An AI system that got better
because we took AI out

They wanted a sales system "with AI" — like everyone does

A car dealer group needed a system to manage its sales pipeline. The initial requirements specified cloud AI driving 14 features — from lead scoring and customer insights to sales coaching.

But one condition mattered more than the technology: sales managers had to be able to explain how the system makes decisions — why this lead is Hot and why that one isn't.

We audited all 14 features — 12 could be solved with pure logic

Going through them one by one, we found that lead scoring works with rules that have clear weights · drafting messages to customers works with templates · a 360° customer profile can be assembled from facts already in the system · sales coaching can be answered from a knowledge base.

Only 2 features truly needed language understanding — and both had low returns.

"AI" in the requirements document had become a label for the solution, not a need of the problem.

LLM flexibility vs accountability for decisions

LLMs understand language well, but they come with hidden costs: black-box decisions that can't be explained · a cost per request · customer data that has to leave the server · and vendor lock-in.

Rule-based logic, on the other hand, is explainable and exactly repeatable — but less flexible, and the business has to maintain the rules itself.

The real question isn't "is AI good enough" but "who should own this business's decision logic?"

Glass-box Decision Engine

We decided that every decision in the system — lead scoring · customer profile · insights — is computed from rules that are written down explicitly, versioned and auditable after the fact, not from LLM inference.

• Intentional — the team owns the formula and can explain why each weight is what it is
• Versioned — every ruleset carries a policy version
• Traceable — every result is recorded with its version + input + score breakdown — every decision can be replayed

LLM observes. Code decides. — let AI do what it's good at (reading, understanding, observing) · but decisions belong to code that the business owns.

What we actually built

A multi-tenant system with Row-Level Security at the database layer · all decision logic as pure functions with versioned rulesets · every scoring recorded as an auditable decision log · running on infrastructure the business controls itself — customer data goes nowhere.

In production — and every decision explains itself

The system passed multi-role end-to-end testing (a sales coordinator following up Hot customers · a manager viewing pipeline analytics) with no critical bugs, and the core decision pipeline has been validated in real production.

More importantly: when a manager asks "why is this lead Hot?" — the system answers with a score breakdown, not "the AI said so."

If logic can do it, don't use AI

Generative AI is good at understanding unstructured things — but it shouldn't own important decisions. Trading language flexibility for explainability + repeatability + ownership of the logic is always worth it for SMEs and for markets that demand accountability.

And the burden of maintaining rules isn't a cost — it's a feature: the business owns its own logic