AI Lead Scoring in Dynamics 365: Which Route Fits Your Data

The score nobody trusts

A new lead lands in Dynamics 365 with a grade on it. Your rep glances at the grade, ignores it, and works the list by gut. Marketing reports a healthy MQL number on Monday, and your sales managers call many of those leads junk. Now someone is pitching AI lead scoring, or an agent, as the fix.

In Dynamics 365, AI lead scoring means one of two things. The first is a model trained on your own qualified and disqualified leads (Configure predictive lead scoring) that assigns “a score between 0 and 100” (Prioritize leads through predictive scores). The second is the Sales Qualification Agent, which rates each lead’s fit against your target customer profile as high, moderate or low and sets a lead rating with the Fit, Intent, Recency, Engagement (FIRE) framework (Agent concepts). Customer Insights - Journeys scoring is a third option, and it is points rules your marketers set (Create lead scoring models).

The opening scene is not a model failure. It is a definition failure. When sales and marketing disagree on what a qualified lead is, no lead scoring route fixes that. Put AI on top of a disputed definition and it automates the dispute at scale, with a confident number attached.

If you already run Dynamics 365 and are shopping for new lead scoring software, Microsoft offers three routes, each with its own license and data prerequisites. The real question is which one your data supports today.

Takeaway: before you pick a route, get one written definition of “qualified” that both VPs sign.

Settle the definition before you settle the tool

A shared definition is not a slogan. It is a short document with four parts:

• Fit: who counts as a target customer. This is your target customer profile.
• Intent: which behaviors signal real buying interest, and which are just curiosity.
• Disqualification: a fixed list of allowed reasons, so “not a fit” means the same thing from every seller.
• Ownership: who owns the handoff, and what happens when a lead is rejected.

Treat this as a template, not a finished answer. Your version will reflect your market.

Consistent disqualification matters more than most leaders expect. Predictive scoring trains on leads that were qualified and disqualified in the window you choose (Source), and Microsoft asks for data balanced between those two outcomes (Lead management FAQs). My inference from those two pages: if one rep disqualifies a lead that another would have kept, the outcomes the model learns from are inconsistent. Microsoft does not say this. It follows from how the model is trained.

Recorded disqualification reasons matter too, but for your definition and your reporting. They are how you see where the two teams still disagree.

Across 20+ years in IT and hundreds of engagements in multiple countries, I have watched this argument come up before any technology question. Teams that skip it end up having it later, at higher cost.

Takeaway: the definition has one owner on each side and lives somewhere every seller can read it.

Route 1: Predictive scoring from your own lead history

What it does. Microsoft says “The model assigns a score between 0 and 100 to leads based on the signals from leads and related entities, such as contact and account.” Leads also get a grade from A to D, where A is most likely to convert to an opportunity and D is least likely. “System administrators can define lead score ranges for a grade, depending on your organizational requirements.” (Source)

What it needs. “Your organization must have created and closed at least 40 qualified and 40 disqualified leads during the time frame that you choose to train the model.” That window “ranges from three months to two years.” (Source) Hitting the minimum is not the whole story. The FAQ adds: “If you have more than the minimum requirement, ensure that the data is balanced between the two outcomes (qualified vs. disqualified or won vs. lost) for better model performance.” (Lead management FAQs)

Licensing, which is your decision. Learn says “You must have a Sales Premium, Sales Insights, or a Sales Enterprise license.” (Enable and configure Sales Insights) The same page states: “Licensing options for Sales Insights features have changed.” Confirm your position with your Microsoft account team, as the page advises.

On Sales Enterprise, the cap is 1,500 scored records per month, and the usage graph covers lead and opportunity scoring combined (Lead and opportunity scoring). Sales Premium has no limit (Digital selling capabilities). Above the cap, Microsoft says only that “The app displays a warning symbol” and that the excess “is shown in red.” If you also score opportunities, that cap fills faster than your lead count alone suggests.

Caveats that change the decision:

• Learn says you can publish a model whose accuracy falls below the AUC threshold, but “it will perform poorly.” Do not publish it. A weak score teaches reps to ignore scores, and that habit outlasts the model.
• “Real-time scoring works only for new leads. The scores for updated leads are refreshed only after every 24 hours.” A lead that just did something interesting may not jump the queue until the next refresh, up to 24 hours later.
• “The system takes about four hours to sync the data with the data lake. If you’ve closed leads recently, the model won’t consider them right away.”

Questions for your CRM owner: Do we have 40 and 40 in a window we trust? Do sellers disqualify by the same rules? What license are our sellers actually on?

Pick this if: you have clean closed-lead history and your sellers work leads inside Dynamics 365 Sales.

Route 2: Customer Insights - Journeys scoring models and qualification

What it is. Journeys “scores leads using demographic attributes and interactions.” Your marketers define how many points a lead gets for each condition, and “points can be negative or positive.” A model holds up to 40 groups, and “Scoring models refresh every two hours.” (Source)

Be clear about this with your team: it is rules, not AI. Nothing learns. The model is exactly as good as the points your marketers assign, which makes it a direct test of whether your definition is real.

What it needs. Customer Insights - Journeys in your organization. An administrator enables the Lead Management feature switch, the marketer builds conditions from demographic attributes and “Dynamics 365 standard interactions” (Create lead scoring models), and you only qualify leads from models that are live (Qualify the best leads).

