7 Buzzword Red Flags in an AI Sales Pitch
Seven phrases should make you slow down in any AI pitch: "AI-powered" with no named model, "agents" that never act unattended, "trained on your data," "no hallucinations," "fully autonomous," "our proprietary AI," and jargon stacked thick enough to end your questions. None of them proves the product is bad — each one marks a spot where the language is doing work the product might not.
A red flag isn't a verdict; it's a place to dig. The pattern behind all seven is the same: a word with a precise technical meaning being used for its emotional meaning, in front of a buyer the seller assumes can't tell the difference. Be the buyer who can tell the difference. Here's the list, each with the counter-question that settles it.
1. "AI-powered" — with no named model
If the deck says AI-powered and nobody will tell you which model — Claude, GPT, Gemini, an open-weights model, something older — treat the claim as unverified. "AI" spans everything from a frontier language model to a decade-old regression script, and the term doesn't distinguish them on purpose.
Counter-question: "Which model is under the hood, and what happens to our pricing and product when you swap it?"
Renting a frontier model is normal and fine — nearly everyone does it. Refusing to say is the tell.
2. "Agents" that never act unattended
The word "agent" has a testable definition: the system takes a goal and completes multi-step work through tools, on its own, and reports back. A shocking share of products wearing the label are chat interfaces or fixed workflow builders with one LLM step — the full taxonomy is in agent vs chatbot vs copilot.
Counter-question: "Show me a task it completes when nobody's watching — every step, every tool it touched."
The difference is a price tier, not a nuance. Agent-labeled products bill like labor replacement; chatbots bill like software. Pay for the one you're getting.
3. "Trained on your data"
The most reliably misleading phrase in the genre. Nine times out of ten the mechanism is retrieval — your documents fetched into the model's context at question time (RAG, explained in plain English here) — not fine-tuning, which actually modifies a model. Retrieval is a good mechanism! But "trained" implies a bespoke asset, custom timelines, and data-rights implications that retrieval doesn't involve.
Counter-question: "Is that retrieval at question time or fine-tuning? If fine-tuning — who owns the resulting model artifact?"
4. "No hallucinations" (or "100% accurate")
Current language models can produce confident, fluent errors. That's not a scandal; it's a known property of the technology, and honest vendors engineer around it — grounding answers in retrieved documents, constraining output formats, adding review steps, measuring error rates. What they don't do is claim zero. A zero-error claim means the speaker is either overclaiming or quietly describing something much narrower than the pitch implies (say, a lookup over a fixed FAQ list).
Counter-question: "What's the measured error rate on our kind of task, and what's the review process when it's wrong?"
Ironically, the vendor who opens with "here's our error rate and here's the guardrail design" is the trustworthy one. Confidence about limits is the strongest quality signal in this market.
5. "Fully autonomous" — pitched as a feature
Autonomy is a dial, not a destination, and serious builders keep humans at the irreversible steps: payments, external sends, deletions, deploys. "Fully autonomous" as a headline feature tells you the vendor is selling to your fantasy of zero involvement rather than to your actual risk profile — because an agent that can act without limits is an agent whose mistakes act without limits too.
Counter-question: "Which actions require human approval by default, and how do we configure that list?"
A vendor with a crisp answer has thought about your downside. A vendor who says "it just handles everything" has thought about their close rate. This is the same expectation-setting that makes or breaks internal rollouts, covered in how to explain agents to your leadership team.
6. "Our proprietary AI"
Almost no application vendor built the model — frontier models cost sums only a handful of labs can spend. What CAN be proprietary, and genuinely valuable, is everything wrapped around the rented model: the agent loop, the integrations, the domain-specific skills, the guardrails. The red flag isn't the word "proprietary"; it's the implication that the intelligence itself is theirs, held together by vagueness when you probe.
Counter-question: "Which parts of the stack did you build, and which do you license?"
Map their answer onto the five layers — model, agent, skills, tools, orchestration, laid out in how the agentic stack fits together — and the pitch usually reorganizes itself into something you can actually price.
7. Jargon stacking — density as a discovery-killer
The subtlest flag isn't any single word; it's velocity. "Our agentic orchestration layer leverages RAG over your knowledge graph with MCP-native tool calling" — delivered fast, with a glance that checks whether you'll admit to being lost. Jargon stacking works because buyers stop asking questions rather than reveal what they don't know. The sentence may even be true. But its function in the meeting is to end scrutiny, and language deployed to end scrutiny is a red flag regardless of accuracy.
Counter-question: "Take that sentence one term at a time — what does each piece actually do in our deployment?"
This is where fluency changes the room. Once you know the eighty terms, jargon density stops being intimidating and starts being information: clear answers mean a real product; a flinch at precision means the vocabulary was the moat. What that flinch costs when it makes it into the contract is quantified in what jargon confusion costs in vendor negotiations.
The pattern under all seven
Every flag on this list is a precise term doing emotional work: "trained" sounds bespoke, "autonomous" sounds effortless, "proprietary" sounds defensible, "agent" sounds like headcount. The defense isn't cynicism — some of the best products in the category are sold under sloppy language by reps who never built the thing. The defense is a working vocabulary and the standing habit of converting every buzzword back into mechanics before it converts into a line item. Twenty minutes of glossary beats a quarter of buyer's remorse.
FAQ
Does a red flag mean I should walk away from the vendor?
No — it means slow down and ask the counter-question. Plenty of good products get sold with sloppy language by enthusiastic reps. The red flag is only fatal when precision makes the answer worse instead of better: that's when the vagueness was the product.
What's the single strongest question to ask any AI vendor?
"Walk me through one real task the system completed for a customer like us — every step, every tool it touched, and where a human intervened." It converts every buzzword into mechanics at once, and vendors with real products usually enjoy answering it.
Why do vendors say "no hallucinations"?
Because hallucination is the objection every buyer arrives with. But current language models can produce confident errors; honest vendors reduce the rate with retrieval and grounding, constrain outputs, and design review steps. A claim of zero is either overclaiming or describing something narrower than it sounds — ask which.
Is "proprietary AI" always a lie?
No — the scaffolding around a rented model can be genuinely proprietary and genuinely valuable: the agent loop, the integrations, the domain skills. The red flag is when a vendor implies they built the model itself and gets vague when you ask which one is underneath. Renting Claude or GPT is normal; hiding it is the tell.