ChatGPT Work + Salesforce: What the Integration Can (and Can't) Do in 2026
Partly — there’s no OpenAI-built Salesforce connector in ChatGPT Work’s official list, but Salesforce ships its own hosted MCP server (GA April 29, 2026) that you add to ChatGPT Work as a custom MCP connector — and it reads and writes. OpenAI’s native connectors cover Google Drive, Gmail, Outlook, SharePoint, GitHub, HubSpot, Linear, Notion and a handful more; Salesforce isn’t on that list. Instead you point ChatGPT Work at Salesforce’s own MCP endpoint (https://api.salesforce.com/platform/mcp/v1/<SERVER-NAME>), which exposes SObject CRUD — read, create, update, delete — plus Flows, Apex actions, and queries, all honoring your field-level security and sharing rules. Auth is per-user OAuth 2.0 with PKCE on a dedicated mcp_api scope, and every transaction runs as the authenticated user. The catch is the shape of the work: a ChatGPT Work agent run is a session you start, and it’s metered — nothing watches your pipeline between runs.
This is the enterprise-agent angle. For the plain consumer-ChatGPT setup, see our ChatGPT + Salesforce guide; here we’re focused on what ChatGPT Work’s multi-step agent runs can and can’t do across a connected Salesforce stack.
What ChatGPT Work can actually do with Salesforce
Once the Salesforce hosted MCP server is added as a custom connector, a ChatGPT Work agent can, inside one run:
- Query the pipeline in plain English. “Which enterprise opportunities have slipped their close date and haven’t had activity in ten days?” — answered from live CRM data, not an exported report.
- Read across objects. Accounts, contacts, opportunities, leads, cases, and your custom objects — whatever your permissions expose — pulled into the run for reasoning.
- Write back. The hosted MCP server supports create, update, and delete on SObjects, so an agent can advance a stage, log an activity, update a field, or create a follow-up task — always scoped to your CRUD, FLS, and sharing rules, and recorded in the audit trail as you.
- Trigger Flows and Apex. Beyond raw records, the server can expose Flows, Apex invocable actions, and Named Queries, so a run can kick off org logic you’ve already built rather than reimplementing it.
- Chain Salesforce with the rest of the stack. Because ChatGPT Work runs across every connector you’ve added, one agent run can read an opportunity from Salesforce, pull the thread from Gmail or Outlook, and draft the recap — a multi-step errand across the connected stack. Salesforce has also shipped an Agentforce Sales app inside ChatGPT (open beta) as a higher-level alternative to the raw MCP connector.
How to set it up
The Salesforce side is an admin job; the ChatGPT Work side is a custom connector:
- In Salesforce, enable a hosted MCP server for your org and create an External Client App for OAuth — the overview docs walk through the
mcp_apiscope and per-user PKCE flow. Start with the read-onlysobject-readsserver to test. - In ChatGPT Work, turn on developer mode and go to Settings → Apps → Create App. Name the connector (e.g.
Salesforce). - Enter the server URL:
https://api.salesforce.com/platform/mcp/v1/<SERVER-NAME>for production (or the.../sandbox/<SERVER-NAME>variant for a sandbox), and paste ChatGPT’s callback URL into your External Client App. Follow Salesforce’s ChatGPT configuration guide. - Sign in with your Salesforce credentials when prompted — OAuth, no shared API key. Then select Salesforce under the + button in the prompt and confirm with something read-only: “list my open opportunities and their stages.”
The limits that matter
- It doesn’t run on triggers. There’s no “when a lead lands, route and email it” or “when an opportunity stalls, nudge the rep.” A ChatGPT Work agent touches Salesforce only when you start a run. A deal can go cold over the weekend and nothing moves.
- Agent runs are session-bound and metered. ChatGPT Work runs are long and autonomous, but they’re manually started and counted against your plan’s allowance. Between runs, nothing is watching your funnel — it’s an errand you dispatch, not a standing watch.
- Not an OpenAI-native connector. Salesforce isn’t in ChatGPT Work’s official connector list, so you’re on the custom-MCP path — an admin has to stand up the Salesforce hosted server and the External Client App first. That’s more setup than clicking “connect Gmail.”
- Everything is a person in the driver’s seat. By design, every MCP transaction runs as the authenticated user and lands in the audit trail. That’s good governance — but it also means recurring rituals (the Monday pipeline digest, the post-call CRM update) only happen when someone re-runs them.
