Ignite 2026 put Salesforce agents inside Copilot and took Contact Center Intelligence to GA. The convenience is real. The governance split underneath it is documented, and almost nobody has read the docs.
Last Updated: November 20, 2026
Copilotforce puts Salesforce context and 38+ sales skills inside Microsoft Copilot, and Microsoft Agent 365 now issues Entra identities to Agentforce agents. Both are real and both are useful. But the acting identity, the authoritative execution record, the retry boundary and the spend meter now live on two different control planes, and the vendor documentation says plainly that Microsoft will not infer the missing pieces. Consolidate read-heavy, human-confirmed work onto the vendor surface. Keep multi-system writes, unattended runs and anything with an audit requirement longer than 30 days in a layer you control.
What actually shipped at Ignite this week
Three announcements matter for anyone running agents across more than one vendor. First, Salesforce and Microsoft announced Copilotforce on 17 November, a plugin that brings Sales Cloud workflows into Copilot, a Sales Cloud app for Teams, and bidirectional zero-copy data between Salesforce Data 360 and Microsoft Fabric. Salesforce lists the Copilot plugin as public preview (beta) and the Teams app as pilot; only the Fabric OneLake data federation is generally available.
Second, Microsoft shipped the plugin as “Salesforce in Copilot” with more than 38 prebuilt sales skills, alongside Dynamics 365 plugins carrying more than 100 skills and an expanded Agent 365. Third, Contact Center Intelligence reached general availability, layering Microsoft voice and digital agents on top of contact centre platforms customers already run.
The pattern is overlay, not replacement. Microsoft has stopped asking you to migrate and started asking to sit on top of what you have. That is a genuinely better deal for buyers than rip-and-replace. It is also how two vendors end up jointly responsible for a single transaction with no joint accountability model.
The seam is identity, and it is already documented
There are two completely different authentication paths in this integration, and most coverage has collapsed them into one.
The runtime path is delegated. Salesforce states that Salesforce in Microsoft Copilot operates through authenticated user access, preserving granular permissions, sharing rules and audit trails in both environments, and that writeback actions require explicit user confirmation and respect existing validation rules. Good design. The seller is the actor, Salesforce enforces Salesforce rules, Microsoft enforces Microsoft rules.
The governance path is not delegated. To pull Agentforce activity into Agent 365, Microsoft's own setup guide has you create a Salesforce external client app with the client credentials flow enabled and a “Run As” user holding the chatbot_api, sfap_api and api scopes. That is a shared service identity. Everything Agent 365 reads from Salesforce is read as that account, not as the employee whose work it describes.
Then the identity data itself degrades at the boundary. The same guide says Active users shows N/A because Agentforce sessions do not include Microsoft Entra identities, and the observability reference is blunt about the consequence: identity-based metrics display as unavailable when the source platform does not provide the signal, and Agent 365 does not infer missing identity information. You get sessions, exceptions, runtime and tool calls. You do not get who.
Even on the Microsoft side, agent identity is a tag rather than a first-class object. Entra audit logs record agent blueprint activity as application events, agent identity activity as service principal events and an agent's user account as user events, distinguished by an agentType property and a blueprintId. Any SIEM rule you wrote for service principals now fires on agents, and any report that counts “agents” needs a join before it means anything.
One seller request, three identities
Entra user token. The Copilot session, model routing and Work IQ grounding are all Microsoft-side records.
Salesforce authenticates the user independently and applies its own sharing rules. The handoff itself is the least-logged hop in the chain.
Captured only if Agentforce Session Tracing was switched on in Data 360 before the run.
Explicit user confirmation plus validation rules. Today this is what stops a bad loop; it is also what stops the workflow scaling unattended.
Read by the external client app's Run As service user, on a manual sync you trigger. Not the employee, not in real time.
Your log store. Nothing in either vendor's stack joins steps 1 and 3 for you.
Whose log is authoritative during an incident
Ask that question before you turn the plugin on, because the answer has three uncomfortable properties.
