Chapter 9: Let the LLM Talk. Never Let It Decide.
A client's sales chatbot changed personality overnight.
Nobody touched the configuration. The vendor upgraded the underlying language model, and the bot that had been confidently guiding buyers turned cautious and hedging. Deals in progress noticed. That was the day the lesson landed for that team: your sales logic cannot be allowed to drift with the model of the day. Whatever the AI is doing in your sales process, the part that decides must be yours, versioned, and boring.
This chapter is the architecture chapter. If you take one picture away from this book, make it this one.
Three layers
The systems that work, and I've now seen the pattern across every successful implementation, separate three concerns that vendors love to blur.
The conversation layer is the language model. It reads the customer's rambling email, asks the follow-up question, translates "robust, but not over-engineered, and it has to cope with winter" into candidate requirements, drafts the proposal narrative in the customer's language. It's the interviewer, and it's superb at it. A constraint solver, for all its guarantees, is a terrible interviewer.
The control layer is the constraint model and the pricing waterfall. Deterministic, versioned, tested. It's the judge. Same inputs, same outputs, every time, with an explanation on demand. Nothing reaches a customer, a contract, or an ERP without passing through it.
The knowledge layer is your bottled expertise from Chapter 7, retrieved and cited. It's what lets the conversation layer speak about your products instead of products in general, with receipts instead of vibes.
The interaction is a loop: the LLM proposes, the solver disposes. Or as I've put it in front of a hundred executives now: AI collects intent. Rules certify reality. Give the language model four verbs (propose, validate, price, explain) all of which are calls into the control layer, and the demo magic becomes production-safe. The customer in Denver asks why they got the bigger engine, and the answer comes back grounded: this engine derates above 1,500 meters; Denver sits at 2,000; here's the rule and the source. Conversation that persuades, validation that protects. Stop making one pretend to be the other.
Principles, tests, and the blank page
Two practical disciplines separate teams that run this safely from teams that got lucky so far.
First, the bot needs principles, not just permissions. One medical-device company's assistant fetched the cheapest imaging configuration for a buyer. Legal in one region, flagged by compliance two weeks later. The bot wasn't wrong. It was unprincipled. Encode commercial doctrine the way you'd brief a new hire: bias toward standard modules when they meet the need within tolerance; never lead with the loophole; when refusing, refuse with a route ("that combination isn't available, here's the closest configuration that is"). Written principles also make the bot testable, which is the second discipline: keep a suite of golden quotes that must always pass and poison quotes that must always fail, and run it on every model change. That's how you find out about the personality transplant before your customers do.
Get the architecture right and the payoffs stack up quickly. The blank-page problem dies first: instead of a configurator that opens with forty empty fields (the moment where sales traditionally opens Excel), the rep gets a valid first draft seconds after the CRM notes land. Documents follow: proposals grounded in the actual configuration, in the customer's language. Then the plumbing: one client binds ERP handoff to the single safe moment when a configuration goes valid, and quote-to-order went from two days of copy-paste to two minutes. AI drafts. Rules verify. Humans decide.
Anti-pattern: The Blender. Prompts calling rules calling prompts, business logic smeared across model instructions where nobody can version or debug it. When yesterday's prompt produces a different machine today, you don't have a system. You have a performance. Don't build a blender. Build a pipeline.
The strange gift of this architecture is what falls out the bottom of it: every conversation, every proposed configuration, every accepted and rejected option becomes structured data about what your market actually wants. Which is a bigger asset than almost anyone realizes.
The architecture every executive should demand: a probabilistic interface on a deterministic core. The LLM proposes; the rules dispose.