Negative scoring is where the model earns its keep

Here is the dirty secret of most B2B scoring models: they are lovingly tuned on the upside and almost blind on the downside. Builders pour care into weighting every positive signal, a demo request worth this many points, a pricing visit worth that many, and then wave through the obvious non-fits with a shrug. The result is a model that is generous to everyone, including the people who will never buy, and a generous model is a leaky one.

Penalise the obvious, concretely

Negative scoring is not a vibe, it is a set of specific deductions you can ship today. A personal email domain, the free-mail address that signals a tyre-kicker rather than a buyer at a real company, should cost fifteen to twenty-five points. A company-size mismatch, an account well outside the range you can actually serve, should cost twenty to forty. A job-seeker signal, the candidate researching your company rather than evaluating your product, should cost around fifteen. These are not arbitrary; they are the difference between an MQL pool you trust and one quietly stuffed with people you will waste follow-ups on.

Prefer a hard disqualify to a deduction

For a true non-fit, a deduction is too soft. Take it off the table entirely. A role mismatch, someone whose job could never sign or influence a purchase of what you sell, is a hard disqualify, or thirty points off if you must keep a numeric trail. The principle: a deduction merely lowers a lead's rank, leaving it in the pool to resurface the moment a few positive signals stack up, whereas a hard disqualify removes it from the MQL pool altogether. For people who can never buy, removal is cleaner than ranking. You do not want the intern who downloaded a guide climbing back into your follow-up queue because they happened to open three emails.

Encode disqualification as deterministic rules

This is where an AI-run funnel either earns its design or betrays it. The disqualification layer must be deterministic rules, not probabilistic guesses, because the entire point is that the agent never wastes a personalised follow-up on a non-buyer. A free-mail domain plus a sub-threshold company plus a non-buying role is not a low score to be weighed against engagement, it is a closed door. Write it as an explicit rule the agent checks first, before any positive signal is even considered, and the agent stops burning your scarcest resource on people who were never going to convert. The negatives are not the unglamorous afterthought of the model. They are where it stops leaking, and a model that does not leak is a model that earns its keep.