It is retroactive-hostile. Sessions that run while Agentforce Session Tracing is off are not recorded and cannot be collected later. There is no backfill. If the integration goes live before a Salesforce admin enables tracing, Agent Platform Tracing and Audit and Feedback in Einstein Audit, Analytics and Monitoring Setup, the first weeks of production simply have no execution record on the Salesforce side.
It is slow. The same guide warns that the first sessions after enabling tracing can take up to two hours to reach Data 360, and Anthropic Claude activity can take five to ten minutes to appear by comparison. Agent 365 collects new sessions when the connection syncs, and synchronisation is manual only, with no prescribed interval. An incident timeline assembled from a manually synced registry is a reconstruction, not a trace.
It is short, and incomplete about itself. Observability data is kept for 30 days and then deleted. And Microsoft states that it does not currently guarantee that connection create, sync and delete actions appear as searchable events in the Microsoft Purview audit log, recommending you record those changes through your own change-management process instead. The control plane that governs the cross-vendor connection is not itself fully auditable yet. That is an unusually honest disclosure and it deserves to be read as a requirement, not a caveat.
Treating the Agent 365 Activity tab as the audit trail
It looks like one. It shows sessions, exceptions, runtime and tool calls for Salesforce, Bedrock, Vertex AI, Claude and Databricks agents in a single pane, which is exactly what a compliance reviewer wants. But it is a 30-day operational view, assembled on manual sync, with user identity blank for Agentforce. It will not survive a regulator asking who approved a write 90 days ago.
Retries, token cost, and the two meters
Nobody bills a cross-vendor retry loop to one place. Microsoft's commercial model is a per-user subscription plus usage-based consumption, and Agent 365 Cost Management tracks consumption across users, services, agents and models with spending policies, limits and alerts. That meter covers Microsoft-side services. Work the Salesforce agent performs consumes Salesforce-side entitlements under Salesforce's own commercial terms. If a Copilot-initiated loop calls an Agentforce skill fourteen times because a validation rule keeps rejecting the payload, each vendor sees a plausible, in-policy usage curve and neither sees the loop.
Microsoft is moving toward a cross-vendor view. Agent 365 Business Analytics measures investment in work completed rather than tokens, and value insights across Microsoft and non-Microsoft AI services are scheduled for December 2026. Useful for a board slide. Not the same thing as a reconciled bill, and not something to design your chargeback model around while it is in Frontier preview.
The reason this has not bitten anyone yet is the human in the loop. Confirmed writebacks cap the blast radius. That guard disappears as the surface goes autonomous: Agent Accounts enter public preview in December 2026 with memory and triggers, and Autopilot already runs unprompted with its own Entra Agent ID and a named manager. The day a trigger fires at 03:00 with no one to confirm is the day retry semantics stop being theoretical.
Assuming the connected-platform registry is live inventory
Microsoft documents two limits that break the assumption: synchronisation is manual, and deletion reconciliation is not guaranteed, so an agent removed in Salesforce may persist in the Microsoft 365 agent registry. Teams build offboarding and access reviews on top of a registry that silently drifts.
Where control actually sits
The split below is what we would put in front of a risk committee. Note how often the honest answer in the last column is “you”.
| Concern | Microsoft side | Salesforce side | Practical owner |
|---|---|---|---|
| Authenticating the acting human | Entra user token, Conditional Access | Independent authenticated user access, sharing rules | Shared; two policy engines must agree |
| Identity of the agent | Entra Agent ID, rolling out for partner platforms through December 2026 | Agentforce agent definition in the org | Microsoft, for governance only |
| Execution record | Agent 365 Activity, 30-day retention, manual sync | Agentforce Session Tracing in Data 360, must be enabled in advance | You, via export and correlation |
| Spend meter | Agent 365 Cost Management | Salesforce-side entitlements | You; no joint meter exists |
| Retry and idempotency | Not specified at the plugin boundary | Validation rules plus explicit user confirmation | You, once runs go unattended |
| Deprecation of the integration | Microsoft | Salesforce | Neither party owes you notice jointly |
Preview status is a governance fact, not a footnote
Count what is actually GA in this story: the Fabric OneLake data federation, Contact Center Intelligence, the Dynamics 365 plugins, plugins and skills governance. The cross-vendor agent layer is not. The Salesforce Copilot plugin is public preview beta, Teams is pilot, third-party observability is a Frontier preview, dynamic agent risk is preview, and Oracle Generative AI Agents and Snowflake Cortex are still marked “coming soon” for observability while syncing agents today. You can be governing an agent you cannot observe.
