The Honest ROI of AI for Small Business: Real Numbers

The ROI of AI for small business is one of the most lied-about numbers in the current vendor pitch landscape. Every vendor deck has a chart showing 5x or 10x return in the first year. Most of them are nonsense.

This post is the version of the math that holds up under audit. What ROI looks like when you actually measure it, the payback periods that are realistic for an SMB, the categories of cost that get hidden, and the trap most owners fall into when they try to do their own calculation.

If you want the broader strategic frame, my AI roadmap for small business post covers where ROI fits in the overall decision. This post is the spreadsheet version.

The honest ROI equation

The basic equation has four components.

Hours saved per month, multiplied by loaded labor rate.

Plus monetized risk reduction or revenue increase, only when clearly attributable.

Minus initial build cost, amortized over a sensible payback window.

Minus monthly operating cost (model API, infrastructure, monitoring) plus internal operating time at the same loaded rate.

That's it. No multipliers for "strategic value". No assumed productivity gains across the rest of the team. No revenue lift from "improved customer experience" unless you can show me the experiment.

The discipline of refusing to add anything you can't measure is what separates real ROI calculations from vendor decks. If a number is in the calculation, you need to be able to point to where it comes from.

Realistic numbers for a triage agent

A specific example, anonymized from a real SMB engagement.

An e-commerce SMB has a support team of four. They handle 1,200 customer tickets a month. Average ticket triage (read, classify, set priority, route to the right agent) takes 4 minutes per ticket. That's 80 hours a month of triage work, before anyone has actually replied to anything.

Build a triage agent. It classifies and routes 90% of tickets correctly, drafts an initial response on the easy cases, and queues the rest for human triage. The team still does the actual replies. Triage time drops from 80 hours to about 18 hours a month (the remaining is exception handling).

Hours saved: 62 per month. Loaded labor rate for a support specialist: $40 per hour, all-in. Monthly savings: $2,480.

Build cost: $22,000 (mid-range for a triage agent with one integration). Operating cost: $180 per month for model API + infrastructure + monitoring. Internal operating time: 3 hours a month from the team lead at $60 loaded rate = $180.

Net monthly value: $2,480 - $180 - $180 = $2,120.

Payback period: $22,000 / $2,120 = 10.4 months.

Year-one ROI (year-one savings minus build cost minus year-one operating cost): $25,440 - $22,000 - $4,320 = -$880. Roughly breakeven in year one.

Year-two ROI is the clean win: $25,440 - $4,320 = $21,120 annual value, against the original $22,000 build.

This is what an honest first AI project looks like. Year one is breakeven or slightly negative. Year two and beyond are clean wins. The compound effect is real, especially when you start adding more agents and the platform work amortizes.

What vendors leave out

The deck version of this same engagement would show something like a 4x ROI in year one. Here's how they get there.

They count the time savings at the senior manager's rate, not the actual person doing the work. So $80 per hour instead of $40. That doubles the monthly value calculation.

They count "downstream productivity gains" of the freed-up team. The math goes: 62 hours saved on triage means the team can handle 50% more tickets, which means the business can support 50% more customers, which translates to revenue. This is sometimes true. Usually it isn't, because tickets aren't the binding constraint on customer acquisition. But it always looks great in a deck.

They don't include the internal operating time. The slide assumes the agent is deploy-and-forget. It isn't. The team lead spends three hours a month tuning, reviewing edge cases, and updating the prompts. At a real loaded rate, this is non-trivial.

They quote the lowest model API tier (Claude Haiku, GPT-4o-mini) and assume it works for everything. In practice, the triage agent ends up using a mix of small and large models depending on the case. Operating costs are 2 to 4x the quoted number.

They ignore retry, monitoring, eval infrastructure, and ongoing model swaps. Real systems need all of these, and they cost both money and time.

The cumulative effect of these honesty gaps is usually a 2x to 4x overstatement of ROI in the first year. The vendor's number isn't a lie exactly; it's a best-case projection with all the caveats removed. Be the buyer who puts the caveats back in.

I've covered some of this in hidden costs of AI implementation, which is the cost-side companion to this post. Read both before signing anything.

ROI by agent type

Different agent types have different ROI profiles for SMBs. A rough guide based on builds I've seen ship.

Internal triage agents. Highest first-year ROI for most SMBs. Clear baseline, clear hours saved, low operating overhead. Payback in 6 to 12 months is realistic. The boring workhorse.

Document extraction agents. Strong ROI when the document volume is high (50+ documents a day) and the extraction would otherwise require human data entry. Payback in 6 to 10 months. Watch out for edge case handling cost; bad architecture can sink the ROI.

Drafting agents (for human review). Variable ROI. The math only works when the keep rate (how often the human accepts the draft with light edits) is above 60% and ideally above 70%. Below that, the human spends nearly as much time editing as they would writing from scratch. Many drafting builds don't clear this bar. Payback 12 to 24 months for the ones that do.

Internal knowledge lookup agents (RAG). Hardest ROI to measure directly. The value is in time saved by team members getting answers faster. SMBs often build these and then can't prove they paid back. Build only when you have a specific use case with measurable baselines.

Customer-facing chat agents. The riskiest ROI category for SMBs. The upside (24/7 support, reduced ticket volume, faster response) is real but hard to attribute. The downside (brand risk, customer complaints, edge case failures) is also real and easier to attribute. Most SMBs should not start here.

The cluster of agent types that produces reliable SMB ROI is the boring one: triage and document extraction. Start there, prove the math, then expand.

What payback period to target

For SMB AI projects, target a 6 to 12 month payback on the first build. Anything shorter is unusual and probably wrong (or it's a small, low-stakes build that wasn't going to move the needle anyway). Anything longer than 18 months for a first build, and the build is probably too ambitious.

The reason this matters: SMB AI builds carry execution risk. The 18-month payback project that's halfway done at month 10 is the project most likely to get cancelled when the budget gets tight or priorities shift. The 8-month payback project that ships at month 4 and is paying off by month 10 is the project that survives an internal budget review.

Right-sizing the first build to the right payback window is the single most important framing decision the owner makes. I've seen owners insist on the ambitious version, watch it stall at month 8, and lose the AI program entirely. The boring, fast-payback first project is what funds the ambitious second project.

When ROI isn't the right framing

Some AI builds aren't really about ROI in the strict sense. They're about capability, brand position, or strategic option value. A few honest cases.

You're building a capability because you know you'll need it in 18 months and you'd rather have it earlier than perfectly timed.

You're building because a customer has asked for it and the contract value justifies the build even if hours-saved math is fuzzy.

You're building to learn. The first agent is partly a way to teach the team how to operate AI before you scale up.

These are legitimate reasons. The framing trap is calling them ROI when they aren't. If the project is really about capability or learning, say so, budget for it, and don't pretend the spreadsheet works when it doesn't. Mixed framing is how SMBs end up with projects that fail both the ROI test and the capability test.

For most first AI projects in an SMB, though, ROI is the right framing and the math works when done honestly. The vendor deck math doesn't, but the honest version does. Use the honest version.