How to Tell If an AI Vendor's ROI Claim Is Real

You have a demo on the calendar this week, and somewhere in that deck is a slide with one big number on it. Forty percent less admin time, three times the booked jobs, and no way for you to check whether either one means anything before you sign.

Here is what the next ten minutes gets you: three questions that sort a real ROI number from a marketing one. Ask them out loud in the sales call and you stop paying for results nobody ever measured.

Nobody hands a service-business owner a way to check these numbers, so you nod, you sign, and you find out in month four. Somebody finally did the checking: in July 2026, Praxivara traced 70 widely cited small-business automation statistics back to their primary sources. Seventeen held up.

Quick verdict

What this is: Three questions you ask out loud in the sales call, plus a fourth you ask only if the first three survive. What it saves you: The cost of a bad twelve-month contract. On a $500-a-month tool that is $6,000, and the subscription is usually the cheap part of a bad decision. This week: Pull the ROI claim from your most recent pitch or renewal notice and run it through all three before you sign anything.

Start with arithmetic, because most claims die there

Say you run an electrical shop. Nine techs, about 310 jobs a month, $840 average ticket. The rep says their AI booking assistant delivers 30% more booked jobs.

Take that literally for one second. Thirty percent of 310 is 93 extra jobs a month. At $840 that is $78,120 in new monthly revenue from a tool that costs $500. If that were reliably true, the rep would not be selling you software. They would be buying electrical companies.

So the number is doing something else. It is probably real for somebody, somewhere, under conditions nobody in the room is going to volunteer. Your job is to find out which somebody, and whether they look anything like you.

Question 1: Who counted it, and do they sell the fix?

This is Praxivara’s rule and it is the best one in the category: before a statistic moves your money, ask who measured it, when, and whether the measurer sells the fix. An undated survey of a vendor’s own customers is marketing with a decimal point. That is most of the category: of the 70 statistics Praxivara traced, twenty-one were vendor claims, twelve were stale, and eight had no traceable source at all.

Watch how far apart those two worlds are. The Census Bureau’s Business Trends and Outlook Survey found AI use among US businesses sitting between 17% and 20% from December 2025 through May 2026, with firms of four employees or fewer under 6%. Vendor surveys in the Praxivara audit reported adoption between 58% and 76% depending on who they polled and how the question was worded. Same country. Same year. Three times the number.

You have seen this on this site already. The “78% of customers buy from whoever responds first” line that software reps love has no published study behind it, just aggregator blogs quoting each other. That is what an untraceable stat looks like from the inside, and it is where eight of those 70 ended up, including the most-quoted cash-flow number in the whole category.

What you say in the room: “Who ran that study, what year, and were the people surveyed your customers?” Three facts. If the rep has to check, the number was never load-bearing.

Question 2: What was the number before?

A percentage is a comparison. Without the starting point it is a shape, not a measurement.

“Cuts response time 40%” from what? Six hours or six minutes? “99% accurate” against what a human does, or against a coin flip? Vendor evaluation research keeps landing on the same flaw: ROI projections built on accuracy or speed with no baseline metrics and no stated assumptions. A claim about accuracy is not a claim about money, and the two get swapped in sales decks constantly.

I hold my own sources to this. When SchedulingKit put a 50% reduction in no-shows on its appointment reminder research, I used the figure and told you to treat it as directional, because it is aggregated from mixed sources and it is the exact number a scheduling vendor most wants to be true. That is the standard. Use the number, name what it rests on.

What you say in the room: “What was the before number, who measured it, and over how many weeks?” A vendor who ran a real pilot answers in about fifteen seconds.

Question 3: Can I call somebody my size?

The tell here is the case study library. Enterprise logos, a hospital system, a national franchise, and nothing from a twelve-person shop in a market like yours. Refusing references is a documented red flag on its own, and prewritten testimonials are not references. A reference is a phone number.

Ask for two customers in your trade, roughly your headcount, live for at least six months. Then ask those customers three things the vendor will never tell you: what broke in the first month, how long it took before it actually worked, and what they still do by hand.

If the answer is that small-business clients “prefer privacy,” you are being asked to be the first one.

The fourth question, if the first three survive

Ask what failed in their last deployment and what they did about it.

Every vendor with real installations has a story here, usually involving a data import, a phone system, or a customer who hated the automated voice. A rep who has one will tell it, because fixing it is the proof. Hesitation on that question reads as immaturity, and immaturity on the vendor side gets paid for on your side.

This is the same discipline as deciding whether a project is worth starting at all, which I laid out in the 5-question checklist. One decides if the problem is worth solving. This one decides if the person selling the solution has ever solved it.

What to do this week

1. Pull the claim and write it down. Open the last pitch deck, renewal email, or one-pager sitting in your inbox. Write the headline number on a sticky note with the vendor’s name beside it. If there is no number anywhere in the material, that is already your answer for that renewal.
2. Send one email with the three questions. Copy this: “Before we move ahead, three things. Who ran the study behind the number in your deck, and what year? What was the before-and-after baseline, measured over how long? And can you put me in touch with two customers in my trade with under 25 employees?” Send it and start a clock.
3. Run the claim against your own book. Take their percentage, apply it to your real jobs per month and your real average ticket, and look at the monthly dollar figure it implies. If the implied number is bigger than your best month ever, the claim is measuring something other than revenue, and you need to know what before you sign.

The rep who has the numbers will answer by Tuesday. The one who doesn’t will send you a case study about a bank.