There is also a commercial reality worth saying out loud. In the same keynote cycle, Microsoft promoted Dynamics 365 Activate for CRM, generally available, which helps customers migrate from Salesforce to Dynamics 365, and argued that Dynamics data is natively integrated with Work IQ in a way partner plugins are not. Meanwhile ServiceNow used the same month to launch an Autonomous CRM Engine aimed at moving enterprises off legacy CRM platforms. Partnership and displacement are running on the same calendar. Architect for the version where one side's enthusiasm cools.
ServiceNow announces its Autonomous CRM Engine, targeting CRM migration cost directly.
Ignite Day 1: Copilotforce announced, Contact Center Intelligence GA, Agent 365 expanded with Entra Agent ID for partner platforms, secretless connections and dynamic agent risk.
Entra Agent ID rollout across Agentforce, AgentCore, Vertex AI, Databricks, Snowflake, Claude, OpenAI, Oracle, SAP, ServiceNow and UiPath completes. Agent Accounts enter public preview. Action center preview. Value insights extend to non-Microsoft AI services.
AI is only as powerful as the data and context that grounds it.
— Judson Althoff, CEO, Microsoft Commercial Business, Salesforce Newsroom
Vendor surface or neutral layer: how we decide
This is not an all-or-nothing choice, and treating it as one is how teams end up rebuilding Work IQ badly in n8n or putting an irreversible three-system write behind a preview plugin. Split the estate by workflow shape.
| Workflow shape | Best owner | Why |
|---|---|---|
| Read-mostly: meeting briefs, pipeline summaries, deal-health reads | Vendor surface | The grounding across Teams, Outlook and SharePoint is the product. You cannot rebuild it cheaply and the failure mode is a bad paragraph, not a bad record. |
| Single-system write with a human confirming | Vendor surface | Confirmation and validation rules are already enforced at the CRM boundary. |
| Writes spanning three or more systems of record | Neutral layer | You need one transaction boundary, one retry policy and one compensating action. Neither vendor offers that across the seam. |
| Unattended or trigger-driven runs | Neutral layer | No human confirmation means you supply idempotency keys, backoff and a dead-letter queue. |
| Audit requirement beyond 30 days | Neutral layer | Agent 365 observability retention is 30 days; your obligations are usually years. |
| Agent-to-agent handoff between two vendors' agents | Neutral layer | Failure attribution is undefined when neither side logs the other's reasoning. |
| Deep dependency on Microsoft 365 permissions and sensitivity labels | Vendor surface | Re-implementing label inheritance outside the tenant is a security regression, not an architecture. |
The neutral layer does not have to be heavy. In most builds it is an orchestrator that owns four things and nothing else: a correlation ID minted at intent, an idempotency key per external write, a retry and compensation policy, and an append-only event log that both vendors' records get joined into. n8n, Temporal or a small service all work; the choice matters far less than the contract. Keep the MCP servers and tool registries governed centrally too, which Microsoft now supports through MCP server governance covering Azure AI Gateway and third-party gateways such as LiteLLM, showing which agents depend on a server before you block it.
This is the same conclusion we keep reaching across model generations and now across vendor surfaces: capability is rarely the constraint, coordination is. More on how we structure that in integration engagements, with worked examples in our case studies and model-by-model behaviour in model coverage.
