Lead scoring without ML: why a five-rule scorecard wins for most teams

Every quarter another sales team tells me they need AI lead scoring. We dig in, and what they actually need is a clean five-rule scorecard with negative weights and a decay function. The ML model is the headline, but the lift is in the basics that almost nobody bothered to set up first.

This is not a contrarian take. It is the same advice you will find buried at the bottom of every honest ML guide. The problem is that the rules-based version sounds boring, so teams skip it and pay for a model that learns the wrong thing from dirty data.

The math problem most teams have first

A predictive model needs enough closed-won and closed-lost examples to learn from. Most B2B teams do not have that volume yet. According to the Prospeo guide on AI versus traditional lead scoring, the practical floor is roughly 1,000 leads per year, 100 closed deals, and 12 to 24 months of clean CRM data before an ML model produces a defensible score. Below that, a well-built rule-based model will out-predict a poorly trained AI model every single time.

That is the part the vendor decks leave out. A model trained on 60 conversions and a year of inconsistent stage definitions does not learn your buyer, it learns your data hygiene problem.

Why the rules version usually wins

Forrester has been blunt about this for years. The Forrester piece on what lead scoring actually is calls out that most production scoring models are built on guesses and random estimates of propensity to buy, not on any real analysis. Swapping that for an ML model trained on the same shaky inputs does not fix the inputs, it just makes the score harder to argue with.

A simple scorecard is auditable, fast to change, and forces the sales and marketing teams to agree on what "qualified" means before they ship it. That alignment is where most of the real lift comes from. The Clueless Company breakdown of failing lead scoring models puts it well: the most common reason scoring projects fail is not bad logic, it is that sales does not trust the score, so reps keep working whoever they were working before.

You cannot build trust in a black box. You can build trust in five rules you can read off a whiteboard.

The five-rule scorecard

Here is the version we recommend for any LeadGrid customer below the ML threshold. It fits in one screen and explains itself.

fit_score:
  icp_company_size:    { match: +20, miss: -15 }
  icp_industry:        { match: +15, miss: -10 }
  decision_maker_role: { match: +20, miss: -10 }
 
intent_score:
  pricing_page_visit:  +20
  demo_request:        +30
  three_emails_opened: +5
  inactivity_per_week: -5
 
threshold_for_sales_handoff: 50

Five rules, two negative weights, one decay rule. That is it. The structure is borrowed from the Reform guide on common lead scoring mistakes, which is direct about the two failure modes a basic scorecard has to fix: tracking too many signals creates noise, and only assigning positive scores floods the CRM with cold leads that nobody removes.

The decay rule matters more than people realise. The Breadcrumbs B2B lead scoring framework for 2026 recommends subtracting points for each week of inactivity, exactly so zombie MQLs do not sit at the top of the queue while a hotter lead from this morning sits below them. Without decay, your scorecard is a high score table, not a queue.

When to graduate to ML

There is a real threshold where ML scoring earns its complexity. You have crossed it when three things are true at once.

You have at least 100 clean closed-won and 100 clean closed-lost records, with consistent stage definitions across that history. The Landbase lead scoring statistics roundup cites Forrester data showing 38% higher conversion and 28% shorter sales cycles for teams running AI scoring, but those numbers come from teams that had the data hygiene to train on. Without that hygiene, the model is learning your bad CRM, not your buyer.

Your sales cycle is complex enough that humans cannot hold the variables in their head. Multi-stakeholder enterprise deals with 18 plus touchpoints across six months are a different problem than a self-serve SaaS funnel with a free trial gate. ML earns its keep on the complex shape, not on the simple one.

You have already shipped the rules-based version and you can name exactly which rule the ML model is going to beat. If you cannot name it, the ML model is not solving a problem you have, it is solving one a vendor has.

The honest hybrid

The pattern that works is rules-based scoring as the base layer, ML as an overlay on top once you have the data to justify it. The rules give you a score reps trust on day one. The ML overlay, once you ship it, adjusts the score using patterns the rules cannot see, but the human-readable score stays in the UI.

This is also what the SiriusDecisions data on lead scoring adoption was pointing at when it found that 68% of companies were running lead scoring but only 40% of salespeople saw any value in it. The teams in the 40% had a model the reps could explain. The teams in the other 60% had a model the reps did not believe.

Build the five-rule version first. Earn the ML upgrade once your data and your sales team are ready for it. Most teams never need to take that second step, and that is fine.

Start free →