When AI ships the impossible

I watched a team ask a generic AI assistant to help with a rush quote. The bot pulled from PDFs and past proposals and returned a polished, confident configuration. It looked beautiful. It was also impossible to build.

The spec paired a high-torque motor with a gearbox ratio we don’t make, used a voltage range that fails regional compliance, and extended an axis beyond the enclosure’s travel. Sales sent it anyway. Engineering caught it late. Cue the fire-drill, the red faces, and the awkward call to the customer.

I’ve done CPQ work since 2000, a lot of it in Tacton CPQ. I’ve seen smart teams try to paper over product complexity with a clever UI, a few guardrails, and now a chat assistant. The pattern is the same every time: confident output, brittle reality. Nice decks don’t move metal.

The real issue isn’t AI, it’s your product reality

The common plan right now sounds like this: “We’ll train the AI on our docs, past quotes, and CRM. Then we’ll watch quoting speed up.” It feels sensible. It’s also how you create workslop - fluent, confident nonsense at scale.

The reframe: your risk isn’t the AI model. It’s the absence of a single, unified, logical representation of your product reality for anything to reason with. If you feed scattered PDFs, tribal rules, and half-remembered constraints into an AI, you won’t get insight. You’ll get chaos with a smile.

A world model solves that. Not as another pile of if-then rules, but as a holistic description of how your product actually works. Think of it as your product’s single source of truth that both humans and AI can query and trust. It’s not a new buzzword. It’s the antidote to made-up answers.

CPQ is not about automation - it’s about correctness.

When you treat configuration as a world to simulate rather than a checklist to validate, the whole system shifts. Instead of reacting to bad choices, you only allow valid choices. Instead of explaining away exceptions, you make them explicit and testable.

AI does not replace logic - it depends on it.

Here’s the mechanism in plain terms. A world model represents parts, attributes, and constraints as a complete system. Compatibility, physics, compliance, pricing logic, and commercial policies live in one place, expressed once. The system doesn’t wait for a user to make a mistake so it can say no. It only proposes routes that lead to a buildable, priceable solution.

Use a GPS metaphor. A good GPS doesn’t just beep when you take a wrong turn. It calculates the route that avoids dead ends from the start. Your configuration logic should do the same. When talking about systems that only uses symbolic logic for configuration, the same principle applies. Symbols still describe a world. The job is to make that world complete enough that no path leads to nonsense.

If you want an AI assistant to be useful, give it this GPS. Let it ask the world model what is valid, rather than guessing from documents. The AI becomes a fast talker sitting on top of hard rules, not a storyteller inventing roads.

If the system cannot explain itself, it will never be trusted.

I hear one counterargument: “Our best reps already avoid the bad paths.” Yes. They carry the world model in their head. But they still burn hours on verification, and when they move on, your accuracy walks with them. Systematize their thinking or keep paying the turnover tax.

Make AI safe: practical rules and next steps

Rules that keep AI honest

  • Model the product’s reality once, use it everywhere. One canonical representation for configuration, pricing, and documentation. No forks. In Tacton CPQ or your chosen platform, link the same structure to all touchpoints. Example: the same constraint that drives selection also drives drawing generation and price blocks.
  • Proactive configuration beats reactive validation. Don’t show options you know will fail. If a motor-gearbox pair is impossible, it never appears. Example: drive attribute compatibility to narrow choices before the user clicks, not after.
  • Name and remove the PDF soup anti-pattern. Training AI on scattered specs and slide decks is how you get workslop. If a rule exists only as a sentence in a doc, it doesn’t exist in your system. Extract it, formalize it, test it.
  • Every rule must be explainable and testable. If you cannot show why a configuration is valid in one sentence and one test, it will rot. Build a small, automated test suite for your core constraints. Run it on every change.
  • Price logic follows product logic, not the other way around. Don’t hack the configuration to fit discount targets. Structure the product truth first, then layer commercial policy. Example: calculate cost and capability from the configuration, then apply price waterfalls.
  • Kill the Rule Graveyard. Duplicated, overlapping conditions are the fastest way to brittleness. Centralize shared logic and eliminate copies. If you need a fork, make it a parameter, not a paste.
Adoption is the only metric that matters.

Immediate actions you can start this quarter

  • Inventory the truth and declare one authority. List every place a product rule lives today: CAD notes, ERP fields, tribal Slack, old Excel. For each rule, pick one system of record. Everything else becomes a reference, not a source.
  • Build a thin, correct slice of the world. Don’t boil the ocean. Pick one high-volume product family. Model the full configuration constraints end to end. Wire the AI assistant to ask this model for valid options via an API, not to summarize documents. Ship it to a pilot team.
  • Expose the why in the UI. When the system makes a choice or hides an option, show the reason. Sentence-level explainability builds trust and speeds onboarding. This is where Tacton CPQ’s attribute-level constraints and explanation texts shine, but any platform can do it if you design for it.
  • Set up change governance like you mean it. Define owners, a change path, and a fast test cycle. If a rule change still needs a project plan, sales will route around you. Weekly updates, small diffs, and green tests or it doesn’t go live.
  • Measure correctness, not just speed. Track first-time-right quotes, engineering escalations, and cycle time to valid. Speed without correctness is just a louder alarm.

I remember a multinational rollout where the first instinct was to fine-tune the AI with hundreds of past quotes. We paused and modeled a single, clean product slice instead. Within two weeks, reps stopped asking engineering about that family. The AI chat stayed, but its job changed: gather requirements, draft texts, and call the configuration service for facts. Complaints dropped. Cycle time did too.

If you take nothing else from this, take the order of operations. Human judgment defines intent and tradeoffs. Explicit logic encodes the product reality. AI accelerates interaction on top of that. Remove any one piece and the stool tips.

The future of CPQ is hybrid intelligence.

One last point on tech choices. You don’t need a new acronym to start. Use what you have if it can represent constraints as a whole, not as scattered validations. The strongest setups I see treat the configuration engine as the world, and everything else consumes it. That includes your shiny AI.

An AI can’t reason about a world it doesn’t understand. A world model is how you teach it.