What to put in place before you switch it on
Six things, in this order, and the ordering is load-bearing.
One: have a Salesforce admin enable Agentforce Session Tracing, Agent Platform Tracing and Audit and Feedback before the first production run, because sessions executed with tracing off are unrecoverable. Two: mint a correlation ID in the Copilot-side request and carry it into the Salesforce payload, in a custom field if necessary; without it, joining the two logs is guesswork. Three: ship both log sources into one store with your real retention period, since the vendor view expires at 30 days. Four: name a human owner per agent and run the Action center review against ownerless and inactive agents once it lands. Five: set spending limits and alerts in Agent 365 Cost Management, and separately agree an expected volume with your Salesforce account team so a loop is visible from both ends. Six: write down what you do if either vendor deprecates the integration, and keep at least one non-plugin path to the same CRM action.
None of that is exotic. It is the ordinary discipline of running a distributed system, applied to a thing the marketing calls a plugin. If you want a second pair of eyes on where your own seams fall, talk to us.
Frequently Asked Questions
Do Salesforce agents actually run inside Microsoft's tenant?
No. Agentforce still executes in Salesforce. What runs in Copilot is a plugin surface that authenticates the user to Salesforce, invokes Salesforce skills and renders the results. Separately, Agent 365 can register those agents and pull session telemetry. The execution stays on Salesforce infrastructure; only identity registration and a summarised activity view land on Microsoft's side.
Which log is authoritative if an agent writes a bad CRM record?
Salesforce's own record of the write is authoritative for what changed. Entra and Copilot logs are authoritative for who asked and what the surface did. Neither is authoritative for the whole chain, and Agent 365's activity view is a 30-day operational summary gathered on manual sync. Unless you correlate the two sides yourself, there is no single authoritative incident timeline.
Who pays for a retry loop that spans Copilot and Agentforce?
Both of you, separately. Microsoft meters Copilot-side consumption through Agent 365 Cost Management; Salesforce meters its own side under its own commercial terms. There is no joint meter today. Microsoft has said value insights covering non-Microsoft AI services arrive in December 2026, but that is reporting, not reconciliation. Set limits on both sides and alert on call-count anomalies, not just spend.
Is Entra Agent ID enough to govern third-party agents?
It is a real improvement and a necessary foundation, but it governs the agent object, not every action. Entra audit logs classify agent activity under existing application, service principal and user event types with an agentType tag, so your detection rules need updating. And for Agentforce specifically, Microsoft documents that user identity is unavailable in collected activity and is not inferred.
Should we drop our n8n or custom orchestration layer now?
Not for anything that writes to multiple systems, runs unattended, or needs an audit trail longer than 30 days. Do move read-heavy, human-confirmed workflows onto the vendor surface, because the Microsoft 365 grounding is genuinely hard to replicate. Keep the neutral layer thin: correlation IDs, idempotency keys, retry and compensation policy, and an append-only event log.
What is the single highest-value thing to configure first?
Agentforce Session Tracing, before any production traffic. Microsoft's connection guide states that sessions running while tracing is off are not recorded and cannot be collected later. There is no backfill. Every other gap in this article can be closed retroactively; a missing execution record cannot. Enable tracing, Agent Platform Tracing and Audit and Feedback, then verify a test session reaches Data 360.
How much of this is generally available today?
Less than the headlines suggest. Contact Center Intelligence, the Dynamics 365 plugins and the Fabric OneLake data federation are GA. The Salesforce Copilot plugin is public preview beta, Sales Cloud in Teams is pilot, third-party observability is a Frontier preview, and Entra Agent ID for partner platforms completes rollout in December 2026. Pilot accordingly and avoid preview dependencies in regulated paths.
Research digest
AI Research Briefing
Honest insights on AI agents, Small Language Models, and local RAG. No hype. Only when we have something worth sending.
- No hype, just measurable outcomes
- Read by 2,400+ engineers
- Unsubscribe anytime