Your CRM's AI assistant isn't working for you: how platform AI pulls your data in and makes it stickier

The AI built into HubSpot, Notion, Slack, Salesforce and Microsoft 365 is a customer-retention tool before it is a productivity tool. Its connectors pull your other data into the platform, its answers are only good about data it holds, and getting data back out is harder than getting it in. Use it when the platform is already your system of record. Don't let it become the reason it is.

I run my own practice on HubSpot and use its assistant most weeks, so this isn't a piece against bundled AI. It's about which way the connectors point. Every one of these assistants is sold as a way to get more from the data you already have. Every one of them is also, quite rationally from the vendor's side, a way to make sure more of your data ends up with them. Both things are true at once, and the second one is the part nobody puts on the pricing page.

Everything below was checked against the vendors' own documentation in late September 2026. These things change monthly, so treat the specifics as a snapshot and the pattern as the point.

What is the AI inside your existing tools actually for?

For the vendor, it's a moat. The analysts have started saying this out loud. Futurum's Keith Kirkpatrick wrote in February that SaaS companies "are unlikely to make it simple or inexpensive to migrate that data from their platform", and named "the use of data gravity as a competitive moat" as the strategy (Futurum). theCUBE Research's 2026 predictions put it more neutrally: "data gravity is pulling AI workloads toward existing database ecosystems, reinforcing incumbent platforms" (SiliconANGLE).

You can see the mechanism in the connector lists. Notion AI will search your Slack, Google Drive, Gmail, Jira, Microsoft Teams, SharePoint, OneDrive, GitHub, Linear, Outlook and calendars, once you're on the Business plan, and it does so by indexing them: "New content may take up to 3 hours to be indexed by Notion AI" (Notion). HubSpot's Breeze Assistant will "find relevant information from connected apps (e.g., Slack, Gmail)" and answer "regardless of where the data is stored" (HubSpot). Microsoft 365 Copilot offers "more than 100 prebuilt connectors" including Salesforce, ServiceNow, Google services, Jira and Confluence (Microsoft), with HubSpot and Notion themselves now in preview. Salesforce's Data 360 advertises "200+ connectors" and zero-copy links into Snowflake and Databricks (Salesforce). Slack notes that "data from connected sources, like Salesforce, will be included in Slackbot responses" (Slack).

Each of those is genuinely useful. Each also means the assistant gets better the more of your other systems it can see, and the platform gets harder to leave at exactly the same rate.

Why does the assistant reward you for moving more data in?

Because its answers are only as good as what it can reach, and the obvious fix for "it doesn't know about X" is to connect X.

That loop is the whole design. You ask the CRM's assistant about a client and it doesn't know about last week's call, because the notes are in your inbox. So you connect the inbox. Now it knows, and it's noticeably more useful, and a copy of your correspondence is indexed inside the CRM. Next week it's the shared drive. The month after, the ticketing system. None of these was a decision to make the CRM your system of record. It just became one, a connector at a time.

The pricing points the same way. Notion folded unlimited AI into its Business plan in May 2025, at $20 a member a month, with lower plans getting "a limited amount of trial usage" (Notion). Slack raised Business+ from $12.50 to $15 a user "to reflect the significant value added through new advanced AI capabilities" (Slack). HubSpot's agents need Professional or Enterprise and run on credits, at a cent a credit and fifty credits per resolved customer conversation (HubSpot). In every case the AI is the reason to be on the higher tier, and the higher tier is where your data lives.

What happened with Slack, and why does it matter?

Slack is the clearest example of the connectors pointing one way, so it's worth the detail.

In May 2025, Salesforce changed Slack's API terms. Apps not distributed through Slack's own marketplace were rate-limited to one request a minute on the methods that read message history, returning fifteen messages at a time, where an internal app can still pull a thousand (Slack). The current terms, reissued in October 2025, say that "you may not: (A) use API Data to train a large language model; (B) bulk export Slack message and file data except where expressly allowed by an additional agreement", that the Discovery API "may only be used for security and compliance use cases", and that apps "may not create persistent copies, archives, indexes, or long-term data stores" of the data they access (Slack). The search API Slack offers instead works, as Computerworld put it, "only from within Slack itself" (Computerworld).

The practical effect fell on the enterprise-search tools businesses had been using to search their own conversations from outside Slack. One of them, Glean, told its customers the change was "locking your data within Slack" and limiting results "to the scope and quality of Slack's search technology" (Hunton). At the same moment, Slack's own AI search, on the plan that had just gone up in price, was being extended to read Salesforce data.

Two fair points on Slack's behalf. There are real security reasons to stop third parties hoovering up message history, and Slack still lets you export your own workspace, though private channels and direct messages are "not available" on the Free and Pro plans and need an application on Business+ (Slack). But look at the direction. Salesforce data flows into Slack's assistant freely. Slack data flowing out to somebody else's assistant became slow, restricted and contractually fenced. That asymmetry is the pattern, and it's not unique to Slack.

