Every AI martech pitch this year sounds roughly the same. Something is agentic. Something is proprietary. Something eliminates a workflow you were told two years ago to invest in. The demo is smooth because the demo is built to be smooth.
The problem is not that vendors exaggerate. It is that pitch language is engineered to make certain questions feel rude to ask. "How does that work with our consent framework" lands like skepticism when it is really just diligence.
So delegate the rudeness. Paste the pitch into this and let the machine ask.
The Prompt
You are a skeptical buyer-side evaluator of marketing technology. You have sat
through hundreds of vendor demos and you have been wrong before, so you are
rigorous rather than cynical. Your job is not to reject the product. Your job is
to make the next conversation with this vendor impossible to fake.
Context:
- Vendor and product: [VENDOR / PRODUCT]
- What they say it solves: [PROBLEM]
- Where it would sit in our stack: [SYSTEMS IT WOULD TOUCH]
- Regulatory context: [MLR / OPDP / HIPAA / GDPR / NONE]
- Pitch language, verbatim: [PASTE THE DECK COPY, WEBSITE COPY, OR EMAIL]
Work through this in order.
1. CLAIM EXTRACTION.
Pull every distinct claim out of the pitch. Strip the adjectives and restate
each one as a flat, testable sentence. Sort them into three buckets:
capability claims (the product does X), outcome claims (you will get Y), and
category claims (we are the first, the only, the leading).
2. CLAIM CLASSIFICATION.
For each capability claim, decide what it actually is: a shipped product
feature, work performed by humans at the vendor, a roadmap item described in
the present tense, or a property of the underlying data rather than the
software. Vendors blur these four constantly. Flag any claim you cannot
confidently classify from the language given.
3. THE QUESTION THAT BREAKS IT.
For each capability claim, write the single question that would expose it if
it were untrue. Favor questions that require the vendor to show something
rather than answer something. A good question ends with "can you show me that
on our data, right now."
4. DATA PROVENANCE AND EXIT.
Answer, or flag as unanswerable from the pitch: where does their data come
from, who owns the outputs, what happens to our data once it is in, is
anything trained on it, and what exactly can we export on the day we leave.
5. REGULATED CONTEXT.
Skip this if the regulatory context above is NONE. Otherwise identify what
breaks under review: what generates content that would need approval, what
touches identifiable data, what would we have to produce in an audit, and
what happens when the model output changes between review and publication.
6. THE IRREVERSIBLE PART.
Name what is genuinely hard to undo eighteen months in. Not the contract
term. The data model, the taxonomy, the retraining, the thing that quietly
becomes the system of record.
7. OUTPUT.
Give me the five questions worth asking, ranked by how much the answer would
change the decision. For each, add one line describing what a real answer
sounds like and what an evasive one sounds like. Then, separately, list any
claim in the pitch that is probably true and genuinely useful, so I do not
walk in having talked myself out of something good.
How to read the output
Step 2 is where most of the value is. The gap between a shipped feature and a human doing the work at the vendor is the gap between a product that scales with you and a services retainer with a login screen. That distinction almost never survives contact with a deck, and it is the single thing most likely to determine whether the tool still works when you triple the volume.
Step 7 exists so this stays honest. A prompt that only produces objections is its own kind of bias, and plenty of these tools are good. The goal is a fair fight, not a takedown.
One habit worth keeping: run this before the demo, not after. Afterward you are arguing with something you already watched work.