The SDR Agent Got Its Own Mailbox. Your CRM Still Thinks a Person Sent It
Your CRM still thinks a person sent it.
Microsoft’s Sales Development Agent sends email from its own Exchange account. That sentence is in the company’s October 8 post on bringing CRM into the flow of work, and it is the whole story. An actor with a mailbox of its own now writes into the fields your reps write to.
The agent is in public preview. It researches prospects, personalizes outreach, manages follow-ups, and hands qualified opportunities to a seller once buying signals say a human should take over. On November 16 it starts consuming Copilot credits. Your comp plan, your SDR capacity model and your sourced-pipeline report were all written when a human sent the email.
The meter is the date that matters. After November 16 you cannot cleanly say what a rep produced, because the rep is no longer the only one producing. Before it, you still can. The window is under four weeks.
I run Salesforce every day, so I will say it in Salesforce terms. A Task with a subtype of Email has an owner. A Lead has a source. An Opportunity has a sourced-by value that a report derives from the first touch. None of those fields has a column for “a person did this.” Each one assumes it. If the agent logs under the rep’s user, or under an integration user that someone mapped to a team, next quarter’s meetings-booked count contains two kinds of work and the same label.
What’s actually shipping before the meter
Microsoft, October 8. The Sales Development Agent is in public preview inside the Dynamics 365 and Copilot sales stack, managed in Teams, connected to leading CRM platforms. It starts consuming Copilot credits on November 16, 2026. The post prints no credit price and no consumption rate; Microsoft’s admin notice says estimates come closer to the date. The one customer number is Adobe’s: after three weeks at limited scale, Prabhath Yeluri, a senior manager in sales strategy, says the agent is “driving meetings with 10+ customers we wouldn’t have otherwise reached.” That is ten customers. It is not a conversion rate, and it names no baseline.
Salesforce, September 11. Hunter, the outbound sales agent, is in pilot with general availability set for November 2026. It is the first agent on a runtime Salesforce built to pursue goals “across days and weeks,” and Salesforce says it builds 60% of Perk’s sales pipeline. Two vendors, two proof points, two units. Microsoft counts customers reached. Salesforce counts the share of pipeline. Neither prints what the same team produced before the agent. That is the number a board will ask for, and the vendors have no reason to supply it.
Google, standing rule. Gmail’s sender guidelines ask senders to keep the spam rate reported in Postmaster Tools below 0.10% and to avoid ever reaching 0.30%. This is the counter-read, and it is not about whether the agent works. A human SDR with a bad list damages one mailbox. An agent scales the damage at the speed it scales the volume. Microsoft’s setup docs answer half the question: each agent instance is provisioned with its own identity, mailbox and email alias, and the setup screen may let you pick a different email domain. That screen is the decision. Put the agent on your primary domain and its spam rate becomes your reps’ spam rate.
Last time I covered this ground it was price and packaging, how the SDR job was sold in pieces (Signal, July 5). This is the accounting layer under that. Whoever owns the pieces now has to say which piece a person did.
What I’d build in the revenue org this week
One build: an activity-origin audit, cut before November 16.
Revenue moment: The Monday pipeline review where someone asks if the meetings number went up because the team got better or because a new sender joined.
Data shape: Salesforce Task and Event (OwnerId, CreatedById, TaskSubtype, ActivityDate), EmailMessage (FromAddress, ActivityId), Lead (LeadSource, CreatedDate), Opportunity (CreatedDate, LeadSource) and OpportunityContactRole. In the warehouse: sfdc.task, sfdc.event, sfdc.email_message, sfdc.lead, sfdc.opportunity, plus a new field on Task, Actor_Type__c, with values human, agent, unknown, and an Agent_Id__c.
System shape: A nightly rule that labels every Task by the user that created it. Rep users are human. The integration user the agent connects through is agent. An email sent from the agent’s alias is agent, whoever created the Task. Anything else is unknown. Freeze a 13-week baseline snapshot before the meter starts: emails sent, meetings booked and email-to-meeting conversion per SDR, and sourced pipeline per rep.
Eval shape: Run the labeling rule over the 13 weeks before the agent existed. It must label no row agent. Any agent label there is a bug in the rule, found before it contaminates anything. Whatever lands in unknown is the noise your baseline already carries: the sequencer, the marketing sync, the flow nobody owns. Then, once the agent runs, report human and agent conversion as two separate rates and never as one.
Latency budget: Daily. The review is weekly.
Cost shape: About three engineer-days and one new field. The expensive unknown is the credit meter. Microsoft tells admins to set spending policies before November 16, so set the guard in the pilot: cap credits per week, and log credits against meetings booked from the first day.
Instrumentation: Three counters. Share of outbound email by actor type. Meetings booked by actor type. Opportunities whose first touch was an agent.
Two weeks in, success is a baseline you can show a finance partner, frozen with a date on it, and a column that never mixes the two. The credible failure is the first query. Every agent Task is created by the rep’s own user, because the connector acts on the rep’s behalf, so the rule calls the agent’s work human. The meetings number rises and the label says the team did it. That is a finding too. It means you cannot tell the two apart today.
Sources
-
Bringing CRM into the flow of work and agents into business process for sales and service teams — Microsoft Copilot blog, October 8, 2026; own Exchange account, public preview, credits from November 16, Adobe quote — https://www.microsoft.com/en-us/copilot/blog/2026/10/08/bringing-crm-into-the-flow-of-work-and-agents-into-business-process-for-sales-and-service-teams/
-
Create a Sales Development agent instance (preview) — Microsoft Learn; own identity, Exchange mailbox, email alias, domain choice at creation — https://learn.microsoft.com/en-us/dynamics365/sales/sales-dev-agent/create-instance
-
MC1492151, New Dynamics 365 Sales and Customer Service experiences across Teams, Copilot, and Outlook — Microsoft 365 Message Center, October 2026; consumption estimates closer to transition, spending policies before November 16
-
Salesforce Expands Agentforce With a New Portfolio of AI Agents Built for High-Value Work — Salesforce, September 11, 2026; Hunter pilot and November GA, long-horizon runtime “across days and weeks,” 60% of Perk’s pipeline — https://www.salesforce.com/news/stories/agentforce-job-ready-ai-agents/
-
Email sender guidelines — Google Workspace Admin Help; spam rate below 0.10%, never 0.30% or higher — https://support.google.com/a/answer/81126
-
The SDR job got sold in pieces. — Golden Data, July 5, 2026 — https://goldendata.app/signal/2026-07-05/
Send me an email and we will talk. If something here landed close to what you're working on, the door is open. No calendar funnel, no pitch deck — I read every note that comes in.
Doing the work rather than deciding what to build? Crafting is the column for that chair.
← All Signal issues · Drafted with Claude · Edited by Paul Brown