Apply Simple Rules First

GTM Lab's first attempt at the qualification decision sends the whole record to a model and asks it to decide.

The model is good. It reads the message, weighs the company, considers the industry, and returns a well-argued answer. It also, roughly one time in twenty, reasons its way past a criterion that was never negotiable. It decides that a company below GTM Lab's size threshold is worth routing anyway, because the use case is such a strong fit and the enthusiasm in the message is obvious.

Every one of those is defensible in isolation, and every one of them is wrong, because the threshold was not a suggestion the system was invited to weigh. It was a rule.

Rules are better than judgment at the things rules are good at

A deterministic rule has properties a model cannot offer, and they are worth naming because they sound boring and are decisive. It returns the same answer every time. It costs nothing to run. It can be tested exhaustively. It can be explained by pointing at it. And when it is wrong, it is wrong in a way you can find, because a threshold that is set badly is visible in the definition rather than distributed across a thousand judgments.

That last property is why the engagement score in Chapter 1 is deliberately built without a model. Every value it produces gets read in a weekly leadership meeting, and the ability to say “this deal scores 61 because the last meeting was 64 days ago” is worth more than any accuracy a model might add. A number you cannot defend is not usable in the meeting where it matters, no matter how good it is.

Do the cheap, certain work first

Order the decision so deterministic rules run before anything else, for three reasons. Most records never need the expensive step: a vendor solicitation from a known domain is settled by a list lookup. The model gets a smaller, cleaner question: instead of “decide everything about this record,” it is asked one thing it is actually suited to. And the non-negotiable criteria stay non-negotiable, because they are evaluated by something that cannot be talked out of them.

A short-circuit is a feature, not a shortcut

Some conditions are decisive on their own. When one is true, the decision is made and nothing else is evaluated. GTM Lab's engagement score has one: if a meeting is on the calendar for a future date, the score is the maximum and no other input is considered. Not weighted heavily. Not a large bonus. The whole answer.

This looks crude next to a blended formula, and it is better. A deal with the next conversation booked is not a stale deal, whatever its stage date says, and a formula that lets three negative terms drag such a deal into the danger zone is producing a number that contradicts the single most informative fact on the record. Short-circuits also make a system legible. “It scored maximum because a meeting is booked” is an explanation a rep can act on. A composite that arrives at a similar number through four offsetting terms is not.

The cost is real and worth stating: a short-circuit can be gamed. A rep who knows a booked meeting produces a maximum score can book a meeting. That is a management problem rather than a formula problem, and the answer is that the score is one input to a conversation rather than the conversation.

Comparing three designs on the same decision

The engagement score. Deterministic wins outright. The inputs are dates, the operation is arithmetic, there is nothing to interpret, and explainability is the whole requirement. A model here would cost money, add latency, introduce variance between runs, and produce a number nobody could defend. It buys nothing.

Qualification, model only. Fails as described in the opening. The deciding evidence includes a hard threshold, and a model asked to weigh a hard threshold will eventually weigh it.

Qualification, rules first with a model on the interpretive step. Wins, and the next chapter is about where exactly that line falls.

The lesson is not that the combination is best. It is that the engagement score and the qualification decision have different evidence, and the evidence chooses the design.

Carry this into your business

Find a decision at your company that goes straight to a model and list the criteria that are genuinely non-negotiable. Move those in front of the model. It is usually an afternoon, and it removes an entire category of confident wrong answers.