Why you care. This is where the MQL handoff happens. “Once a lead meets this value, then the lead is flagged as Sales-Ready or Marketing qualified.” From there you can “Assign leads to your sales team using Dynamics 365 Sales assignment” or “Trigger a sales sequence.” The threshold is a sales leadership decision, even though marketing builds the model.

Caveats that change the decision:

• “If a lead has already been Marketing Qualified and then the lead scores diminish below the qualification criteria threshold, the lead remains as Marketing Qualified.” Cold MQLs stay MQLs until a person acts.
• “To add or modify new qualification criteria, you need to stop the current qualification model.” Changing the bar is an event, not a tweak.
• Scoring models left over from outbound marketing, which was removed in May 2026, “can’t be used in real-time” (Outbound marketing removal). If your marketers point to an old model, it has to be rebuilt.
• Microsoft lists a known issue: with parent contact scoring enabled, “form submission interactions count twice in the lead scoring model,” and processing slows (Known issues).

Questions for your CRM owner and marketing lead: Who owns the threshold? What happens to leads that miss it? Who re-checks MQLs that have gone cold? Who gets an alert when the routing automation fails?

Pick this if: your bottleneck is the marketing-to-sales handoff, not seller prioritization.

Route 3: The Sales Qualification Agent, starting in research-only mode

The agent comes in two modes, and the mode choice is the most consequential decision in this article.

Research-only mode researches each lead, checks its fit against your target customer profile, and rates it high, moderate or low (Agent concepts). It generates a draft outreach email “for you to review and send.” When fit is “high or medium” (Learn’s handoff page calls the middle band medium), “the agent hands over the lead to a seller as per the assignment rule configured by the admin.” When fit is low, “the agent disqualifies the lead and notifies the supervisor.” (Work on leads handed over)

Research and engage mode goes further. It sends outreach and follow-up emails on its own, checks BANT criteria and detects positive intent from replies. “Leads that demonstrate purchase intent and meet most of your criteria are handed over to sellers for qualification, while those that don’t are disqualified.” Microsoft also states: “The agent doesn’t replace your judgment or decision-making process.” (Sales Qualification Agent overview)

The irreversible calls, which are yours to make. “Only one of the following modes can be deployed in an organization.” “You can upgrade from Research-only mode to the Research and engage mode later, but not the other way around.” And “You can’t delete an agent once configured. Contact Microsoft support to delete the agent from your organization.” (Set up and configure)

Start in research-only mode. That is the right call, not a cautious one. You keep a seller on every outbound email, you learn whether the agent’s fit ratings match your definition, and you keep the option to upgrade. Starting in research and engage mode gives that option away: Learn allows no move back to research-only, and only Microsoft support can delete the agent.

What you budget for and approve:

• A Copilot Studio license (Source).
• Tenant capacity, either “prepaid capacity or pay-as-you-go capacity.” (Sales Qualification Agent overview) Copilot Credits are pooled “across the entire tenant” (Billing rates), so other agents draw from the same pool. When capacity runs out, the agent pauses (Agent FAQs), and Learn says the capability “is unavailable until more capacity is added” (Consumption-based billing).
• Server-side synchronization with Exchange. Without it, “the agent doesn’t generate outreach emails or engage with leads.”
• Your lead data model: “The agent only works on the out-of-box lead entity.” For research, it passes “the lead’s company name, website URL, and any Dataverse fields that you manually define in the configuration settings to Bing Search.” (Responsible AI FAQ)

One point to keep straight in your reporting: the agent gives a high, moderate or low fit and a FIRE-based lead rating. It does not produce a 0 to 100 score, and its ratings are not interchangeable with predictive scores.

Pick this if: your target customer profile is written down and your bottleneck is research time per lead, not lead volume.

Which route your data supports today

That is why the lead scoring software question resolves to a pilot choice: treat each route as its own decision and pick one for the pilot.

Takeaway: choose one route for the pilot, not all three.

Run the AI lead scoring pilot like a sales leader, not a science project

Capture your baseline before anything is switched on. Use your own figures; this article gives no targets. Track at least these:

• lead-to-opportunity conversion
• time to first seller touch
• share of leads disqualified, by reason

Without that baseline, any AI lead scoring pilot ends in a debate about impressions. With it, you have a before and after, and a decision backed by data, not feel.

Then keep a seller reviewing every decision the route makes:

Log every disagreement between the seller and the output. Those disagreements are not noise. They show you exactly where your definition is still unclear, which is often the most valuable finding of the pilot.

One practitioner caveat. The Sales Close Agent, a preview, is already gone: Learn says “Starting September 30, 2026, you can’t create new instances of Sales Close Agent,” and “Existing instances of Sales Close Agent will be removed on October 30, 2026.” That date has passed. (Sales Close Agent overview) Pilot on what Learn documents today, and keep your definition and baseline independent of any one tool. Those two assets survive a product change. A tuned configuration does not.

Takeaway: the pilot passes when sellers act on the output and your baseline metrics move. A dashboard full of scores is not a pass.

Get a second pair of eyes before you sign off

The route you pick matters less than the definition behind it and the baseline you measure against. Start with the route your data supports today, and start the agent in research-only mode, because you can upgrade from it later but never move back, and you can’t delete the agent yourself.

If you want a check before you approve a pilot, Citizen Development Academy runs a 30-minute review. We walk one real lead from form to first seller touch and check five points: shared lead definition, capture and dedupe, scoring readiness, routing and SLA, and handoff feedback and reporting.

Request a 30-minute Lead-to-Revenue Review with Citizen Development Academy