On this page (14)
Agentforce and Microsoft Copilot Studio both let you build AI agents, but they live in different places. Agentforce runs inside Salesforce and works on CRM records with Salesforce permissions, so it fits customer-facing service, sales and CRM workflows. Copilot Studio is Microsoft's low-code agent builder for Microsoft 365, Teams and the Power Platform, so it fits employee productivity and workflows built on Microsoft data such as SharePoint, Outlook and Dataverse, though it can also publish customer-facing agents to websites and messaging channels. Companies that run both platforms rarely choose one; they decide which agent owns which job and how the two share data without bypassing either system's access rules.
How do they compare?
| Question | Agentforce | Microsoft Copilot Studio |
|---|---|---|
| Where it runs | Inside Salesforce, including service channels, the console, Slack and Experience Cloud sites | Microsoft 365, Teams, websites and Power Platform |
| Primary data | CRM records, Data 360 and Salesforce knowledge | SharePoint, OneDrive, Outlook, Dataverse and connected sources |
| Who builds | Salesforce admins and developers with Agentforce Builder, Agent Script and flows | Business makers and Power Platform developers |
| Actions | Flows, Apex, CRM actions and connected APIs | Power Automate flows, connectors and APIs |
| Access model | Salesforce profiles, permission sets and sharing | Microsoft Entra ID permissions and the source systems' sharing |
| Best first uses | Case handling, lead response, account summaries, service self-help | Internal knowledge questions, request intake, document-heavy employee tasks |
When should Agentforce own the job?
- The agent talks to customers or partners through service channels, portals or messaging.
- The work reads or updates Salesforce records such as cases, leads, opportunities or accounts.
- Answers must respect Salesforce sharing rules, for example which advisor or territory may see an account.
- Sales and service teams already work in Salesforce or Slack all day.
When should Copilot Studio own the job?
- The users are employees working mainly in Teams, Outlook and Office documents.
- The knowledge lives in SharePoint or other Microsoft sources rather than the CRM.
- The workflow already runs on Power Automate or Dataverse.
- The agent's job is internal: HR and IT questions, policy lookups, request intake.
How do companies run both?
The common split puts customer-facing and CRM work in Agentforce and employee productivity in Copilot Studio. Where an employee in Teams needs CRM context, connect the two through governed integration, such as APIs or connectors that call Salesforce as the signed-in user, rather than syncing records into Microsoft storage. Keep one owner for each agent, one place where its instructions live and one log of what it did.
What governance applies to both?
- Clean up over-broad access first: SharePoint sites shared with everyone and Salesforce profiles with View All both turn into AI exposure.
- Define which data each agent may use and which actions it may take, in writing.
- Keep a person approving outbound customer messages until accuracy is proven.
- Test with real questions and edge cases before launch, and monitor answers after.
- Measure a specific outcome, such as time to resolution or request volume handled, against a baseline.
We work across both ecosystems: Agentforce and Claude on the Salesforce side, and Copilot Studio and Microsoft 365 Copilot on the Microsoft side. The useful first step is a short inventory of the jobs you want agents to do and where the data for each one lives.
Questions to settle before you build in either
- Which system holds the data and records the agent must act on?
- Who will own the agent after launch, and in which team?
- How will you test it against real conversations before anyone relies on it?
- What should always be handed to a person, and how does that handoff work?
- How will usage be metered and paid for as volume grows?
Answering these usually settles the platform choice: build the agent where its data and its owners already live.
How do you plan a first agent?
- Pick one job with a clear measure, such as answering a common service question or triaging internal requests.
- List the data and actions the agent needs, and confirm the right people can already access them.
- Write the agent’s instructions, boundaries and handoff rules in plain language.
- Build a test set of real questions, including awkward and out-of-scope ones.
- Launch to a small group with a person reviewing outputs, then widen as accuracy holds.
- Monitor answers, escalations and the outcome measure every week.
| Check | Agentforce | Copilot Studio |
|---|---|---|
| Access cleanup | Profiles, permission sets and sharing reviewed | SharePoint and Teams sharing reviewed |
| Knowledge quality | Salesforce knowledge articles current and approved | SharePoint and document sources current |
| Actions | Flows and APIs tested with the agent’s permissions | Power Automate flows tested with the agent’s connections |
| Handoff | Routing to service queues or owners | Routing to the right internal team |
| Monitoring | Conversation review and outcome reporting | Usage analytics and conversation review |
Frequently asked questions
Can one agent work across both platforms?
An agent should live in one platform and reach the other through governed integration that respects the user’s permissions. Building the same agent twice, or copying data between platforms so one agent can see it, creates gaps in access control and audit.
Which is cheaper to run?
It depends on licensing you already own and how usage is metered for your volume. Estimate cost per conversation or per action for your expected volume on each platform before committing.
How long does a first agent take?
A narrow agent with good knowledge and clean permissions can reach a pilot in weeks. Most of the time goes into preparing data and testing, not building.
Do we need a partner for each platform?
Not necessarily. What matters is that whoever builds understands both access models and designs the boundary between them, so neither agent bypasses the other system’s controls.
A 90-day plan for a first agent
| Period | Focus | Decision at the end |
|---|---|---|
| Days 1–30 | Pick the job, clean up access, prepare knowledge, write instructions and tests | Is the data good enough to start? |
| Days 31–60 | Build, test against real questions, pilot with a small group and human review | Is accuracy high enough to widen? |
| Days 61–90 | Widen access, monitor outcomes, decide the next job | Expand, adjust or stop, based on the outcome measure |
Red flags
- Copying CRM data into SharePoint, or the reverse, so an agent can see it.
- An agent with no named owner.
- No test set of real questions before launch.
- Customer-facing answers sent without review before accuracy is proven.
- Usage costs not estimated for expected volume.
Most companies end up using both platforms. The work that matters is deciding which jobs belong where, keeping each agent close to the data it governs, and measuring outcomes rather than activity.

