5 Sentences AI Agent Vendors Say That Actually Mean They Cannot Do It
Every AI agent sales call has the same five evasions. Here is what each one really means, the question that flushes it out, and what to do when you hear it.
We have sat through a lot of AI agent sales pitches, both as buyers scouting the market and as an agency talking to prospects who just came off calls with someone else. Almost every pitch uses the same handful of evasions when the customer asks a hard question. If you have not heard these before, you will as soon as you start looking.
Here are the five most common. Each one has a decoder, a follow-up question that flushes out the real answer, and a rough guide for what to do when you hear it.
#1. "We are model-agnostic"
What they mean: "We do not know which model we are actually using, and switching providers would break our product."
Model-agnostic sounds like a strength. It usually is not. A tool that is truly agnostic has done real work to make the same task produce the same output on GPT, Claude, and Gemini. That is expensive engineering, and almost nobody has done it well. What "model-agnostic" almost always means in practice is that the vendor picked one model, prompted around it, and would rather not commit to the specifics in front of a buyer.
Follow-up question: "Which specific model runs my workflow today? What is your fallback if that model changes? Have you tested the workflow on the fallback?"
What to do: If they cannot name the current model and the fallback, and confirm both were tested, treat "model-agnostic" as marketing copy. Real answers about model choice are worth thousands of dollars a year at scale. Vague answers mean you will be surprised by the bill and by the quality changes.
#2. "It just works out of the box"
What they mean: "Setup is easy in the demo. Setup on your actual data is a professional services engagement we bill separately."
The demo runs on a sanitized dataset the vendor built for the demo. Your data is messier. Every real integration involves at least a week of "getting the fields right", another week of "handling the edge cases", and usually a third week of "training a human on when to override the agent". None of that is "out of the box".
Follow-up question: "How many hours of setup does a typical customer at my volume actually need before the agent is producing useful output? What did the last three customers spend, in hours and dollars?"
What to do: Never sign an "out of the box" pitch without a fixed-price setup fee. If the vendor cannot give you one, walk. And multiply their estimate by 2 in your own head. If they say "3 hours" they mean "12". If they say "12 hours" they mean the whole first month.
#3. "You can customize it however you want"
What they mean: "The base product does not do what you need, and you will pay us to build it, or you will build it yourself against a limited API."
"Customize" is one of the most-abused verbs in software sales. In an AI agent context it usually means one of three things: adjusting a system prompt, mapping fields to your CRM, or writing code against a webhook. Each of these has real limits. If your specific need falls outside those limits, the answer is a paid professional-services quote you did not budget for.
Follow-up question: "Show me the exact interface where I customize this specific behavior. What have your customers customized that ended up needing a paid engineering ticket instead?"
What to do: Ask to see the customization interface live, not in a slide deck. If it is a system prompt textbox, that is a light customization. If it is a Zapier integration, that is a medium one. If it is "our team writes custom code for you", that is a $10,000 to $50,000 engineering engagement the vendor will happily quote after you sign the base contract.
#4. "Our customers see 10x improvement in their workflows"
What they mean: "One customer in a case study we cherry-picked saw a good result. Your mileage will vary."
10x is a suspicious number in almost every context. It usually comes from comparing the vendor's software to the worst possible baseline (a human doing it manually, badly, in the wrong tool) rather than a reasonable baseline (a human doing it in a normal tool with some templates in place).
Real gains are usually in the 1.3x to 2x range for tasks that were already reasonably efficient, and 3x to 5x for tasks that were done poorly to begin with. 10x is almost always a marketing number.
Follow-up question: "Show me the specific customer, the specific workflow, and how they measured the 10x. What did their previous baseline look like? Was their previous process actually working?"
What to do: Ask for a reference customer at your size and industry, not the case study. If the vendor cannot introduce you to a real reference call, "10x" is a slide, not a promise.
#5. "We take security very seriously"
What they mean: "We have SOC 2 and a security page. Please do not ask specific questions."
Security-seriously language is a tell. Vendors who actually think about security walk you through how they handle each specific concern. Vendors who "take it seriously" hand you a compliance badge and move on.
For an AI agent, the specific concerns are: what data does the agent read, where is it stored, who has access, is the data used to train models (yours or the provider's), and what happens if you cancel. If they cannot answer all five in one call, they are not ready for a business that has any customer data with sensitivity to it.
Follow-up question: "Walk me through what happens to a customer email that comes into my inbox and gets processed by your agent. Where does it live at each step? Who sees it? Is any of it used for model training? What happens to it 30 days after I cancel?"
What to do: Any hesitation on any of those five is a signal. Real answers are boring and specific. "The email is processed in our AWS us-east-1 region, encrypted at rest with KMS, retained for 90 days for debugging then purged, our support team can access it with your explicit permission, we do not use customer data for training, and on cancellation we delete within 30 days." That is what a real answer looks like. If you get "we take security very seriously" instead, they are not ready for you.
#The pattern
All five sentences work the same way. The vendor uses a positive-sounding phrase that appears to answer the question. The phrase does not commit to anything specific. If you accept it, the sale moves forward. If you ask a follow-up, you get a real answer or you get a squirm.
The good news: the follow-up questions are cheap. They cost you 30 seconds of sales-call time. They will save you a lot of money in month three.
#Two related reads
If you have been running these decoder questions and finding a lot of squirms, you are probably in the situation where the honest answer is to slow down. Two related pieces on when to hold off:
- When You Should NOT Hire An AI Agent For Your Business — the seven situations where the answer is no
- How To Tell If An AI Agent Will Actually Save Your Business Money — the buy-decision math you should run before you sign anything
#Where we come down
We would rather lose the deal than have a client sign for something we cannot actually deliver. When we sell an agent build, we tell you the model, the setup cost, the customization limits, the realistic gain range, and exactly how your data flows through the system. That is the least a buyer should expect from any vendor, and it is the fastest way to know whether the vendor across the table is worth the meeting.
If you want a second opinion on a specific vendor quote or pitch you are evaluating, come talk to us at /get-started. We will read the proposal and tell you which of these five sentences show up in it, and whether the follow-up answers hold up. No sales conversation on our end unless it makes sense for you.
