Build it: a 100-point model you can ship this week
You do not need a data-science team to ship a working model. You need a defensible point allocation, a recency rule, documented reasoning, and the humility to treat version one as a hypothesis. All of it fits in an afternoon, and a deterministic rules model you can read beats a black box you cannot trust, especially when an agent is acting on its output unsupervised.
A split that works
A practical hundred-point B2B model allocates roughly twenty-five points to demographic fit, twenty-five to firmographic fit, forty to behavioural engagement, and reserves ten points of possible deductions for negative signals. Notice the shape: behavioural engagement carries the largest single weight because it is your live intent signal, while fit is split across two buckets so neither the person nor the company can carry the match alone. The deductions are deliberately capped, a floor on how much damage one negative does, so a single penalty does not nuke an otherwise strong lead, while the stack of them still removes the genuine non-fits. Map your two axes onto this: the fifty fit points feed the gate, the forty behavioural points feed the timer, and you carry both numbers forward rather than summing them into one.
Weight recency, or measure the past
Intent that ignores time is not intent, it is history. An action taken in the last seven days should be worth roughly three times an action taken thirty days ago. Without this decay, a lead who was hot last quarter and has gone silent keeps a high behavioural score, and your model quietly routes today's attention to last season's interest. Recency weighting is what keeps the timer pointing at now. Build it in from the start, because an intent score that does not decay is a clock with no hands.
Document every point, or you cannot debug it
Write down the rationale for every value. Why is a pricing-page visit worth what it is worth? Why does a free-mail domain cost fifteen and not twenty-five? An undocumented model is unfalsifiable: when conversion lift disappoints, you have no way to reason about which weight was wrong, so you tweak blindly and call it intuition. A documented model is a debuggable one. The point table is not the deliverable; the reasoning behind it is.
Plan version two before you launch version one
The first model is a hypothesis about which signals predict conversion, not a finished asset, and you should hold it that loosely from day one. Decide in advance what evidence would make you change a weight, what you will measure, and when you will revisit it. Launching version one while already expecting version two is not indecision, it is the correct posture toward a model that lives downstream of a buyer behaviour that will not sit still. Ship it this week, but ship it knowing it is a draft you intend to beat.