Lead score grades and threshold calibration

What it does

Grades turn a raw score into a label — Cold, Warm, Hot, or whatever the client insists on — by mapping point ranges to names on the scoring model. The model's Insights and Scores tabs then show you how the population actually distributes, which is how you pick the boundaries instead of guessing them.

Key facts

Grades are configured on the scoring model itself, alongside the conditions, and are optional. A model with no grades still produces a numeric score.
Each grade is a name plus a range expressed with less than, more than, or between operators — for example 0 to 12 points as Cold.
Grades are recalculated with the score, on the same background cycle (roughly every two hours), not on record save.
The Insights tab on a published model gives you: total leads scored, the score achieved by the top 10% of leads, the overall score distribution, and the volume of leads sitting in each grade.
The Scores tab is the row-level view — every scored lead with its score and grade, and a route through to the lead record so you can sanity-check individual calculations.
Score and grade are attributes on the lead, so both are usable in segment definitions and journey branching, not just in the scoring UI.
Grades, like conditions, are locked once the model is published. Retuning a boundary means republishing.
There is no built-in benchmark or "recommended" threshold. The top-10% figure on the Insights tab is the closest thing the product gives you to a starting point.

When to use / skip

Add grades when humans read the output. Sales users do not want a number between 0 and 80; they want three buckets and a rule about what to do with each. If the only consumer of the score is a qualification criterion or a segment filter, grades are optional overhead — the threshold does the work and the label adds nothing.

Where grades genuinely pay off is the handover conversation. "Everything graded Hot gets a call within a day" is a commitment a sales manager can agree to. "Everything over 47 points" is not, because nobody has any intuition about 47.

Skip elaborate grade schemes. Five grades means five service-level promises the sales team will not keep. Three is usually the ceiling, and two — sales-ready or not — is a perfectly respectable answer on a first release.

Configuration decisions

Number of grades and their names. Agree these with sales, not marketing. The labels end up in views and dashboards and are painful to rename later because the model has to be republished.
Where the boundaries sit. Decide whether you set them on business logic (any demo request is Hot) or on distribution (top decile is Hot). Distribution-based boundaries age better but need a month of real scores first.
Whether grade or score drives qualification. Qualification criteria are defined against a score threshold, so if the business talks in grades you'll need to keep the two aligned by hand.
Coverage of the range. Decide what happens to leads that fall between two grades or below the lowest one — leave them ungraded deliberately, or make the bottom grade open-ended.
Retune cadence. Set an owner and a rhythm — quarterly is typical — because boundaries drift as campaign volume changes.

Gotchas

Boundaries chosen before go-live are almost always wrong. With no history, the whole population lands in the bottom grade for the first few weeks and the client concludes the model is broken. Warn them in advance.
Because you cannot edit a published model, every boundary tweak is a republish, and a republish recalculates everything. Reports that trend grade counts over time will show a step change on every retune.
A gap between ranges leaves leads with a score and no grade. Views filtered on grade then silently drop them.
The Insights tab reflects only leads within the model's business unit scope, so distribution can look thin in a multi-BU tenant even when the model is working correctly.
Grade names propagate into segments and journey conditions. Renaming one leaves stale references behind that fail quietly rather than erroring.
If two models score the same lead, there are two scores and two grades. Nothing in the product reconciles them — you decide which one sales sees, usually via a view or a form layout.

Consultant notes

Build the model without grades first, let it run for a few weeks, then add grades using the actual distribution. It's an extra republish and it saves an argument.
Show the Scores tab in UAT and pick five leads the sales team knows well. If the model ranks them the way the reps would, you've bought credibility for the whole feature. If it doesn't, better to find out then.
Push back on requests to expose the raw score on the lead form for end users. Reps who see the number start debating the arithmetic; reps who see a grade get on with the call.
Before go-live, confirm the grade values used in segments, views and any Power BI report match the model exactly — including case and spacing.
Put the top-10% figure in the go-live pack. It's the one number that reliably survives the "is this thing working" question three months later.

Worth revisiting after the first full quarter of live scores, when the distribution is real rather than theoretical.