Microsoft's new Dynamics 365 sales agent is free until 16 November. The real test takes longer than that.

On 8 October 2026, Microsoft's Message Center notice MC1492151 moved its Sales Development Agent into public preview inside Teams, Microsoft 365 Copilot and Outlook. It runs free under Microsoft's Frontier early access programme until 16 November 2026, when it starts drawing down Copilot credits like any other metered agent in the stack.

Five weeks is not a trial period. It is the amount of time Microsoft has decided it can let an AI agent qualify and disqualify your leads before anyone has to pay for the privilege, and five weeks is nowhere near long enough to know whether its calls were the right ones.

What did Microsoft actually announce on 8 October?

A staged rollout, not one launch. The Sales Development Agent and Dynamics 365 CRM support inside Teams both moved to public preview, while other pieces stayed in Frontier or private preview.

MC1492151 covers three separate tiers, and conflating them is the easiest mistake to make. The Sales Development Agent reached public preview, free until the 16 November billing date. Dynamics 365 CRM in the flow of Teams reached public preview on the same notice, requiring an admin to activate the Channel Agent and licences for Microsoft 365 Copilot, Dynamics 365 Customer Service and Teams. A separate set of Sales and Service skills for Copilot shipped at three different maturity levels in the same release: generally available for Cowork, Frontier only for Code, and private preview for Autopilot.

None of this is a general availability date. It is a preview with a billing cliff attached, which independent coverage of the notice flags as the detail admins are most likely to miss when they turn the agent on for a pilot team.

What does the agent actually write back to the CRM?

According to Microsoft's own documentation, it connects as a licensed CRM user with its own identity, logs every interaction and syncs qualification, disqualification and ownership transfer actions straight into Dynamics 365 records.

Microsoft's Agent 365 documentation is specific about the setup: add the agent as a user in the CRM and assign it a licence before it does anything, the same step you would take for a new hire. Once connected, it logs interactions and syncs engagement actions, including emails sent, leads qualified or disqualified, and ownership transfers, back into Dynamics 365, where the sales team can see them.

That is the right design choice, and it is worth saying so plainly. Giving the agent its own identity rather than writing under a shared service account, or worse, impersonating the rep it is working for, is exactly the rule we argued for in agents writing to the system of record. An auditable actor beats an invisible one. The identity problem is solved here. The decision rights problem is not.

Why five weeks is the wrong test window

A disqualified lead is not reversible in the way it looks on paper. Nobody audits a closed lost reason, so a wrong call just disappears, and five weeks is too short for the agent's early decisions to have proved themselves right or wrong.

We have written before about applying a reversibility and blast radius test before handing an agent a decision, in never hand an agent a decision you have not bounded. A disqualify action passes that test on paper: the record still exists, a human could reopen it. In practice it fails it, because nobody goes back to check a lead marked dead. The agent's mistake does not surface as an error. It surfaces as pipeline that was never there, which is the hardest kind of loss to notice.

The free preview window makes this worse, not better. Teams will turn the agent on during Frontier access precisely because it costs nothing, run it against real leads for a few weeks, and reach the 16 November billing decision having judged it on volume of outreach sent rather than on whether its qualify and disqualify calls matched what a human would have done. Outbound-sourced leads routinely take longer than five weeks to either convert or clearly die, so the one signal that actually matters will not exist yet when the bill arrives.

What to actually check before the credits start

Pull every lead the agent touched during the free window and check the disqualified ones by hand, decide your own sign off rule now, and make sure its actions are attributed to its own identity rather than a shared login.

• Pull every lead the agent disqualified during the preview, not a sample chosen after the fact, and have a rep who was not involved check whether they would have made the same call.
• Set your sign off rule before you need it. Decide now whether a disqualify is final on write or needs a human to confirm it within a set number of days, rather than deciding the first time a wrongly closed lead turns up months later.
• Confirm the audit trail actually shows the agent, not a generic service account. If you cannot tell which qualification decisions were the agent's from the CRM record alone, the identity design Microsoft built is not switched on correctly in your tenant.
• Price the next five weeks on outbound volume, not on how the free period felt. Copilot credits meter usage, and a prospecting agent run against a full target account list consumes very differently from the inbound assistants most teams have already budgeted for.

None of this is specific to Microsoft. The same questions apply to Salesforce's Agentforce SDR agents and to any vendor selling an AI agent that writes qualification decisions into your system of record. We covered the broader decision rights problem, and the autonomy levels worth distinguishing between, in the revenue orchestration field manual.

If you are piloting an AI SDR agent against a live pipeline, the question worth answering before any bill arrives is not whether it sends more outreach. It is whether its qualify and disqualify calls hold up against what your own reps would have done. We help revenue teams build that check in before they switch an agent on, not after. See how in our AI automation work, or read how we think about the handoff in from outsourced SDR to engineered cadence.