ChatGPT CRM Integration: Salesforce, HubSpot and Custom APIs

ChatGPT connects to a CRM in three ways: a vendor-published plugin or MCP connection (HubSpot documents a connection built on its own MCP server and Salesforce an Agentforce Sales integration in open beta since December 17, 2025; neither is named on OpenAI's pages as of October 17, 2026, so confirm either in your Plugin Directory first), a custom MCP server or plugin that your team builds against the CRM's API, or a separate application on OpenAI's API that reads and writes records under its own rules. A vendor-published connection is the fastest route and runs under each user's existing CRM permissions; an API build gives the most control and can run without a person in the chat. Whichever route you choose, the CRM stays the system of record, ChatGPT drafts, and a person approves anything that writes.

This guide compares the routes and the controls around them: permissions, read and write scope, data boundaries and ownership. For what sales follow-up automation is worth, read AI sales follow-up automation for small business; this page is about the connection itself.

What can ChatGPT do with a CRM?

In practice, five jobs: a pre-call brief on an account from its records and recent activity; pipeline questions (which deals have not moved in 30 days); drafting follow-up emails and call notes in the company's voice; logging a meeting as a note or task; and data hygiene, such as duplicate contacts or deals missing a close date. The boundary to set first: the CRM holds the truth, ChatGPT proposes, and a named person approves changes to records, outbound messages and anything that touches money.

Three ways to connect ChatGPT to a CRM

A fourth route, a GPT Action in a custom GPT, still works but is ending: OpenAI retires custom GPTs on December 11, 2026 and actions do not migrate to plugins. Our guide to custom GPT Actions and safe integrations covers the transition. Note also that HubSpot and Salesforce do not appear by name on the OpenAI pages we opened on October 17, 2026; the facts below come from the vendors' own documentation, and the fallback for any CRM is a custom MCP server or an API build.

HubSpot and ChatGPT: what HubSpot's own documentation describes

HubSpot's knowledge base article on connecting a HubSpot account to ChatGPT, updated September 28, 2026, describes a HubSpot-maintained connection that runs on HubSpot's MCP server and is available to HubSpot customers on all plans with an OpenAI account. The facts a buyer needs, as of October 2026:

• Who can connect. Super Admins and users with App Marketplace Access permissions can connect without prior approval; everyone else needs a Super Admin to approve the connection first and choose the data permissions and which users may install it.
• Setup. In ChatGPT settings, choose HubSpot and Connect (HubSpot's steps still say Apps, now Plugins); sign in to HubSpot, choose the account, select the permissions and confirm. Users see only what their HubSpot permissions allow, and when a Super Admin adds objects, each user must reconnect.
• Read and write. Read covers contacts, companies, deals, tickets, quotes, invoices, engagements and more; create and update covers many CRM objects, marketing assets and engagements (calls, meetings, notes, tasks, emails); no delete access is listed; some objects are read-only. HubSpot recommends setting write tools to "Needs Approval" so edits are confirmed with a yes or no in chat.
• Limits. Bulk creates and updates handle 10 records at a time; custom validation rules, including pipeline stage rules, are not applied; with Sensitive Data turned on, ChatGPT cannot access engagement data.
• Audit and data. Write actions appear in the HubSpot audit log attributed to both the user and the connection; EU-hosted accounts route through EU servers; HubSpot states that ChatGPT Enterprise, Business and Education data is not used for training.

The validation-rule gap is the one to plan around: a stage change ChatGPT makes is not checked by the rules sales operations built, so keep write tools on approval and spot-check the audit log weekly.

Salesforce and ChatGPT: what Salesforce documents

Salesforce's newsroom announced an open beta of what it calls the Agentforce Sales app in ChatGPT on December 17, 2025 for Agentforce for Sales Add-on and Agentforce 1 Edition customers, listed at the time in ChatGPT's app directory (replaced by the Plugin Directory on July 9, 2026). Salesforce says it respects existing permissions, accesses only information the user already has access to, and handles data through the Agentforce Trust Layer under Salesforce security standards. The announcement shows it pulling a live list of uncontacted leads, updating an opportunity to Closed Won in conversation, and generating account plans saved back to Salesforce. A Salesforce blog post dated September 15, 2026 says hundreds of organizations use it and that access to records is governed by existing Salesforce permissions. OpenAI's pages do not name it and Salesforce's help pages still carried the beta label when we checked on October 17, 2026, so confirm status and edition requirements with Salesforce before you budget.

For a Salesforce administrator the checklist is familiar: confirm the edition and add-on, test in a sandbox first (Salesforce documents a sandbox test for this integration), decide which profiles may connect, and keep ChatGPT's approval mode at "Always ask" for writes.

Zoho CRM and other systems

If your CRM is Zoho, Pipedrive or an industry system, look for a vendor plugin in the ChatGPT Plugin Directory first and read its permission page as above. OpenAI's own admin article names sales-adjacent apps such as Apollo.io, ZoomInfo, Monday.com and Airtable among those it can restrict to company domains, a way to confirm an app exists. If there is none for your CRM, the integration is a custom MCP server or an API build against the CRM's REST API, with the same gateway, least-privilege credential and approval design, the route that does not depend on a vendor's roadmap.