Which way do the connectors point?

Here is the same test applied across the platforms an SME is most likely to be running, from their own documentation as of late September 2026. "In" means what the assistant will read from elsewhere. "Out" means how you get your data back, and whether another AI tool can read it.

Sources for the table: HubSpot export, HubSpot MCP, Notion export, Notion MCP, Notion models, Salesforce models, Salesforce MCP, Salesforce export, Microsoft models, Microsoft MCP.

Two things stand out. The "in" column is long everywhere and getting longer. The "out" column is where the caveats live: the export that leaves out the activities, the format you can't restore from, the tier you have to be on, the application you have to make. And the far column is the good news. HubSpot, Notion and Salesforce all shipped official MCP servers this year, which means another assistant, one you chose, can read what they hold. That's the connectors pointing outward for once, and it's worth rewarding.

Isn't the law fixing this?

In the EU, partly. In the UK, not really, and not yet.

The EU Data Act has applied since 12 September 2025. Providers of platform and software services "must make open interfaces available and, at a minimum, export data in a commonly used and machine-readable format", and switching charges, including data egress fees, are to be removed entirely from 12 January 2027 (European Commission). It applies to non-EU providers serving EU customers (Latham & Watkins), which is why several of the export improvements above appeared when they did.

The UK's Data (Use and Access) Act 2025 sounds similar and isn't. Its data-access provisions commenced in August 2025, but they "confer powers on the Secretary of State and the Treasury to establish Smart Data schemes through regulations", sector by sector (CMS). There is no general right for a UK business to get its data out of a SaaS platform. What UK customers get is the spillover from vendors building for EU rules. Useful, but not something to rely on in a renewal negotiation.

When is the bundled AI the right call?

It's the right call when the platform is already your system of record, chosen for reasons that had nothing to do with AI. If HubSpot is where your customers live, an assistant that reads HubSpot is cheap, well integrated and probably better than anything you'd bolt on. It's the right call when the data it needs is data the platform already holds. And it's the right call when the alternative was buying a separate tool to do the same job, which just moves the gravity somewhere else.

The problem starts at the connector, not the assistant. The test I'd apply to each one is simple: would you still connect your inbox to this platform if its AI didn't exist? If the answer is no, the connector isn't a productivity decision. It's a decision about where your data lives, and it deserves to be made as one.

This is a different argument from the one I made about single-platform assistants, which was about choosing Copilot or ChatGPT as your company's assistant. Here the assistant came free with a tool you'd already bought for other reasons, and the lock-in arrives through the side door. It's the practical case for the four tests of vendor-agnostic AI, applied to software you weren't thinking of as AI software at all.

What should you check before you connect anything?

Six things, most of which take an afternoon.

1. Run the export. Not read about it. Run it, and look at what's missing. Do the notes come out? The comments? The history? Could you restore from it, or is it a pile of files?
2. Read the API terms as if you were a competitor. The Slack precedent is the one to keep in mind: what a third-party tool can read, how fast, and under what conditions. If an assistant you didn't buy from them can't read your data at full speed, the platform has decided who gets to be intelligent about it.
3. Check for MCP or an equivalent open interface. HubSpot, Notion and Salesforce have one. It's the single best signal that the vendor is comfortable with your data being read by something it doesn't own.
4. Ask whether you can bring your own model. Salesforce yes, HubSpot no, Notion a menu rather than a key, Microsoft only in its agent-building tool. This matters less for the assistant's quality than for what it tells you about where the control sits.
5. Note which connectors copy and which read in place. Notion indexes; Salesforce's zero-copy links don't. A copy inside the platform is a second system of record you now have to govern.
6. Watch for the AI as the reason for the tier. If the assistant is why you're on Business rather than Plus, or Professional rather than Starter, the assistant has become part of the switching cost. Price that in at renewal.

The alternative to letting one platform absorb the others is the one I argued for in build vs buy vs integrate: a thin layer of your own, an orchestration tool or an AI gateway, that reads from all of your systems and lets you choose the model. It's more work than clicking "connect", and it means the intelligence about your business belongs to your business. Given how the last eighteen months have gone, I'd want that to be true before the next renewal, not after.

The bottom line

The AI in your CRM, your wiki and your chat tool is useful, and it is also the most effective retention feature those vendors have ever shipped. Every connector makes the assistant a little better and the platform a little harder to leave, and the export button rarely keeps pace with the connect button. Use the bundled AI where the platform is already your system of record. Before you connect anything else, run the export, read the API terms, look for an open interface, and ask whether you'd make the connection if the AI weren't there.

If you're looking at a platform's AI features and can't tell whether they're working for you or on you, a thirty-minute discovery call will get you a straight answer. See AI consulting for how I approach the question, and process automation for the "integrate instead" version.