twimco.Let’s talk ↗︎
Menu
Field notes / AI & judgment

HubSpot now lets AI connectors read Sensitive Data. Hold a thirty-minute meeting before anyone opts in

Until last week, HubSpot kept the fields you marked sensitive away from AI connectors. A Super Admin can now open them up, read-only, and the hardest part of that decision has nothing to do with settings.

Editorial engraving of a bright textile-mill meeting room where four colleagues work through a short printed agenda beside a laptop with one setting switched off.

HubSpot's October 30, 2026 rollup lets Super Admins allow AI connectors to read Sensitive Data, in public beta. ChatGPT, Claude and Copilot can read Sensitive Data properties and engagements; Gemini and the Remote MCP server get engagements only. Access is read-only, user permissions still apply and Highly Sensitive Data stays blocked. Before opting in, agree which connectors, which people, which properties and who reviews it, because the data leaves HubSpot's screens and lands in a chat.

What changed on October 30

HubSpot lets accounts mark some fields as Sensitive Data: the information your business has decided needs more care than an email address or a deal stage. Until now, AI connectors couldn't read those properties or the engagements tied to them. HubSpot's October 2026 developer rollup changes that, in public beta.

A Super Admin can now allow supported connectors to read Sensitive Data. ChatGPT, Claude and Copilot can read Sensitive Data properties and engagements. Gemini and HubSpot's Remote MCP server can read engagements only. Access is read-only. Users still need the right CRM object or property permissions to see anything through a connector, and properties marked Highly Sensitive stay off limits. If a connector is already connected, it has to be disconnected on both sides and reauthenticated, and new permissions can take up to two hours to show up.

That's a careful design. It's also a decision that a single Super Admin can now make on a Tuesday afternoon because a sales manager asked for it. We'd rather it was made in a short meeting with the right people in the room. Here's the agenda we'd use.

Decision guide

What each connector can read in the beta

ConsiderationConnectorSensitive Data reach
ChatGPT, Claude, CopilotSupported AI connectors.Sensitive properties and engagements, read-only
GeminiSupported AI connector.Engagements only, read-only
Remote MCP serverRoute for custom agents and tools.Engagements only, read-only
Any connectorHighly Sensitive Data properties.Still restricted
Source: HubSpot October 2026 developer rollup, published October 30, 2026.

Item 1: why do we want this? (five minutes)

Start with the job, not the switch. Someone wants a chat assistant to see sensitive fields for a reason. Get them to say it in one sentence.

"I want to ask Claude for a summary of a customer's history before a renewal call" is a real job. "It would be nice if the AI knew everything" is not. The first tells you which properties matter and who needs them. The second is how you end up granting everything to everyone.

It also helps to say who the answer is for. A summary that an account manager reads before a call is a different risk from a summary that gets pasted into a proposal, or one that an internal agent uses to decide what to send next. Write the job and the reader on the agenda, at the top, where everyone can see them for the rest of the meeting.

If nobody can name a specific job, stop the meeting here. You haven't lost anything; the beta will still be there next month.

Sequence of work

Thirty minutes before the opt-in

  1. 0-5 min

    Name the job in one sentence.

  2. 5-10 min

    Pick the fewest connectors that do it.

  3. 10-20 min

    Trim who sees which properties.

  4. 20-25 min

    Write the team's rules.

  5. 25-30 min

    Name an owner and a review date.

Illustrative decision framework for a Super Admin and the team.
Read-only describes what the connector can do to HubSpot. It says nothing about what happens to the answer afterwards.

Item 2: which connectors, and only those (five minutes)

The rollup is clear that different connectors get different reach. ChatGPT, Claude and Copilot can read both sensitive properties and engagements. Gemini and the Remote MCP server get engagements only.

Pick the smallest set that does the job from item 1. If your team already standardizes on one assistant, allow that one. Every additional connector is another company's product holding a copy of the answer, another set of settings to audit, and another reauthentication when something changes.

The Remote MCP server deserves its own line on the agenda. It's how custom agents and internal tools tend to reach HubSpot. If someone on your team has built an agent on top of it, ask what that agent does with engagement history before you widen its access. An agent that reads engagements to draft a follow-up is reading call notes, emails and meeting logs, which is often where the most sensitive details end up anyway, typed in by a rep in a hurry. Ask whoever built it to show you one real example of what the agent receives and what it produces from it.

Item 3: who, and which properties (ten minutes)

This is the longest item for a reason. HubSpot says connector access follows the user's existing CRM permissions. That's good, but it means your permissions are now doing double duty. Anyone who can see a sensitive property on a record can now, if you opt in, pull it into a chat window.

So look at your permissions with that in mind:

  • Who can see each Sensitive Data property today, and does every one of them need to?
  • Are there shared or generic logins with broad access? Those are worth closing regardless.
  • Which properties are marked Sensitive but should really be Highly Sensitive, which stays blocked?

Read-only describes what the connector can do to HubSpot. It says nothing about what happens to the answer afterwards. Once a summary is in a chat, it can be pasted into an email, saved in a document, or shared in a workspace that has its own retention rules. Your HubSpot permissions don't follow it there.

We made a similar point when Shopify opened its admin to AI connectors: the permission screen is the start of the decision, not the whole of it.

Item 4: what we'll tell the team (five minutes)

If you go ahead, write down two or three plain rules and send them to everyone who has access. For example:

  • Use the connector for summaries and preparation, not for copying sensitive fields into documents.
  • Don't paste connector answers that contain sensitive details into shared channels.
  • If an answer looks wrong or includes something you didn't expect to see, tell the admin.

The last rule matters more than it looks. People are the best early warning you'll get that permissions are wider than intended.

Item 5: who reviews it, and when (five minutes)

Name one person who owns the decision. Put a date on the calendar, thirty days out, to look again: is anyone using it, for the job you agreed, and has anything surprised you? Keep a short note of what you turned on and why, so the next admin doesn't have to guess.

The review doesn't need to be elaborate. Ask three or four people who use the connector what they've asked it this month, and look at whether the answers needed the sensitive fields at all. If nobody's job got easier, the right move is to switch it back off and close the meeting notes with a sentence about why.

Also decide what would make you turn it off. A connector you can't explain, a property that showed up where it shouldn't have, or a change in the beta's terms are all reasonable triggers.

That's thirty minutes. It's slower than flipping the switch, and it means the switch, when you flip it, is one you can defend.

If you're building agents or assistants on top of HubSpot data and want the permissions designed with the same care as the prompts, our AI applications and agents team does that work. Get in touch and tell us what job you want the assistant to do.

Related work: Sutton Place.

twimco.

We build alongside the people doing the work.
Bring us the problem you’re thinking about.

Related work

Sutton Place ↗︎