AI ROI: The Only Formula That Works for a Small Business (2026)

TL;DR: The Formula

Every vendor deck quotes a triple-digit ROI study. Your books have never met that study. The gap between "AI has great ROI" (true, sometimes) and "this AI purchase had great ROI for us" (measurable, rarely measured) is the entire subject of this post.

The formula that survives contact with a real small business:

AI ROI = (hours returned × loaded hourly cost + attributable revenue − total system cost) ÷ total system cost

Three rules make it honest: measure the hours before you automate, count every cost including your own setup time, and check the number at day 90 — long enough for the novelty to wear off, short enough to kill what's not working before it fossilizes into a forgotten subscription.

Why "What's the ROI of AI?" Is Usually an Unanswerable Question

Asked in the abstract, it's like asking the ROI of "software" or "hiring." The answerable version is per-workflow: what did invoice chasing cost you in hours before, what does it cost now, what did the change cost? That question has a number.

And here's the uncomfortable part: most businesses can't answer it because nobody measured the before-state. They adopted the tool, felt vaguely faster, and now the only ROI math available is the vendor's. As of 2026 I've watched this exact vacuum sell a lot of subscriptions. The fix is embarrassingly cheap — one ordinary week of tally marks on the workflow you're about to automate.

The Metric That Beats ROI: Hours Returned at Day 90

Dollar-ROI is the output, but the input metric I make every client track is hours returned per week, measured at day 90. Two reasons it wins:

• It catches the quiet failure. Small-business AI projects rarely explode — they just quietly stop being used. A subscription nobody opens after week three shows zero hours returned, and the metric says so out loud. (The full list of ways this happens is in why AI implementations fail.)
• Hours convert to dollars honestly. Multiply by the loaded hourly cost of whoever was doing the work and you have a defensible dollar figure — no attribution gymnastics required.

The Back-of-Napkin AI ROI Calculator

Searching for an "AI ROI calculator" mostly finds vendor lead magnets whose output is always "buy the tool." Here's the three-input version that fits on a napkin and doesn't want your email:

• Input 1 — hours/week the workflow takes now. Measured, not guessed.
• Input 2 — loaded hourly cost of whoever does it (salary plus overhead — or your own effective rate, which for most founders is the scariest number on this page).
• Input 3 — total system cost: subscriptions, build fee, integration time, and your own configuration hours at that same loaded rate.

Worked example, real numbers: a follow-up-and-invoice-chasing workflow eats 5 hours a week at a $60 loaded rate — about $1,290/month of labor. Automate it with a $125/month tool subscription (the Claude Team floor, to pick a real price) and even at only half the hours actually returned, you're paying $125 to get back ~$645 every month — roughly 5x, monthly, forever. Run the same hours against a $4,500 one-time install instead: breakeven around month four, and everything after is margin. That asymmetry — small recurring cost against a permanent labor line — is why the honest ROI on well-chosen AI is boring-good rather than press-release spectacular.

My own version of this math, receipts included: I replaced $2,200/year of SaaS with one self-hosted app that costs me $15-25/month to run. That decision keeps paying out every single year.

The Costs Everyone Forgets to Count

• Your own hours. Setup, configuration, prompt-fiddling, fixing the AI's drafts — at your loaded rate. An automation that saves your assistant 3 hours but costs you 2 is not a 3-hour win.
• The subscription pile. AI tools accrete. Audit monthly with a kill rule; three half-used $40 tools are $1,440/year of negative-ROI drag.
• The trust tax. Skip approval gates and one bad customer-facing output costs more than a year of subscriptions. "AI drafts, you approve" is an ROI decision, not a style preference.

When the Math Says Don't

Sometimes the napkin says no, and believing it is the skill. Low-volume workflows (an hour a week isn't worth automating), judgment-heavy work with expensive wrong answers, and anything you can't describe as rules — the honest answer is "not yet." Automating the wrong thing first is the most common way businesses convince themselves AI "doesn't work." Picking the right thing first is a solved problem: what to automate first.

What to Do Next

Today's micro-action: pick your most annoying recurring workflow and start the tally — every time someone touches it this week, one mark and a minute count. Seven days from now you'll have the before-number that 90% of businesses buying AI never collected.

Want the measurement done for you, across the whole business? That's literally the product: my AI Readiness Assessment is $2,500 flat — two weeks, a written roadmap with the ROI math per workflow, and the fee is credited in full against any install within 90 days. Or start free: the AI Readiness Checklist asks the same first questions in 5 minutes, and the consulting cost guide shows what the paid versions of this help run in 2026.

Related guides: why AI implementations fail, what to automate first, what is an AI audit, how much does AI consulting cost.