5 AI Marketing Claims to Verify Before Signing a Contract

How to use this checklist

The five claims below show up in nearly every AI marketing pitch. Each one can be true; each one is also the kind of claim that survives a sales conversation without actually being verified. The checklist exists to give buyers exact questions that move from marketing-language to demonstrable evidence — the practical operationalization of the patterns covered in why most agencies oversell AI. Run this list before the contract, not after the first invoice.

Claim 1 — "Trained on your brand voice"

The pitch says the system has been or will be trained on the brand. The verifier: ask for the training data set and the evaluation. An honest answer involves a corpus of approved samples (usually thirty to a hundred artifacts), a documented eval that scores generated output against held-out samples, and a drift test that runs on a schedule. A vendor that cannot describe these three artifacts is doing zero-shot prompting with the brand name in the system prompt — which is fine if priced as such, and a problem if priced as fine-tuning.

Claim 2 — "Outperforms human writers"

The verifier: ask for the comparable test setup. An honest answer cites a specific A/B test with a stated metric (CTR, conversion, time-to-publish), a stated sample size, and a stated definition of "human writer" (the in-house team? a freelance benchmark? a previous agency?). Vendors that decline to specify the comparison are usually citing internal demos rather than client outcomes. The honest variant of this claim is almost always narrower in scope than the pitch makes it sound.

Claim 3 — "24/7 autonomous"

The verifier: ask for the audit log of one weekend's outputs. An honest answer either shows operator approvals on each shipped artifact or names the cron job and pre-approved templates that produced the outputs. There is a real version of this claim where AI handles a tightly-scoped task continuously under operator review — and there is a fictional version where "24/7" means a scheduler runs at 3am. The audit log distinguishes them in under five minutes.

Claim 4 — "AI-generated qualified leads"

The verifier: ask how "qualified" is defined and where the intent signal comes from. An honest answer distinguishes between scraped contact data (a list of emails matching firmographic criteria) and intent-derived signal (behavior, search, declared interest). Most pitches conflate the two. A list of CFOs at companies in the right industry is contact data; a list of CFOs who have engaged with relevant content this quarter is intent-derived. The price difference between the two should be one or two orders of magnitude — and frequently is not.

Claim 5 — "No prompt engineering needed"

The verifier: ask to see the system prompts. An honest vendor shows them. The honest restated claim is "no client-facing prompt engineering" — meaning the buyer never writes a prompt, but a vendor team has done substantial prompt and chain engineering on the backend. A vendor that genuinely has no prompts is either using a fine-tuned model (verifiable by asking which one) or has built a system so generic that brand voice is not actually a feature. Either is fine if disclosed.

What to do with the answers

Run the five questions through one or two pitches and a pattern emerges fast. The vendors that volunteer these artifacts before being asked are the ones worth a pilot — this is the same selection lens we apply at our AI practice and the one we recommend real estate teams shopping AI vendors use during shortlisting. The vendors that route around the questions or shift to abstract capability talk are signaling something specific about what is and is not under the hood.

Bring us a pitch deck and we will run this checklist with you on a 30-minute call. We will flag what is verifiable, what is hand-waving, and what is worth asking in writing.