Autonomous Sales Agents in Dynamics 365 Sales [2026]

Autonomous sales agents in Dynamics 365 Sales work in the background rather than in a chat window: they research accounts, qualify inbound leads against your criteria, and prepare draft outreach — so the rep opens a record that already has the work done, and spends the time deciding instead of gathering.

What they do

Grounding, and why it is not optional

A model asked about pricing will produce a price. It will be plausible, it will be formatted correctly, and there is no reason it should be right.

The control is grounding: commercial terms come from verified Dataverse price lists and product catalogues, not from generation. And any draft containing a commercial term routes to a person before it reaches a customer. Those two rules together are what make the deployment defensible.

A quoted price that was never approved is not a model error you can explain away to the customer who received it. Treat the human step as part of the architecture not as caution you will relax later.

Registering custom actions

Agent capability beyond the built-in actions is registered through Copilot Studio, backed by a Dataverse Custom API, a Power Automate cloud flow, or an OpenAPI webhook.

using System;
using Microsoft.Xrm.Sdk;
using Microsoft.Xrm.Sdk.Query;

namespace DataDaur.D365Sales.Agents
{
    public class LeadResearchAgentPlugin : IPlugin
    {
        public void Execute(IServiceProvider serviceProvider)
        {
            ITracingService tracingService = (ITracingService)serviceProvider.GetService(typeof(ITracingService));
            IPluginExecutionContext context = (IPluginExecutionContext)serviceProvider.GetService(typeof(IPluginExecutionContext));
            IOrganizationServiceFactory serviceFactory = (IOrganizationServiceFactory)serviceProvider.GetService(typeof(IOrganizationServiceFactory));
            IOrganizationService service = serviceFactory.CreateOrganizationService(context.UserId);

            if (context.InputParameters.Contains("Target") && context.InputParameters["Target"] is Entity lead)
            {
                if (lead.LogicalName != "lead") return;

                string companyName = lead.Contains("companyname") ? lead["companyname"].ToString() : string.Empty;
                tracingService.Trace(
quot;Agent Triggered for Lead: {lead.Id}, Company: {companyName}");

                // Generate AI Pre-Call Research Brief
                Entity researchBrief = new Entity("datadaur_agentresearchbrief");
                researchBrief["datadaur_name"] = 
quot;Research Brief: {companyName}";
                researchBrief["datadaur_leadid"] = new EntityReference("lead", lead.Id);
                researchBrief["datadaur_briefsummary"] = 
quot;Automated synthetic background check for {companyName}. Verified firmographic data & modern ERP tech stack.";
                researchBrief["datadaur_confidenceScore"] = 0.94;

                Guid briefId = service.Create(researchBrief);
                tracingService.Trace(
quot;Research Brief Created Successfully: {briefId}");

                // Update Lead Status to 'Researched & Ready for Review'
                Entity leadUpdate = new Entity("lead", lead.Id);
                leadUpdate["datadaur_agentstatus"] = new OptionSetValue(100000001); // Researched
                service.Update(leadUpdate);
            }
        }
    }
}

Give each action a description precise enough that the agent knows when it applies. Vague action descriptions are the main reason an agent calls the wrong tool, and the fix is almost always in the description, not in the code.

Identity and scope

1. Entra ID application user. The agent authenticates as its own identity, with a security role scoped to the tables its actions touch.
2. No standing write access to closed records. An agent should not be able to alter a closed opportunity or a posted document.
3. Attribution in the record. Agent-created activity is marked as such, so pipeline reports do not silently mix the two.

Measuring it honestly

The tempting metric is volume: leads touched, emails drafted, hours saved. All of those go up immediately and none of them mean anything.

The metric that matters is whether the agent's qualification agrees with what the pipeline later shows. If agent-qualified leads convert at a materially lower rate than rep-qualified ones, the agent is generating work instead of removing it, and the criteria need fixing before the scope widens.