Dynamics 365 Lead Scoring: Predictive Models and AI Agents (2026)

Abstract illustration of a row of translucent teal glass spheres of rising size with the largest sphere glowing orange, representing scored and ranked leads

Last updated: November 16, 2026

By ZapAI Team

Dynamics 365 lead scoring now comes in three flavors, a predictive model that learns from your closed leads, rule-based points in Customer Insights – Journeys, and an AI agent that researches and emails leads for you. Our Dynamics 365 Sales consultants set all three up for mid-market US sales teams, and this guide covers how each works, what data and licenses they need, how to judge accuracy, and the rules we use to pick one, checked against Microsoft Learn on November 16, 2026.

Dynamics 365 lead scoring uses a machine learning model, trained on at least 40 qualified and 40 disqualified leads, to score open leads 0 to 100 and sort them into grades. Enterprise scores 1,500 records monthly. Premium has no cap.

Here’s a hypothetical you may recognize. Six inside reps. 900 open leads. Everyone works whatever landed most recently, and the good ones go stale by Thursday.

Scoring is supposed to fix that. Sometimes it does. And sometimes it hands reps a confident number built on two years of junk data, so they learn to ignore it within a month.

We wrote this for sales leaders and CRM admins deciding what to switch on. License prices live in our Dynamics 365 Sales pricing breakdown, so we skip the price sheet here and focus on how scoring works and how to deploy it without wrecking trust.

What Counts as Lead Scoring in Dynamics 365?

Three different tools share the name. Predictive scoring in Dynamics 365 Sales learns from history, Customer Insights – Journeys adds and subtracts points by rules you write, and the Sales Qualification Agent researches each lead and decides whether a seller should get it.

  • Predictive lead scoring rates open leads by how closely they resemble leads you qualified before.
  • Predictive opportunity scoring does the same for open deals, using won and lost history.
  • Journeys scoring models count demographics and engagement, like email opens or event registrations, and flag a lead as sales-ready past a threshold.
  • The Sales Qualification Agent reads CRM data and public web sources, checks fit against a target customer profile, and drafts or sends outreach.

They don’t compete much. Plenty of teams run two at once. The trick is knowing which question each one answers.

How Does Predictive Lead Scoring Work?

An admin creates a model in Sales Insights settings. The model trains on leads that were created and closed in a window you choose, learns which attributes separate qualified from disqualified leads, and then scores every open lead that matches the model’s filters.

Minimum data is firm. Per Microsoft’s predictive lead scoring setup guide, you need at least 40 qualified and 40 disqualified leads created and closed inside the training window, which runs from three months to two years and defaults to two. Short of that, the Get started button stays gray. No workaround.

Data syncs to the data lake in about four hours. Training takes minutes. Be patient.

Scores run 0 to 100. Grades sit on top. Microsoft’s lead management FAQ says the model produces the score while grades are just four buckets an admin configures, shown as A through D. Same FAQ, two timing rules. New leads get scored within about five minutes of saving, and updated leads refresh every 24 hours.

Reps see why, too. The score widget on the lead form lists the top positive and negative influencing factors, things like industry, email domain or lead source, and that explanation is the feature reps actually use, because a bare number with no reason attached gets ignored the second it contradicts their gut.

What the model can and can’t learn from

Out of the box it trains on standard attributes. You can add custom columns from the lead, contact and account tables. Microsoft’s edit and retrain page, updated May 29, 2026, lists what it won’t take.

  • Attributes on custom tables.
  • Date and time columns.
  • System columns like lead score or version number.

Two settings matter more than they look. Ignore empty values stops a mostly blank column from acting like a signal. Intelligent fields check email domains and name validity, so the model learns that a gmail address and a first name of asdf rarely turn into pipeline.

One model or several?

Up to 10 models per scoring type, published or not. Build separate ones when regions, business units or product lines sell differently, and use a filter column to split leads between them, with a calculated column if the split needs more than one field.

