Resources · 3 min read
Anti-invention: an agent that never improvises in front of a customer
A model that doesn't know the answer fills the gap with a guess. Here's how to build the system that won't let it.
Why models invent
A language model is trained to continue text plausibly, not to say “I don't know”. If a customer asks about a size that isn't in the context, the plausible thing is to answer a size — so it does. It's not a defect you switch off with a setting: it's the nature of the tool, which is why the control can't live in the model. It lives in the system around it.
Layer 1 — Instructions aren't enough, but they matter
Telling the agent “don't invent prices” reduces the problem; it doesn't eliminate it. Instructions define default behavior: what tone it uses, when it asks instead of asserting, and when it escalates to a person. They're necessary and they're the weakest layer — everything that truly matters gets reinforced in the other two.
Layer 2 — Data comes from your base, not the model's memory
Every answer about price, stock or features is built from a query to the business's system on that turn — the model writes over retrieved data, it doesn't remember. If the query doesn't return the fact, the agent has nowhere to assert it from. This is the heart of grounding: turning “answering” into “querying and writing”.
Layer 3 — Output validation and escalation
Before sending, the answer is validated against hard rules: does it mention a price that didn't come from the base? Promise a discount that doesn't exist? Touch a forbidden topic? What doesn't pass doesn't go out — it gets reformulated or escalated to a person with full context. The customer sees a correct answer or a person; never a guess.
How you verify it works
With a suite of real scenarios: trick questions, sold-out products, sizes that don't exist, customers who insist. It runs before every change — of prompt, catalog or model — and grows with each month's real conversations. Anti-invention isn't an architecture promise: it's a measured result.
Rather go through it on your own case?
A 15-minute audit of your real operation. You leave with an honest read on what to automate first, and what to leave alone.
Book a 15-min audit