Think about what a new sales hire gets in week one. A product walkthrough. The deck. A pricing sheet with the discount thresholds. The battle cards for the three competitors that come up. A list of the objections everyone hears and what has actually worked against them. Recordings of calls that went well. A definition of who you sell to, and, more usefully, who you do not.
Now think about what most AI sales tools get. A field labelled “tone of voice”, three hundred characters, into which someone has typed friendly but professional.
We would not let a human answer a prospect after that briefing. It is not obvious why we expect better from a model.
What the text box can and cannot encode
The box does one thing well: register. It can make output shorter, warmer, less exclamatory. That is genuinely worth having, and it is why the first draft from these tools often reads fine at a glance.
What it cannot encode is everything that makes a reply correct rather than merely well-mannered.
- How the sentence sounds
- A company name and a one-line description
- Maybe a value proposition, compressed past the point of meaning
- Whether the thing they asked about exists yet
- What it costs, and what the discount floor is
- How you differ from the competitor they just named
- The objection they raised, and the answer that has worked before
- Whether they are even a fit, and how to say so gracefully if not
Given only the first column, a model does what anyone would do when asked a question they cannot answer: it produces something plausible. That is where the invented feature comes from, and the confidently wrong price, and the cheerful agreement to an integration that does not exist. Not a flaw in the model, a briefing problem.
What changes with real documents
Give it the same library a new hire gets and the failure mode changes shape. Instead of inventing, it answers from the document, and where the document is silent, that silence is visible and can be handled, rather than filled in.
The practical difference shows up most in three places. Pricing questions get the real number and the real caveat. Competitor questions get your actual positioning instead of a generic differentiator. And disqualification gets handled properly, a prospect who is not a fit gets told so, politely and early, rather than being enrolled in four more touches.
A knowledge base is not a dumping ground. The documents that help are the ones a person would have been given: current pricing, the live battle cards, objection handlers that reflect what works now. Feeding it a folder of outdated collateral produces an AI that is confidently out of date, which is harder to catch than one that is obviously uninformed.
Where a human still signs off
Better briefing raises the floor; it does not remove the need for judgement. Sends stay gated for human approval, and it is worth keeping them there while you read the drafts, not because the drafts are bad, but because reading them is how you find the gap in the documents. Every reply that comes out wrong is pointing at something missing from the library.
In practice teams run in draft mode for a few weeks, fix what the drafts expose, and only then loosen the reins, and usually only for the categories they have watched most: standard qualifying questions, scheduling, follow-ups on things already discussed.
The test
Take the hardest question a prospect asked your team this month. Ask your AI tool. If the answer is fluent and wrong, you do not have a model problem. You have a tool that was never given anything to be right about.