Business process flows add a catch. Leads that abandoned the selected flow drop out of training and scoring entirely. And Microsoft’s real-time scoring notes say brand new leads ignore the business process flow, so two models with the same filter but different flows can score a fresh lead with either one. Give each model its own filter. Saves a headache.

How Is Predictive Opportunity Scoring Different?

Same engine, different outcome. Opportunity scoring trains on won and lost deals instead of qualified and disqualified leads, needs at least 40 of each, and can weigh attributes differently at each stage of your sales process.

That stage-level option is called a per stage model. Microsoft’s opportunity scoring setup page says it’s off by default, can only be turned on when you add a model, and needs at least 40 closed opportunities in the last stage of the chosen business process flow. Worth it? Usually, once you have the volume. Estimated close date means one thing at Qualify and something else at Propose.

Opportunity scores also feed pipeline conversations. Our sister post on Dynamics 365 Sales forecasting covers how scored deals show up in forecast reviews.

Many small teal glass beads flowing into a clear glass prism that emits four stacked bars of light, one of them orange, representing training data becoming lead grades

Which License Do You Need, as of November 2026?

Sales Enterprise or Sales Premium. Enterprise includes predictive lead and opportunity scoring with a shared cap of 1,500 scored records a month, Premium removes the cap, and Microsoft’s pricing page doesn’t list predictive scoring under Sales Professional at all.

Learn’s digital selling capabilities page spells out the cap. Pass it and the usage panel shows a warning. The Dynamics 365 Sales pricing page lists the rest.

Item, as of November 2026Sales ProfessionalSales EnterpriseSales Premium
List price, per user per month, annual$65$105$150
Predictive lead and opportunity scoringNot listed1,500 scored records a monthNo limit
Copilot Credits includedNone listedNone listed1,000 per user per month
Sales Qualification AgentListed, runs on Copilot CreditsListed, runs on Copilot CreditsListed, draws on included credits first

Is 1,500 enough? Do the arithmetic before you buy. Leads and opportunities count against the same cap, and every new lead your web forms create gets scored, so a team pulling in 1,200 leads a month plus 400 open deals is already over. Do rescored updates count too? Microsoft doesn’t say. Watch the usage panel in month one.

How Do You Set Up Predictive Scoring?

You get two routes. Quick setup on the Sales settings page creates and publishes default lead and opportunity models in a couple of clicks, while advanced settings let you pick the business process flow, filters, training window and attributes before anything goes live.

We use the advanced route every time. Here’s the order.

  1. Confirm the license and that advanced Sales Insights features are on.
  2. Count closed leads in your window, qualified and disqualified separately.
  3. Open Sales Insights settings, then Predictive models, then Lead scoring.
  4. Pick the business process flow, status option set, filter column and training window.
  5. Train, then open View details and read the Performance tab.
  6. Edit attributes, switch on Ignore empty values where blanks dominate, retrain.
  7. Publish, keep Retrain automatically on, and add the score widget to any custom lead form.
  8. Repeat for opportunities.

Step seven trips teams on custom forms. The widgets only land on the out-of-the-box Sales Insights form, so reps on your own form see nothing until someone adds the control. Check Dataverse roles too. Setup needs the Sales administration roles, and our Dataverse security roles guide explains why a copied custom role can miss privileges Microsoft added later.

Retraining runs every 15 days when the toggle is on, and it’s on by default after publishing. It isn’t blind. Per the retrain page, the new version goes live only if its accuracy is at least 95 percent of the active model’s, or if the active model is more than three months old. Otherwise the old one stays. Sensible guardrail, and almost nobody knows it exists.

How Do You Know the Model Is Any Good?

Read the Performance tab before you publish. It shows accuracy, recall, conversion rate, F1 score, area under the curve and a confusion matrix, all calculated on the most recent 20 percent of closed records held back from training.

Microsoft’s model accuracy page gives a worked example worth studying. A sample of 1,000 opportunities scores 75 percent accuracy and 92 percent recall. Sounds decent. But the conversion rate in that same sample is 70 percent, which means a model that called every single deal a win would be right 70 percent of the time without learning anything at all.