How do you keep data boundaries?

• Decide what leaves the CRM. Vendor-published connections expose what the user's permissions allow; a custom MCP server or gateway can allow-list fields, which is the only way to keep notes, pricing exceptions or health information out of the conversation.
• Know the training and retention terms. OpenAI's enterprise privacy page says that by default it does not train its models on data accessed from apps and that ChatGPT Business and Enterprise data is not used for training by default; workspace admins control retention, and deleted conversations are removed within 30 days. Enterprise and Edu workspaces can feed conversation logs to compliance tools through the Compliance Platform.
• Control who connects. OpenAI's admin controls for plugins and apps say apps are enabled by default in ChatGPT Business, so disable what you have not reviewed; set the workspace approval default ("Always ask" or "Allow read actions") and per-app overrides; Enterprise can ask OpenAI to block company-domain accounts from connecting apps outside authorized workspaces.
• Keep the CRM's own log. HubSpot's audit log and Salesforce's permission model are your evidence of who changed what; the integration should never write under a shared service login when a user's identity is available.

Record matching, read-only pilots and write approvals

1. Start read-only. Briefs, pipeline questions and drafts for two to four weeks, with the vendor connection's write tools disabled or on approval.
2. Test record matching. Ask about accounts with similar names, duplicate contacts and closed deals; note where ChatGPT picks the wrong record, because a write would land there too.
3. Define the first write. Usually logging a note or task after a call. Keep stage changes, amount changes and outbound email for later.
4. Require confirmation. "Always ask" in ChatGPT and "Needs Approval" on HubSpot write tools; the confirmation text should name the record and the field.
5. Review the audit log weekly for the first month, then monthly, and compare writes against the approvals.
6. Add the next write only after the log shows the first one is accurate.
7. Write the owner down, with the reconnection steps for when a vendor adds objects or scopes.

Failure handling, testing and ownership

CRM integrations fail in three ways: a revoked or expired token (nothing works), a vendor change in objects or scopes (some tools disappear until users reconnect), and a quiet wrong match (a note lands on the wrong account). The first two show up as errors; the third only in the audit log and sampled conversations. Our guide to AI agent monitoring and maintenance turns this into a monthly checklist with an owner, a kill switch and a cost line.

How NetSys scopes a CRM integration

Our OpenAI and ChatGPT integration service starts with the vendor route when one exists: we review the vendor's permission page with your CRM administrator, set the ChatGPT approval modes, run the read-only pilot and the record-matching tests, and document who may connect. Where no vendor connection fits, we build the MCP server or gateway with field allow-lists and per-user identity, and hand over the permissions list and the audit routine. Automations that must run without a person are designed as agents with an approval queue under our AI workflow automation service.

Scope a CRM integration with an engineer, starting from one job your sales team repeats every day.

Frequently asked questions

Does ChatGPT integrate with HubSpot?

HubSpot documents its own connection to ChatGPT, built on HubSpot's MCP server, for customers on all its plans: a Super Admin approves it, users connect with their own HubSpot login, and it can read and, with approval, create and update records. HubSpot's article is the reference, since OpenAI's pages do not name it; if it is not available in your workspace, a custom MCP server or API build is the route.

Does ChatGPT integrate with Salesforce?

Salesforce documents an Agentforce Sales integration for ChatGPT, in open beta since December 17, 2025 for Agentforce for Sales Add-on and Agentforce 1 Edition customers, that respects existing Salesforce permissions and can update records in conversation. OpenAI's pages do not name it, so confirm current status and edition requirements with Salesforce; a custom MCP server or API build covers other editions.

Can ChatGPT update CRM records?

Through a vendor-published connection with write tools enabled, or through a custom MCP server or API build, yes. Keep writes behind ChatGPT's "Always ask" approval and the vendor's own approval setting, start with low-risk writes such as notes and tasks, and review the CRM audit log.

Is our CRM data used to train ChatGPT?

OpenAI says that by default it does not train on data accessed from apps and does not train on ChatGPT Business or Enterprise workspace data. Personal Free, Go, Plus and Pro accounts have training on unless the user opts out, which is one reason to connect a CRM only from a company workspace.

Which ChatGPT plan do we need for a CRM integration?

A workspace plan. ChatGPT Business (formerly Team) gives admins app enablement and approval settings and lets admins and owners publish custom MCP servers; ChatGPT Enterprise adds role-based app assignment, the Compliance Platform and the verified-domain restriction. Vendor connections may add their own plan conditions, as Salesforce does.

What is the difference between a vendor-published connection, a custom MCP server and an API integration?

Who built it and where it runs. A vendor-published connection is maintained by the CRM vendor and used in chat under each user's login; a custom MCP server is built by you or your provider, exposes only the tools you choose and is published by a workspace admin; an API integration runs outside ChatGPT in your own systems, billed per token, without anyone typing.