How to spot a credible AI proof: reading a vendor's case study (or ROI claim) before you sign

Sooner or later the slide appears. A vendor shows you a polished case study — the client's logo, an enthusiastic quote, one big number in the middle: “+40% productivity”, “300% ROI in six months”. Or it's even leaner: a single round figure, promised as though it were a fact. The question that matters isn't whether the number is high. It's a different one, and it has to be asked before you sign: how do I know this proof is credible?

You don't need a multinational's procurement department to answer it. You need a lens — the same one the analysts and finance directors who assess these numbers for a living use. In this article we lend it to you: seven things to look for in a case study or a promised return, to tell proof that holds up from proof that's just well-laid-out marketing. Because the bad news is that the second category is enormous — and the good news is that recognising it is easier than it looks.

Verifiability beats polish

The first instinct is to look for the logo: if they name a real, verifiable client, I'll trust it. Fair enough — a named, checkable client remains the strongest credibility signal there is. But the gap with a well-built anonymous case study is smaller than people assume. Blind-but-verified testimonials earn buyers about 60% trust against the 64% of named ones: just four points apart (UserEvidence, “The Evidence Gap Report” 2025; ProofMap, on anonymous case studies). On one condition, though: that the anonymous version offsets the missing name with rich specifics — exact percentages, timelines, industry and company size, quotes attributed to a name and a role. A generic description — “a leading company in the sector” — is worth next to nothing, logo or no logo.

The lesson for the reader: don't dismiss an anonymous case just because it's anonymous, and don't trust a named one just because it's named. Look at the density of the detail. An anonymous case that tells you industry, size, period, numbers and a quote with a name and a role is more credible than a prestigious logo stuck onto an empty sentence.

It also helps to know what kind of proof you're looking at. The 2025 Edelman Trust Barometer ranks the formats by the “believability” B2B buyers perceive, highest to lowest: original or primary research (70%) > expert opinions (64%) > peer insights (62%) > customer testimonials (55%) > third-party awards and endorsements (42%) (Edelman, “2025 Edelman Trust Barometer Special Report: Brand Trust, From We to Me”, June 2025). Translated: the badge or award on the homepage — last on the list — weighs less than the client's number and quote; and a methodology written up as research weighs more than either. When a vendor shows you mostly logos and accolades, they're offering you the weakest form of proof there is.

Why the numbers, on their own, aren't enough

There's a precise reason the big number alone doesn't settle it: the people who assess these purchases for a living discount self-reported ROI. Only 14% of the finance chiefs surveyed say they've seen clear, measurable impact from AI spend so far (CFO.com, “So far, few CFOs see substantial ROI from AI spending”, RGP survey of 200 US finance chiefs, 2025). Not because the models don't work, but because they've learned to distrust figures modelled and projected by the seller itself. Another data point confirms it: 71% of CFOs believe standard ROI metrics are poorly suited to capturing AI's value (EY survey, cited in CFO.com's reporting). If even the most technical audience doesn't trust the naked number, neither should you.

The point isn't to ask for a bigger number. It's to ask the vendor to state how they measured it. A “300% ROI” with no method is an advertising line; the same 300% with a note of what went into the cost, over which period, against which baseline, becomes something you can assess. It's the difference between the headline and the method — and credibility always lives in the second.

The model to imitate: Forrester's Total Economic Impact

There's a de-facto standard for vendor ROI studies, and it's worth knowing because it gives you the yardstick to measure all the others: Forrester's Total Economic Impact (TEI). Not because every vendor can afford one, but because its four pillars are exactly the questions to put to any proof, even the most homespun:

• An independent interviewer. The interviews are conducted by Forrester's analysts, not the vendor. Whoever gathers the proof isn't whoever sells it.
• Real, live customers. Companies that actually run the technology, live, for a stated minimum period — often six months or more. Not a three-week pilot passed off as a settled result.
• A four-part financial model. Cost, Benefit, Flexibility and Risk — not benefit alone. A case study that shows you the gains and hides the costs is telling you half the story.
• Risk-adjustment. The benefits are scaled down according to the uncertainty and caveats the interviewees themselves state. The final number is deliberately conservative, not optimistic.

Don't expect every vendor to hand you a TEI. But use its four points as a lens: who gathered this data? how long has the client been live? are they showing me the costs too? is the number conservative, or the most optimistic projection possible? Those four questions alone are enough to separate serious proof from a brochure.

The two flaws that sink a proof

There are two methodological errors that, once you've learned to see them, you won't be able to un-see. They're the two silent killers of a case study's credibility.

The first: no counterfactual, no baseline. A before/after comparison with no baseline attributes every change to the vendor's intervention — including the ones that would have happened anyway (Better Evaluation, on comparing to the counterfactual). If the case-study company also hired people, changed its prices or enjoyed a favourable season over the same period, how much of that “+40%” is down to the tool? A credible case names the baseline period and says explicitly what else was changing in that window — headcount, pricing, seasonality. One that presents a clean before/after, as if the rest of the world had stood still, is asking you for an act of faith.