So accuracy alone flatters. We look at three things instead.

  • AUC, since it measures whether the model ranks a real win above a real loss.
  • False positives in the confusion matrix, because those are the leads reps chase for nothing.
  • Whether the top factors make sense to your best rep. If they don’t, something’s off in the data.

Below Microsoft’s AUC threshold, the app marks a model Not ready to Publish. You can publish anyway. Don’t. Microsoft says plainly that such a model performs poorly, and a weak model trains reps to distrust scores for good.

Bad data that poisons a model

Models learn your habits, good and bad. These are the patterns we find most when a model won’t reach publishable accuracy, and every one of them is a process problem wearing a data costume.

  • Leads closed as disqualified in bulk during a cleanup, which teaches the model that old leads lose.
  • Test records from the implementation still sitting in the training window.
  • Reps qualifying leads only to convert them into opportunities they already won elsewhere.
  • Duplicates. Same company, three leads, three different outcomes.
  • A lead source field that half the team never fills in.
  • Deals closed in Business Central or a spreadsheet while the opportunity stays open forever.

That last one is common for distributors. If orders live in ERP, wire the outcome back to the opportunity, which our post on Business Central and Dynamics 365 Sales integration walks through. Microsoft’s own fix for messy history is blunt. Pick a shorter training window, like six months or a year, when the data was cleaner.

A glowing orange glass orb hovering above a scattered field of teal glass cubes, a few cubes lit brighter beneath it, representing an agent researching leads

What Does the Sales Qualification Agent Do?

It’s an autonomous agent inside Dynamics 365 Sales that researches assigned leads, checks fit against a target customer profile you define, and either drafts outreach for a seller or sends it, follows up and hands over only leads that show purchase interest.

Two modes exist, per the Sales Qualification Agent overview. No preview label on it, either.

CapabilityResearch-only modeResearch and engage mode
Research leads and check target customer profileYesYes
Check budget, authority, need and timelineNoYes
Generate outreach emailsYes, seller sendsYes, agent sends
Detect purchase interest and follow upNoYes
Hand over promising leads to sellersYesYes

Read the fine print before picking. The agent setup guide, updated September 18, 2026, says only one mode can run per organization, you can upgrade from research-only to research and engage but never back, and a configured agent can’t be deleted without a call to Microsoft support. That’s a one-way door.

Prerequisites go well beyond a license. You need a Copilot Studio license, the modern Sales Hub interface, data policies that allow the Copilot Studio, Dataverse and public website connectors, server-side sync with Exchange, and in-app notifications. Engage mode also needs a shared mailbox. The agent runs as an app user with the AISalesPerson role, and custom seller roles need extra table privileges listed on that page.

Our take? Start in research-only mode. Reps get research and a draft email, they decide what goes out, and you learn whether the agent’s idea of a good fit matches yours before it talks to a single prospect on your behalf. Copilot features beyond the agent are in our Copilot in Dynamics 365 use cases post.

What does the agent cost to run?

Copilot Credits. Per Microsoft’s consumption billing page for Sales, the agent bills through prepaid Copilot Studio capacity or pay-as-you-go on an Azure subscription, prepaid gets used first, and when the quota runs dry the capability simply stops until someone adds more.

Microsoft hasn’t published a per-lead figure. What exists are feature rates on the Copilot Studio billing rates page, such as 5 credits per agent action and 2 per generative answer, plus a premium token charge when a reasoning model runs the show. A lead that needs web research, a profile check, an email and two follow-ups will burn several of those. How many? Nobody outside Microsoft can tell you yet.

So meter a pilot. Run 100 leads, read consumption in the Power Platform admin center, divide, then multiply by your monthly volume. Our Copilot Studio pricing guide covers credit packs and pay-as-you-go if Premium’s included credits won’t stretch.

