Predictive lead scoring

What it does

Trains a model on your own closed leads and gives every open lead a score from 0 to 100, with the top influencing factors behind it. Sellers see the score in views and in a widget on the lead form; the point is ordering the queue, not predicting the future precisely.

Key facts

Configured under Sales Insights settings > Predictive models > Lead scoring. Sales Enterprise includes predictive scoring capped at 1,500 scored records per month; higher volumes need the appropriate premium entitlement.
Minimum training data is 40 qualified and 40 disqualified leads created and closed within the chosen training window. Below that, the Get started button does not appear.
The training window is selectable from three months to two years.
Model definition includes a business process flow filter, the state option set and which values count as qualified and disqualified, plus a filter column and values to restrict which leads the model applies to. Calculated columns are the trick for filtering on more than one field.
Leads that abandoned the selected business process flow are excluded from training, from scoring and from the 40/40 minimum count.
Maximum of ten models per org, published and unpublished combined. The app warns you if a new model would score the same leads as an existing one.
Optional automatic retraining every 15 days; manual retraining is available at any time.
Scores are written to the msdyn_predictivescore table, shared with opportunity scoring. They are no longer written onto the lead table itself, which matters for anyone writing views, charts or flows against them.
Publish readiness is gated on the model's AUC score against a threshold. You can publish a model that falls short, and it will score badly.
Closed leads take roughly four hours to reach the analytical store, so a lead you closed this morning is not in this morning's training run.
Custom business process flows need change tracking enabled to sync, or the model has nothing to filter on.

When to use / skip

This is worth doing when the client has a genuine volume problem — more inbound leads than the team can call in a day — and enough closed history to train on. In that situation the score is a cheap way to order the work list and the influencing factors give sellers a reason to trust it.

Skip it when lead volumes are low, when the sales team already knows every lead personally, or when the data is thin or new. Forty qualified and forty disqualified leads is the technical floor, not a sensible one: a model trained on the bare minimum will produce scores that a sales manager can pick apart in a meeting, and once trust goes it does not come back. Also skip it where the client's disqualification discipline is poor — if leads are left open forever rather than disqualified, the model learns from a biased sample and will systematically overscore.

Configuration decisions

Which business process flow scopes the model, and whether the client's real process is actually represented by it.
What counts as qualified and disqualified, particularly if the client has customised the lead status reason set.
The training window: longer gives more data, shorter reflects a business that has changed. Pick deliberately rather than accepting the default.
Whether one model covers everyone or you segment by region, product or business unit using filter columns — remembering the ten-model ceiling.
Whether to include custom and intelligent attributes in the model rather than the default standard set.
Whether automatic retraining every 15 days is appropriate, or whether the client wants a human to look at accuracy before each republish.

Gotchas

Models created before the 2020 wave 2 change apply to all leads and override newer ones. If an old model is lurking in a long-lived environment, delete it before you build anything new or your new model will appear to do nothing.
Scores living in msdyn_predictivescore rather than on the lead catches out anyone who assumes they can just sort a view by a lead column. Use the supplied widget and views, or plan the joins.
Disqualification hygiene silently poisons the training data. Check how many leads are simply left open before you promise anything about accuracy.
The 1,500 scored records per month entitlement on Sales Enterprise is easy to blow through at high inbound volume, and the conversation with procurement is better had during design.
Leads that fell out of the selected business process flow vanish from the model's world entirely, including from the minimum record count, which is why the Get started button sometimes refuses to appear in an org that visibly has hundreds of closed leads.
The four-hour sync delay means a demo where you close a lead and expect it to influence scoring will not work.

Consultant notes

Check the 40/40 minimum against the client's actual closed leads before the feature reaches a statement of work. It is the fastest way to kill an unrealistic expectation early.
Show the top influencing factors, not just the number. Sellers accept a score they can argue with and ignore one that appears from nowhere.
Agree with the sales manager what a score actually triggers — call order, routing, nothing at all. A score with no process attached gets ignored within a fortnight.
Look at the Performance tab and the AUC before publishing, and tell the client plainly if the model is not good enough yet. Publishing a weak model to hit a go-live date costs more than waiting.
Revisit accuracy a quarter after go-live, when there is real post-implementation data, and retrain rather than assuming the launch model still fits.

Worth revisiting if the scored-record entitlements change, or when scoring gets folded further into the agent-driven qualification story.