Lead scoring models in real-time journeys

What it does

A scoring model adds up points against a lead record based on the lead's own attributes and its marketing interactions, and writes a number back that you can grade, segment on, and qualify against. It's the real-time journeys replacement for the old outbound lead scoring models, rebuilt on the interactions store rather than the legacy marketing insights database.

Key facts

Turned on at Settings > Overview > Feature switches > Lead Management. That one switch lights up both Scoring models and Qualification under the Lead management area. Nobody sees the menu until it's flipped.
Conditions come in two flavours: Attributes (demographic/firmographic fields from the lead, or from related tables such as contact and account) and Interactions (email opened, event registration, website visit, and so on).
Interaction conditions are built as volume plus window plus points — "at least 5 emails opened in the last 28 days, +10" — and can be narrowed to a specific asset using interaction metadata.
Only standard Dynamics 365 interactions are scoreable. Custom interactions you push into the interaction store are not available as scoring conditions.
A model is capped at 40 groups of conditions. That's generous for a sane model and a hard stop for the ones clients design in a spreadsheet.
Scores recalculate on a background cycle, roughly every two hours. This is not a real-time trigger-driven score, whatever the module name suggests.
A model has two states: Draft (editable, scoring nothing) and published/live (scoring, conditions read-only). You can view the logic after publishing but not edit it.
Fields with field-level security enabled cannot be used in scoring conditions at all, regardless of the modeller's security role.
Models are business unit scoped — a model scores only leads belonging to the business unit it sits in.
Two out-of-box roles govern access: Lead Score Viewer (read and visualise) and Lead Score Modeler (create, edit, publish). Neither is granted by default to standard marketing users.
Model settings include Count parent contact interactions, which rolls the parent contact's engagement into the lead's score. Off by default.

When to use / skip

If the client has a real sales team that works leads and complains about volume, scoring earns its place quickly — it's the cheapest way to stop sales calling everyone who downloaded a PDF. If leads go straight to a queue and get worked in arrival order regardless, scoring is decoration and you should say so.

The honest test is whether anyone will act on the number. A score that nobody uses to route, prioritise or gate a handoff is a report nobody reads. On B2C-shaped projects where the lead table is barely used and everything lives on contacts, skip it — there is no contact scoring equivalent here, and forcing lead records into existence just to score them is a bad trade.

Also skip if the client's interaction history is thin. Scoring on three weeks of email opens produces a model that ranks nothing. Give it a quarter of engagement data before you build anything with grades attached.

Configuration decisions

Model scope. Do you build one model for everything, or separate demographic-fit and engagement models that qualification later combines? Two narrow models are usually easier to defend and easier to retune than one that mixes fit and behaviour.
Point scale. Agree the arithmetic up front — whether interactions can go negative, what the practical ceiling is, and whether a single high-value action (demo request) should be able to carry a lead on its own.
Window lengths. Short windows make the score decay naturally; long ones make it a lifetime tally. There is no separate decay setting, so the window is your decay mechanism.
Parent contact rollup. Decide whether lead engagement should include the parent contact's activity. It matters a lot when forms create leads against existing contacts.
Business unit strategy. With more than one BU in scope, decide whether each BU owns a model or you consolidate. Cross-BU scoring is not a thing you can configure your way out of.
Who holds Lead Score Modeler. Publishing is a business decision with a data footprint; keep the role narrow.

Gotchas

Publishing is effectively a one-way door for that version. To change a condition you clone or unpublish and republish, and the scores recalculate from the new logic — historical scores are not preserved as a trend.
After merging duplicate leads, the model stops scoring the surviving primary record until you either delete the scores on the merged (non-primary) records or republish the model. This one bites in every environment with a dedupe process and is very quiet about it.
Field-level security is often applied late, by a security workstream that doesn't know scoring exists. A model built on a field that later gets FLS applied is a support ticket waiting to happen.
The two-hour refresh means a demo where you submit a form and expect the score to jump will disappoint. Build the demo around already-scored data.
Interaction conditions only see interactions the platform actually recorded. Emails sent from outbound (before removal), or from a third-party sender, contribute nothing.
Business unit scoping catches multi-country rollouts late — the model looks fine in the build BU and scores nobody elsewhere.

Consultant notes

Tell the client the score is a prioritisation aid, not a prediction. The moment it's described as "AI" in a steering pack you'll spend a workshop unwinding that.
Demo the Insights tab, not the model builder. Executives care about the distribution and how many leads land in each grade; they do not care about condition groups.
Push back on models with more than about a dozen conditions on day one. Start small, watch the distribution for a month, then add. You cannot edit in place, so iteration cost is real.
Before go-live, check the security roles are assigned, the feature switch is on in every environment including the one you'll do UAT in, and that the merge/dedupe runbook includes republishing the models.
Get the business to write down what score means "call this person". If they can't, the qualification criteria conversation later will stall.

Worth another look if Microsoft ships contact-level scoring or shortens the refresh cycle — both change the design advice here.