The other sales agents

The pricing page also lists Sales Opportunity and Sales Research agents under all three plans, and the Sales Close Agent as preview, while Learn still labels the Opportunity Agent setup page preview. Then there’s the Sales Development Agent. Message center post MC1492151, updated November 11, 2026, says it’s in public preview and moves to usage-based Copilot Credit billing on December 1, 2026. Pilot those in a sandbox. Not production.

When Do Rule-Based Options Make More Sense?

When you have too little history to train a model, when marketing engagement matters more than firmographics, or when your qualification rules are already clear. Business rules, Power Automate and Customer Insights – Journeys scoring models all score by logic you write and can explain line by line.

Journeys is the serious option. Microsoft’s Journeys scoring model guide, updated October 13, 2026, lets you combine demographic attributes with interactions like email opens, form submits and event registrations, give each condition positive or negative points, and add up to 40 groups per model. Scores refresh every two hours. Grades like Hot or Cold are optional.

Handoff is where it earns its keep. Per the Journeys lead qualification page, a lead that crosses your threshold gets marked Marketing Qualified or Sales-Ready, which can trigger Dynamics 365 Sales assignment or a sales sequence. One quirk. A lead stays Marketing Qualified even if its score later drops.

Sequences belong to the sales accelerator, which our sister post on the Dynamics 365 Sales accelerator covers. For the wider Journeys picture, see our Customer Insights guide.

Lighter options exist. A business rule can set a rating from two conditions. A Power Automate flow can add points when a form arrives. Cheap. Transparent. Both rot when nobody owns the rules, so our Power Automate team gives every scoring flow an owner and a review date.

Predictive Scoring vs Journeys Scoring vs Agent Qualification

Each answers a different question. Predictive scoring asks which leads look like past winners, Journeys asks who is engaging right now, and the agent asks whether this specific lead fits and wants to talk.

FactorPredictive lead scoringJourneys rule-based scoringSales Qualification Agent
How it decidesMachine learning on closed lead historyPoints from rules you writeResearch plus target profile and, in engage mode, budget, authority, need and timeline
Data needed to start40 qualified and 40 disqualified leads in the windowClear rules and tracked interactionsTarget customer profile, product info, knowledge sources
LicenseSales Enterprise with 1,500 records a month, or PremiumCustomer Insights – JourneysAny Sales plan plus Copilot Studio license and Copilot Credits
RefreshNew leads within minutes, updates every 24 hoursEvery two hoursPer lead, as the agent works it
ExplainabilityTop influencing factors per leadFully transparent rulesResearch summary and fit reasoning
Running costIncluded in licenseIncluded in licenseCredits per action, no published per-lead figure
Best forSteady inbound volume with clean historyMarketing-led funnels with nurture journeysLead volume your team can’t research by hand
Main riskLearns bad habits from bad dataRules drift and nobody updates themCredit burn and outreach you can’t take back

Before You Turn It On, a Data Readiness Checklist

Run this list before any scoring goes live. In our experience it takes a half day for a mid-market org, and it decides whether reps trust the number in month two.

  • At least 40 qualified and 40 disqualified leads in a six-month window, ideally far more.
  • Roughly balanced outcomes. Microsoft’s FAQ asks for balance between qualified and disqualified.
  • No bulk disqualifications from data cleanups inside the window.
  • Test and demo records deleted.
  • Duplicate detection rules on for leads.
  • Lead source, industry and company size filled in on most records.
  • One business process flow per sales motion.
  • Closed outcomes from ERP written back to the opportunity.
  • A named owner for the model and a monthly review slot.
  • Custom lead and opportunity forms with the score widget added.
  • For the agent, a written target customer profile your sales manager signs off on.

Broader go-live prep sits in our Dynamics 365 CRM implementation checklist. Here’s a hypothetical to make it concrete. A distributor with 2,000 closed leads in two years finds 700 of them were disqualified in one afternoon during a 2025 cleanup. Train on two years and the model learns that leads created before the cleanup lose. Train on the nine months since, and it learns something real.

A vertical column of five flat glass tiles, four glowing teal and the top tile glowing orange, representing a data readiness checklist

ZapAI Decision Rules for Lead Scoring

These rules are our practice opinion, built from Microsoft’s documentation and our own projects. Treat them as a default. Your sales motion might argue otherwise.