- Follow-through stops at the run. A ChatGPT Work agent can draft the recap email and stage the CRM update, but sending from your mailbox, booking the call, and posting to the deal channel as one unattended motion isn’t what a metered session is built for.
If you want Salesforce-adjacent work that runs on its own: Carly
The moment you want something to happen around Salesforce without you starting a run — a new lead routed and emailed within minutes, a nudge when an opportunity stalls in a stage, a Monday pipeline digest built and sent, a case escalation posted to the right Slack channel — you’ve crossed past what a metered agent session is for.
That’s where Carly fits. Carly is an AI executive assistant built to act on triggers, not just answer inside a run:
- Fires on events and schedules, 24/7, in the cloud. When an opportunity changes stage or a lead lands in Salesforce, Carly reacts — summarizes it, emails the owner, updates the record, posts to Slack — while your laptop is closed. No run to start.
- Reads and writes Salesforce natively. Salesforce is a native Carly integration, so Carly updates records, logs activity, and creates tasks directly — no External Client App to stand up first.
- Sends, not just drafts. Carly drafts and sends email across Gmail and Outlook, books meetings, manages tasks, and records meetings — the follow-through that stops at the session with ChatGPT Work.
- Builds the workflow by interviewing you. Tell Carly “when an opportunity moves to Negotiation, draft a recap email to the account owner and create a follow-up task” in plain English; it interviews you and builds it — no developer mode, no MCP URL, no prompt engineering.
Carly connects to 200+ tools across 40+ categories natively, plus any other tool via your own API key — paste it on carlyassistant.com/integrations. AI agents start at $35/month, and steps in a workflow that don’t use AI run free and unlimited. See integrations.
ChatGPT Work vs Carly
| ChatGPT Work (Salesforce MCP) | Carly | |
|---|---|---|
| Query pipeline and records | Yes, in a run | Yes |
| Update / create records in Salesforce | Yes, in a run (via Salesforce’s MCP server) | Yes, natively |
| Trigger Flows / Apex | Yes, if exposed | Via the integration |
| Reacts to a Salesforce event by itself | No | Yes, on the trigger |
| Monday pipeline digest, on schedule | No | Yes |
| Sends email as part of the flow | No (drafts only) | Yes (Gmail + Outlook) |
| Runs while laptop is closed | No (runs are started + metered) | Yes (cloud, 24/7) |
| Connector type | Custom MCP (Salesforce not OpenAI-native) | Native integration |
| Setup | Admin stands up MCP server + OAuth app | Describe it in plain English |
| Pricing | Paid ChatGPT Work plan (metered runs) | AI agents from $35/mo |
ChatGPT Work’s Salesforce connection is a CRM copilot you dispatch for a run. Carly is a teammate that acts on Salesforce events as they land.
Frequently Asked Questions
Does ChatGPT Work integrate with Salesforce?
Yes, through Salesforce’s own official hosted MCP server, which reached GA on April 29, 2026 — not through an OpenAI-native connector. You add it to ChatGPT Work as a custom MCP connector in developer mode, authenticate with OAuth, and can then read and write Salesforce records in plain English inside an agent run.
Can ChatGPT Work update records in Salesforce, or only read them?
Both. Salesforce’s hosted MCP server supports SObject create, update, and delete alongside reads, plus Flows and Apex actions — all scoped to your CRUD, field-level security, and sharing rules. Start with the read-only sobject-reads server to test, then move to a server that permits writes.
How do I connect ChatGPT Work to Salesforce?
An admin enables a Salesforce hosted MCP server and creates an External Client App for OAuth (mcp_api scope, per-user PKCE). In ChatGPT Work’s developer mode, go to Settings → Apps → Create App, enter the server URL https://api.salesforce.com/platform/mcp/v1/<SERVER-NAME>, and complete the OAuth sign-in. Salesforce’s ChatGPT configuration guide has the full steps.
Can ChatGPT Work react to a new Salesforce lead or a stalled deal automatically?
No. ChatGPT Work agent runs are session-driven and metered — they touch Salesforce when you start a run, not when a Salesforce event fires. For “when a lead lands, route it and email the owner” or “when a deal stalls, nudge the rep,” you need a trigger-based assistant like Carly, which integrates natively with Salesforce and runs in the cloud around the clock.
More: Best AI CRM tools · Best AI agents for productivity · ChatGPT + Salesforce · Claude + Salesforce
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
