# Cognitive Routing > Utility routing optimises for price. Cognitive Routing optimises for depth. How Pilot5 measures 20+ AI models by domain, and routes on that evidence URL: https://pilot5.ai/blog/cognitive-routing Category: positioning Published: 2026-09-10 Read time: 5 min Keywords: Cognitive Routing, AI model routing, LLM routing, The Expert, model selection, multi-model deliberation, AI benchmarking, defensible AI answers --- ## Everyone is optimising routing for price. We optimised it for depth. A while back we noticed something uncomfortable about how we were using AI. We were spending more time verifying answers than using them. One model, one confident answer, and no way to know what it was built on, or whether a different model would have said something better. The output was fluent. That was the problem. Fluency is not evidence. The industry’s response to this has been routing. Send each query to the best model available, at the lowest cost that will do the job. That is useful engineering, and it solves a real problem. But it solves the wrong half of the problem. Utility routing optimises for price. It does nothing for the depth of the answer, and nothing for whether you can defend that answer to anyone else. So we built the other direction. We call it **Cognitive Routing**, and it is live today inside The Expert. ## Utility routing and Cognitive Routing ask different questions Utility routing asks: which model is cheapest for this query? Cognitive Routing asks: how do we produce the most defensible answer to this question? Different question. Different category. Everything below follows from that one difference. ## What happens when you ask The Expert a question **1. Your question becomes a structured brief.** It is not sent raw. The Companion reformulates it into a ten-point brief: the decision, the situation, the constraints, what a good outcome looks like. Scoped, framed and made actionable before anything runs. Most disappointing AI output is not a reasoning failure, it is a scoping failure. A model that answers a question nobody was asking will do it very convincingly. **2. The question is documented before it is answered.** Live web research runs across several engines in parallel, merged and de-duplicated, with dead or invented links dropped. Verified institutional sources are pulled in, from 400+ sources across 21 knowledge domains. Your own documents are read alongside them. The model does not receive an isolated question. It receives a question that is framed, documented and sourced. **3. You choose the cognitive lens.** Architect, Strategist, Engineer, Counsel or Contrarian. Structure, horizon, feasibility, risk, dissent. You are not only picking the best model for the question. You are picking the perspective it thinks through, which changes what it looks for and what it flags. **4. The model is selected on evidence.** Not the default. Not the cheapest. The one that has demonstrably performed best on your class of question. That last point is the part worth explaining, because it is the part nobody else can copy quickly. ## How we know which model is strongest Here is where the deliberation engine stops being only a product feature and becomes an instrument. Every time Pilot5 runs a full deliberation, five models analyse the same question independently, without seeing each other, then critique each other’s work across several rounds. Same brief. Same sources. Same constraints. Different models. That is a controlled comparison, run at scale, on real questions that real people actually needed answered. **Pilot5 measures the comparative performance of more than twenty AI models this way**, and it measures them where it matters: by knowledge domain and by question type. Not one global leaderboard, because a global leaderboard hides the thing you need. The model that reasons best about a contract clause is not the one that reasons best about a supply chain trade-off, a pricing structure, a regulatory exposure or a systems architecture. Those are different skills, and the gaps between models are wide and inconsistent. Public benchmarks cannot tell you this. They measure standardised tasks, they are published in advance, and they are optimised against. Our measurements come from adversarial comparison on live questions, where the models were not competing for a score and had no idea they were being compared. So when The Expert routes your question to one model, that choice rests on evidence gathered from deliberations of the same class. **Routed on evidence, never on price.** And a further architectural rule holds throughout: within a panel, never two models from the same provider. Provider diversity is a structural guarantee against correlated blind spots, not a marketing line. The loop closes on itself. Every deliberation improves the routing. Better routing produces better answers. Better answers bring more deliberations. ## What comes back One answer, with its full provenance. Every claim tagged as sourced or inferred, so you always know what is grounded in evidence and what is the model’s reasoning on top of it. A confidence score. The evidence shown, not summarised away. If the answer is not enough, escalate to the full panel in one click. Five models, five roles, the recommendation and the strongest case against it, with the dissent preserved in full. ## Why this matters The value of an AI answer is not how quickly it arrives or how little it costs. It is whether you can act on it, and whether you can defend it to a board, a client, a regulator or a co-founder who disagrees with you. That requires knowing what the answer was built on. It requires knowing why that model, and not another. And it requires someone to have actually measured the difference instead of assuming it. **The quality of the answer starts before the AI.** Cognitive Routing is live today, inside The Expert. **Disagree to decide.** Try it at [Pilot5.ai](https://pilot5.ai). Free credits at signup.