local agent running 24/7 on a Mac Mini typically takes under 3 minutes.

Bucket B: Revenue and Risk (is it making or protecting money?)

• 4. Response time to leads: Faster replies convert better. Track median minutes to first response before and after.
• 5. Follow-up completion rate: The percentage of leads, invoices, or renewals that got every scheduled follow-up. Most owners baseline at 40%-60%. An AI worker gets this to 95%+, and that gap is often where the real money is.
• 6. Recovered revenue: Late invoices collected, lapsed customers re-engaged, abandoned quotes closed. Attribute it carefully: only count what the worker touched and what closed within a defined window (we use 30 days).

Bucket C: Quality and Cost (is it reliable and cheap?)

• 7. Error rate: Wrong, off-brand, or escalated-by-mistake outputs per 100 tasks. Target under 3 by day 90.
• 8. Escalation rate: Tasks the AI correctly hands to a human. This should be 10%-20%. Zero escalations isn't success; it usually means the worker is guessing.
• 9. Cost per task: Total monthly running cost divided by tasks completed. Local workers commonly land at $0.02-$0.15 per task versus $1.50-$6.00 for a human handling the same item.

Step 4: The ROI Formulas (Copy These)

Keep the math simple enough to explain in a meeting. Here are the three formulas we use:

• Net monthly benefit = (hours reclaimed × loaded hourly rate) + recovered revenue + avoided error costs − monthly running cost
• 90-day ROI % = ((total 90-day benefit − total 90-day cost) ÷ total 90-day cost) × 100
• Payback period (days) = total upfront cost ÷ (net monthly benefit ÷ 30)

Here's a worked example from a four-person service business I'll keep anonymous. They automated inbound lead handling and invoice follow-up:

• Upfront: $699 Mac Mini M4 + $1,500 setup = $2,199
• Running cost: $32/month (electricity plus a small API fallback budget)
• Hours reclaimed: 38/month × $29 = $1,102
• Recovered revenue (late invoices collected, leads that would have gone cold): $1,640/month by month three
• Net monthly benefit at steady state: 1,102 + 1,640 − 32 = $2,710

Month one was weaker because of tuning, so total 90-day benefit came to about $5,900 against a 90-day cost of roughly $2,300 (upfront plus three months of running cost and a small tuning allowance). That's an ROI of about 157% at day 90, with payback around day 34. By month six, the same deployment was projecting north of 600% annualized, because nothing about the cost line grew while volume did.

Step 5: Your 30/60/90-Day Scorecard

ROI isn't a single measurement. It's a curve. Here's what "on track" looks like at each checkpoint, based on what we've seen across deployments:

Day 30: Stabilize, don't judge

Expect negative or near-zero ROI. You've paid setup and hardware but the worker is still learning your edge cases. The only metrics that matter this month are no-touch completion rate and error rate. Review the log every few days and fix the top three failure patterns. Most day-30 problems trace to missing context, not model quality. If you want a quick refresher on how the fundamentals fit together, From Zero to AI Employee covers it.

Day 60: Find the leak

By now the trend line should be obvious. If hours reclaimed is under 40%, look at escalation rate. A rate above 30% usually means the worker lacks access to a tool or document it needs (a calendar, a pricing sheet, a CRM field). Fix the access and the number jumps. This is also when to start tracking revenue-side metrics in earnest.

Day 90: Decide and expand

Calculate your 90-day ROI and payback period, then decide: expand, tune, or stop. Above 150% ROI, add a second workflow. Between 50% and 150%, tune scope and re-measure for 30 more days. Below 50% with a clean baseline, the workflow was probably a poor fit, and you should pick a better one rather than abandon the whole idea.

The Local-First ROI Advantage Nobody Talks About

Here's the part I feel strongest about. With a cloud AI tool, success makes your bill bigger. Automate twice as much, pay roughly twice as much in tokens, seats, or task credits. Your ROI curve flattens or even bends down as you scale.

With a local AI worker on a Mac Mini M4, your marginal cost per additional task is close to zero. I've run 50 lightweight agents on a single M4, and the electricity bill barely noticed. That means your ROI curve keeps climbing as volume grows. It's also why privacy-sensitive businesses see an extra benefit: your customer data never leaves the machine, so you aren't pricing in a compliance risk. Our piece on how local AI agents handle HIPAA and data privacy breaks down that avoided-risk value.

One more practical point. Local models like Llama 4 Scout and Gemma 4 handle routine classification, drafting, and routing at zero marginal cost, and you can reserve a cloud model such as Claude Sonnet 4.6 for the rare hard case. In our deployments, that hybrid pattern keeps over 90% of tasks at $0 in token spend, which is why most clients land at $0-$30 a month instead of the $200+ they feared.

Frequently Asked Questions

How long does it take for an AI worker to pay for itself?

In the deployments we track, a local AI worker on a Mac Mini M4 typically breaks even between day 21 and day 45 when it automates at least 25 hours of recurring work per month. Because there are no per-seat or per-token cloud fees, monthly running cost stays near $20-$45.

What is a good ROI for an AI worker in the first 90 days?

Aim for 150%-400% at day 90, meaning every dollar invested returns $2.50-$5.00 in labor savings, recovered revenue, and avoided errors. Under 100% at day 90 usually points to a scoping problem, not a technology problem.

Do I need to track revenue impact, or just time saved?

Track both, but start with time saved because it's the easiest to verify. Add revenue metrics (lead response time, follow-up completion, recovered invoices) by day 30. Time savings alone typically understate total return by 30%-60%.

How do I establish a baseline before deploying an AI worker?

Spend two weeks logging how long the target workflow takes, how often it happens, who does it, and how often it goes wrong. Multiply hours by a fully loaded hourly rate. Without this baseline, you can't prove ROI, only guess at it.

Why does a local AI worker have better ROI than a cloud AI subscription?

Cloud AI costs scale with usage, seats, and tokens, so the more you automate, the more you pay. A local worker has a fixed cost: a one-time hardware purchase plus a few dollars of electricity. ROI improves as volume grows instead of eroding.

Your Next Step: Start the Baseline This Week

You don't need a perfect dashboard. You need one workflow, a two-week baseline, and nine numbers in a spreadsheet. Do that, and at day 90 you'll have a defensible ROI figure instead of a hunch. The owners who measure are the ones who expand from one AI worker to three, because they have the receipts.