On this page (15)
Salesforce now brings CRM data and actions into the major AI assistants rather than asking everyone to work inside Salesforce. Sellers can work with Salesforce from ChatGPT through the Agentforce Sales app, from Gemini Enterprise through Agentforce Sales agents, and from Claude through Claudeforce's Salesforce in Claude plugin; developers can also connect any of them through Salesforce's hosted MCP servers and APIs. In each case Salesforce says the assistant sees only what the user can already see in Salesforce. Most of these surfaces were still in beta in 2026, so check status and licensing for your org before planning a rollout.
What connects to what today?
| Assistant | How Salesforce connects | Status and requirements |
|---|---|---|
| ChatGPT | Agentforce Sales app in ChatGPT: view and prioritize leads, update records and hand tasks to agents from the conversation | Open beta since December 2025, for Agentforce for Sales add-on and Agentforce 1 Edition customers; Salesforce has also said Agentforce Coworker will come to ChatGPT |
| Gemini Enterprise | Agentforce Sales agents inside Gemini Enterprise for lead engagement, meeting briefs, deal risks and CRM updates | Open beta announced April 2026; Salesforce and Google also plan zero-copy access to Google lakehouse data |
| Claude | Claudeforce: the Salesforce in Claude plugin with prebuilt sales skills, plus Claude as a reasoning model inside Agentforce | Announced August 2026; Salesforce listed Claudeforce in beta at Dreamforce in September 2026 |
| Any assistant or custom app | Salesforce-hosted MCP servers and APIs, running as a signed-in user or integration user | Hosted MCP servers available for Enterprise Edition orgs and above; check terms for your org |
| Microsoft Teams | Agentforce Coworker in Teams, alongside Microsoft's own Copilot Studio agents | Announced by Salesforce for later in 2026 |
Do permissions carry over?
Salesforce's position across these integrations is consistent: the assistant acts as the signed-in user, sees only records that user can access, and routes changes through Salesforce so validation rules and sharing still apply. That makes your Salesforce access model the main control over what any assistant can read or change. If a profile or sharing rule is too broad today, every connected assistant inherits the problem.
How should you choose which assistant a team uses?
- Start from where the team already works. A company standardized on Google Workspace will get more use from Gemini Enterprise; one where sellers live in ChatGPT or Claude should start there.
- Check licensing first. The ChatGPT and Gemini integrations were tied to specific Agentforce licenses, and Claudeforce has its own terms.
- Pick one or two workflows, such as meeting prep or pipeline updates, and compare accuracy and time saved before rolling out widely.
- Avoid running every assistant for the same task; it splits training, governance and support.
- Confirm each vendor's data handling and retention terms with security and legal before connecting customer data.
What should IT and security review?
- Which Salesforce users and profiles may connect each assistant.
- Which actions can write to Salesforce, and whether any need approval.
- How assistant activity appears in Salesforce audit and event logs.
- Data residency, retention and model-training terms in each AI vendor's enterprise agreement.
- How access is removed when someone leaves or changes role.
A rollout plan that keeps security comfortable
- Start with one team and one read-only use, such as summarizing accounts or drafting follow-up emails.
- Connect as the signed-in user or a dedicated integration user, never an administrator account.
- Write down what data the assistant may see and what it must never see, and test both.
- Log usage and review a sample of outputs every week for the first month.
- Add write actions, such as updating records, only after the read-only phase has held up.
- Expand team by team, with a short training session on what to check before acting on an answer.
Which use cases should come first?
| Use case | Why it is a good start | What to watch |
|---|---|---|
| Meeting preparation | Read-only, high value, easy to check | Answers drawing on stale or incomplete account data |
| Account and opportunity summaries | Saves time across many records | Summaries that hide important exceptions |
| Follow-up email drafts | Rep reviews every draft before sending | Tone and accuracy of commitments |
| Pipeline review questions | Helps managers find risks quickly | Consistency with the official forecast |
| Record updates from notes | Reduces manual entry | Start only after read-only use is trusted, with clear rules for what may change |
What policy should cover all assistants?
- Which assistants are approved, for which teams and which tasks.
- What data may be used, and what must stay out, such as regulated or contract-restricted information.
- Which actions require a person to review before anything is sent or changed.
- How activity is logged and who reviews it.
- How access is granted, reviewed and removed when roles change.
How do you measure whether it helps?
Pick one measure per use case before rollout: time spent preparing for meetings, time to log notes after calls, or the share of opportunities with complete next steps. Measure for a team that uses the assistant and one that does not yet, then decide whether to expand.
Frequently asked questions
Can different teams use different assistants?
Yes, and it is common. The control that matters is the same for each: the assistant should act with the permissions of the person using it, and access should be reviewed when people change roles. Keep one written policy that covers all of them rather than one per tool.
Will the assistant see records the user cannot?
According to Salesforce, these integrations act as the signed-in user and see only what that user can access. Test this with a few accounts and profiles before rollout rather than relying on it untested.
Should we wait until these integrations leave beta?
For broad rollout, often yes. A small pilot with one team and read-only use is a reasonable way to learn now without committing the whole company.
How does this relate to Agentforce?
Assistants help individuals work with CRM data in the tools they already use. Agentforce agents run inside Salesforce to handle defined jobs, such as answering service questions or qualifying leads, with their own instructions and guardrails. Many companies will use both.
A 90-day pilot plan
| Period | Focus | Decision at the end |
|---|---|---|
| Days 1–30 | Permission review, written policy, one team, one read-only use | Is the output accurate and useful enough to keep? |
| Days 31–60 | A second read-only use and weekly output reviews | Do people use it without prompting, and does the measure move? |
| Days 61–90 | Consider one write action with review, plan the next team | Expand, adjust or stop, with evidence either way |
Red flags
- Connecting assistants before reviewing who can see what in Salesforce.
- Several teams adopting different assistants with no shared policy.
- Write actions enabled on day one.
- No one reviewing outputs after the first week.
- Licensing and data terms not checked with legal and security.
These integrations are moving quickly, so treat any rollout as a series of small, reversible steps. The permission work you do in Salesforce now protects you whichever assistants your teams use later.