If this is trueWe recommendWhy
Fewer than 150 qualified and 150 disqualified leads in the last yearRules first, in Journeys or a simple flowThe Microsoft floor is 40 each, but we’ve rarely seen a stable model that close to it
Steady inbound volume, clean history, Sales EnterprisePredictive scoring, after counting monthly volume against the 1,500 capFree with the license, and scores explain themselves
Over 1,500 scored records a monthPrice Premium against splitting scoring across fewer recordsThe cap is environment-wide and new leads count
Marketing owns the funnel and runs nurture journeysJourneys scoring for sales-ready, predictive scoring for rep priorityEngagement and fit are different signals
More leads than reps can researchSales Qualification Agent in research-only mode, metered 100-lead pilotMode choice is one way and credit burn is unknown until measured
Regulated industry or strict outbound rulesHold off on research and engage modeAn agent sending email in your name needs legal review first
Model stuck at Not ready to PublishShorten the window, fix the data, add attributesPublishing a weak model costs more trust than it saves time

Bias disclosed. We’re a partner and we get paid to set these up. Even so, a team with 30 leads a month and three reps probably doesn’t need any of this yet. A shared view sorted by lead source and created date does the job.

Wrapping Up

Pick the tool that matches the question you’re asking. Predictive scoring ranks leads by resemblance to past wins and needs clean history, Journeys scores engagement by rules marketing can read, and the Sales Qualification Agent does the research nobody has time for, at a credit cost you should meter before you scale. Check the 1,500-record Enterprise cap, read AUC and false positives instead of accuracy, and fix the data first.

ZapAI is a Microsoft Dynamics 365, Power Platform, Fabric and AI consulting firm in Irving, Texas, working with mid-market businesses across Dallas-Fort Worth and nationwide. We implement Dynamics 365 Sales lead scoring alongside the rest of the Dynamics 365 platform, from data cleanup to agent pilots. Want to know if your data can carry a model? Ask us for a lead scoring readiness review, and we’ll count your closed leads, read the history and tell you which option fits before you buy anything.

Questions Sales Leaders Ask About Lead Scoring

What’s the minimum number of leads Dynamics 365 needs to build a scoring model?

Forty qualified and forty disqualified leads, created and closed inside the training window you pick, which runs from three months to two years. Opportunity scoring needs 40 won and 40 lost. More helps.

Does Sales Enterprise include predictive lead scoring?

Up to a point. Enterprise scores 1,500 lead and opportunity records a month combined, and the usage panel flags anything past that. Sales Premium has no monthly limit, as of November 2026.

So what’s the difference between a lead score and a lead grade?

The machine learning model generates the score, 0 to 100. Grades are four buckets, A to D, that an admin defines by score range so reps can sort quickly without debating whether 61 is good.

Is the Sales Qualification Agent generally available?

Microsoft announced general availability in October 2025, and the Learn overview carries no preview label. Other sales agents differ. The Sales Close Agent is listed as preview on the pricing page, and the Sales Development Agent is in public preview with Copilot Credit billing starting December 1, 2026, per message center post MC1492151. Statuses move fast around Ignite, so check Learn before you plan around any of them.

Realistically, how much does the agent cost per lead?

Nobody can say yet. Microsoft publishes credit rates per agent action and generative answer, not per lead, so run a metered pilot of about 100 leads and divide.

Can we use predictive scoring and Journeys scoring together?

Absolutely, and we usually recommend it for marketing-led teams. Journeys decides when a lead is sales-ready based on engagement. Predictive scoring then tells reps which sales-ready leads look most like past wins.

Why isn’t a lead getting scored?

Check five things. The model is published, the lead is open, it matches the model’s filters, it uses the model’s business process flow, and it was created inside the two-year window. Custom forms also need the widget added before reps can see anything.

Leave a Reply

Your email address will not be published. Required fields are marked *

Book a free consultation