You left a vendor pitch yesterday feeling persuaded. The deck looked good, the demo ran cleanly, the salesperson was friendly without being pushy. You said you would think about it overnight. This morning you cannot remember a single piece of specific evidence for any of the headline claims. You have a follow-up call tomorrow and you would like to know what to ask without becoming the buyer everyone dreads.
This is the everyday reality of buying AI in 2026. Vendor pitches mix three layers in roughly equal measure, and the layers are not labelled. There is marketing language, which is the framing. There are verifiable claims, which can be tested. And there are unverifiable claims, which sound concrete but cannot be checked because they have no methodology underneath them. The skill is hearing the three layers as the pitch unfolds and asking one or two questions that pull the third layer down into the second.
What are the three layers of a typical AI vendor pitch?
The first layer is marketing language: ‘enterprise-grade’, ‘next-generation’, ‘intelligent automation’. It frames the conversation and makes no testable claim. The second layer is verifiable claims, with a named customer, specific outcome, methodology, and time window. The third layer is unverifiable claims, productivity multipliers and percentage savings stated without any of that scaffolding. The work during a pitch is hearing which layer each sentence belongs to.
The third layer is the dangerous one because it sounds like the second. ‘Forty per cent cost reduction’ feels like evidence in the moment, but without a baseline, a methodology, a period and a customer name, it is closer to marketing than to data. The US Federal Trade Commission’s Operation AI Comply enforcement action in 2024 charged several AI vendors over claims of exactly this shape: specific-sounding numbers, no substantiation underneath. The agency’s reasonable basis standard requires evidence to exist before the claim is made, and the same standard is useful for a buyer to apply quietly in their own head.
Why does the layering matter for your business?
Because the buying decision compounds. A vendor’s claims become the basis for your budget, your team’s expectations, and your post-implementation success criteria. When a claim turns out to be marketing rather than evidence, the gap shows up as missed savings and slower payback. Deloitte’s 2024 AI ROI survey found satisfactory payback typically lands in years two to four, not the seven to twelve months vendors quote.
The pattern is not theoretical. McDonald’s ran a two-year IBM drive-thru AI trial that was terminated in 2024 after the system failed under real conditions, despite encouraging pilot results. IBM Watson for Oncology was promoted globally on accuracy claims that internal documents later showed were unsupported, with the model trained on synthetic cases rather than real patient data. Air Canada was held liable by a Canadian tribunal in 2024 for misstatements its chatbot made to a customer. The companies that bought these systems were not careless. They were given confident claims and did not have the discipline to separate the layers at the pitch stage.
How do you move a claim from unverifiable to verifiable?
Four questions do almost all of the work. Compared to what baseline. Measured how. Over what period. With which customers. Each one takes a confident-sounding number and asks for the scaffolding underneath it. A vendor with real evidence answers them in more detail. A vendor without retreats into broader phrasing or pivots to a different example.
Compared to what baseline asks against what. The forty per cent reduction is measured against last year’s process, against an industry benchmark, against the vendor’s own previous tool, or against a manual baseline from before the prospect even had a system. The answer changes the meaning of the number completely. Measured how asks for methodology. End-to-end cycle time including rework, or only the bit the new tool touches. Per-transaction labour hours or staff utilisation. The healthcare claims-audit example is instructive: vendors quoted seventy per cent time reductions for AI screening, but when measured end-to-end including the human review of flagged exceptions, the real saving was fifteen to twenty-five per cent.
Over what period addresses temporal honesty. A first-year saving that includes one-off implementation effects is not the same as a steady-state run rate, and the difference matters across a multi-year contract. With which customers addresses generalisability: results from a large financial-services firm with mature data governance may not transfer to a fifty-person services business with a Xero stack and one operations manager. The right answer names two or three customers of comparable size and sector. The wrong answer says ‘our customers report’ without naming any of them.
When to ask the questions, and when to let things move
In a first-meeting pitch, ask the four questions on the one or two claims you would actually rely on if you signed. Not every claim. A vendor will make twenty claims in forty minutes, and treating all twenty as if they require courtroom evidence will exhaust both of you and tell you nothing useful. Pick the two or three that would change your decision and apply the scaffolding to those.
When a verifiable question gets an evasive answer, do not push twice. One calm follow-up is appropriate, with even more specificity than the first ask: ‘I understand results vary. Can you name one customer who landed near forty per cent and one who landed nearer twenty, so I can see the range.’ If that gets another redirect, signal that you noticed, then move on. The signal sounds like ‘I see, let’s pick this up with your reference customers’. It tells the vendor that the evasion was registered and that verification will happen independently, without becoming the moment the meeting turns sour. Pushing harder rarely produces honesty in real time. Moving on without comment leaves the room thinking you missed it.
There are also moments to let marketing language pass. The opening five minutes of any pitch are positioning, and that is fine. The vendor needs to set context. The mistake is letting marketing-shaped phrasing continue past the first technical section unchallenged. Once the vendor moves to claims about outcomes, the second and third layers are now in play, and the four questions become the right instrument.
How do you close the pitch in a way that surfaces real evidence?
One clean ask, said warmly at the end. ‘Before we go further, I want to verify three claims I will rely on if we sign. For each, send me the underlying evidence, the customer who achieved it with contact details for a reference call, and your measurement methodology.’ Name the three claims explicitly so the vendor cannot cherry-pick which to substantiate.
This ask does four things. It forces a binary, the vendor either has the evidence or admits they do not. It creates a written record of what was asked and what arrives, which becomes useful during negotiation. It signals that the buyer understands the difference between marketing and evidence, which changes how the vendor pitches their next call. And it filters: vendors with real customers and real results will send detailed material within a week, while vendors whose claims rest on marketing positioning will send generic case studies or quietly miss the deadline. Either response tells you something you needed to know.
Related: the companion post in this cluster on reading ‘AI-powered’ claims on vendor websites covers the same discipline applied to the marketing surface before you ever get to a pitch. Vendor case-study survivorship is the layer underneath, why the case studies that do exist often paint a more flattering picture than the average customer experience.
If you want a second pair of eyes on a vendor pitch you are sitting in front of, that is exactly the kind of call I take. Book a conversation.