The second: survivorship bias. Publishing only the clients who did well makes success look like the norm rather than the exception (HubSpot, on survivorship bias in sales). The “success stories” page is, by definition, the selection of the survivors: you'll never see the projects that failed, were abandoned or clawed their way back. The more seasoned buyers discount “greatest hits” pages for exactly this reason. The right question to ask out loud is blunt: out of how many total clients are these the results? And the ones that didn't work — what was different about them? An honest vendor has an answer; one that only sells the shop window changes the subject.

Beyond the written case study

There's a proof stronger than any PDF, and often you only have to ask for it. A live reference call — a phone call with a real client — or a small customer advisory board is structurally more solid than any written asset, because you get to ask unscripted questions: about the real implementation problems, the true cost once it's all in, the quality of the support when something breaks (Sword and the Script; Customer Experience Dive). A case study is a monologue edited by the vendor; a reference call is a conversation they can't control. If a seller hesitates to put you in touch with an active client, that's a data point in itself.

It's worth knowing, too, how the big evaluators defend trust. Gartner Peer Insights verifies the reviewer's identity — corporate email domain, role, tenure — before a review counts for anything, because it treats unverifiable identity as the primary point where trust breaks (Gartner Peer Insights). It's the same principle you can apply on a smaller scale: a quote with a name, a role and a verifiable company is worth infinitely more than an anonymous “an executive in manufacturing”.

The buyer's underlying question

If you think all this suspicion is excessive, the numbers say it's the norm. 67% of B2B buyers have ruled out a vendor precisely because the proof offered felt untrustworthy (UserEvidence, “The Evidence Gap Report” 2025). Not “unconvincing”: untrustworthy. Weak proof isn't neutral — it costs you the sale. And there's a gap that explains why: buyers name hard numbers as the most trustworthy proof (51%), yet 40% say vendors don't provide them. That's the “evidence gap”, the distance between what the buyer wants to see and what the seller puts on the table. When you assess a proof, what you're really measuring is how far that vendor has closed — or ignored — this gap.

The checklist: how to read an AI proof before you sign

Let's put it in operational order. Here's what to demand and verify in a case study or an ROI promise, before you put a signature down:

• A named client — or anonymous but rich in specifics. With quotes attributed to a name and a role, not a generic “a leading company”. The absence of detail is itself a signal.
• A stated interview method and a minimum live period. Who gathered the data and how long the client has been live — at least six months, TEI-style. A result from a few-weeks pilot is not a settled result.
• Cost shown net of the vendor's fee. Benefit without cost is half the equation. A real ROI starts from the net, not the gross.
• Baseline period named and concurrent changes flagged. What else was moving in that window — headcount, pricing, seasonality — so you don't credit the tool with what would have happened anyway.
• A number that's risk-adjusted or caveated. Be wary of the clean round figure. An honest number carries its own reservations; a polished one hides them.
• Methodology written up as primary research, not a testimonial. The single most credible format (remember the Edelman hierarchy) is the one that explains how it was measured, not the one that merely cheers.
• Ask for a live reference call. The proof the vendor can't edit. If they grant it without hesitating, that's already half the answer.

How we read — and how we'll build — a proof, at Innesti

We start from a simple principle: you declare the method before the headline. That's how we think you should read someone else's proof — and by the same yardstick we use to judge a vendor's numbers, we judge, and will build, our own. We won't show you a big number without telling you how we measured it, over which period, against which baseline and at what net cost: because a figure that doesn't survive the lens we've just handed you wouldn't deserve your trust. Credibility, for us, isn't the shine of the slide — it's the verifiability of what sits underneath.

Before you even assess a single vendor's proof, though, it's worth knowing which process you want to automate and with what realistic expected results: that's where what to demand of a case study follows from. Our AI-readiness assessment helps line up the right processes and sensible expectations. And if you're already reading proposals, two sibling articles complete the lens: the one on how to assess the reliability of the SaaS vendor (SOC 2, ISO 42001, what to ask about security and your data) and the one on how much AI really returns (the realistic ROI benchmarks against which to test any promise). The first tells you whether to trust the vendor; the second whether the number is plausible; this one, how to read the proof itself.

We've turned the first step into a free, self-serve assessment: a few questions and a pointer on where to start, with what results to expect and what proof to demand from vendors. Take the AI-readiness assessment — then, if it makes sense, let's talk.

This article is purely for orientation and does not constitute purchasing advice nor an assessment of any individual vendor. The percentages cited (trust in testimonials, the hierarchy of proof formats, CFO scepticism on ROI, the evidence gap) come from the industry sources named in the text — Edelman, UserEvidence, CFO.com/RGP, EY, Forrester — and reflect data current to July 2026: they should be re-verified against the original sources before you base a decision on them. The methodological criteria (counterfactual, survivorship bias, reference call) are general evaluation principles, not guarantees.