Lead scoring without a training pipeline

Every B2B company eventually builds lead scoring, and most of them build it twice. The first version is a set of CRM rules (ten points for a director title, five for a company over 500 people) that rots within a quarter. The second is a machine-learning model, which needs a data scientist, a training pipeline, months of history and a retraining schedule, and still can't tell a sales rep why a lead scored 63.

We've been running a third approach in production for an enterprise SaaS company. It grades every demo request A to D, explains each grade, and changes when someone edits a paragraph of text.

Grades instead of scores

A score of 63 means nothing to a salesperson. A grade with a definition does:

A: Heads of the buying function at companies with 1,000 or more employees, or its directors and VPs at any business.

Each grade is a short, plain-language standard describing who the lead is: role, organisation and how they found you. A rep can read it, a marketer can argue with it, and a VP can change it.

Where the standards come from

Nobody invents the standards in a workshop. They're drafted from your own history:

1. Snapshot every past lead as it looked when it arrived, before sales enriched the record. Otherwise the model learns from information it won't have for new leads.
2. Profile each lead into a few comparable traits: function, seniority, company size band, lead source, country tier.
3. Measure what converted. For the last year of demo requests, which combinations of traits became sales-qualified opportunities, and how often?
4. Group combinations with similar conversion rates into grades and write each grade down as a sentence.

In the deployment we run, each grade converts several times better than the one below it. Your numbers will differ. What matters is the shape: a few grades, clearly separated.

A decision model applies the standard

For every new lead, a decision model reads the lead's facts and its profile and answers one typed question: which grade does this lead match? It returns a probability for every grade, so you can see when it's sure and when a lead sits between B and C.

We use Jev, TypeSafe AI's decision model, for this. It's built for classification rather than chat. The answer is a grade and a probability instead of a paragraph, and it costs a fraction of a cent per lead.

Changing the model is editing text

This changes how teams work. The standards are configuration rather than a trained model. When sales says "agency leads are never worth it, even big ones", someone edits the sentence for grade C, bumps the version, and the next lead is judged by the new definition.

There's no retraining and no feature engineering ticket. Because each version writes to its own table, you can compare grades from version 1 and version 2 on the same leads before switching.

Keep it honest with outcomes

Plain-language standards are easy to change, which means they're easy to change badly. The guardrail is the same history the standards came from: for every version, track how each grade actually converts over the following weeks. If grade A stops converting like grade A, the standard needs work.

Where it runs

On SegmentStream Cloud this is a set of ordinary pieces in one workspace repo:

• a managed HubSpot connection and ingestion of contacts and the history of their properties;
• models that snapshot each lead and measure conversion by trait;
• an evaluation connection that grades each new lead;
• an app your sales team opens, with grades, probabilities and explanations.

It runs in production on SegmentStream Cloud today; see lead scoring for how it is built and how our engineers can set it up on your own CRM.

The Lead scoring page shows how it fits with your CRM and the rest of the workspace.

Want it on your CRM? Start a pilot: our engineers set up the first version in two weeks, at a fixed price. Self-serve opens to the waitlist this fall.