Your CRM Now Takes Orders From ChatGPT, Claude and Gemini. Here Is What to Do With That

For two years the way to get an AI agent into your business was to build one. That is still the right answer for anything that runs on its own, but a quieter shift happened in August: the agents arrived inside the software you already use, switched on with a checkbox.

What shipped

• HubSpot. The August 2026 product update added a connector for Gemini and expanded the existing ChatGPT and Claude connectors so that all three can query and create contacts, companies, deals, tickets, notes, tasks and meetings. The Claude connector also exposes HubSpot through a remote MCP server. Availability: all hubs and all tiers, including the free one.
• Slack. On August 26 Slack announced Big Mode, a full-screen workspace for long research and writing inside Slack; Skill Sets, bundles of AI capabilities an admin can assign to a team; and Deep Research across company data and the web. At launch these are for Enterprise+ customers.
• Salesforce. The same day, Salesforce and Anthropic announced Claudeforce: Salesforce data and 37 prebuilt sales skills inside Claude, Claude as a reasoning engine inside Agentforce, and Claude as the default model in Slack. Pilot customers now, open beta expected in September.

For a small or mid-size company the HubSpot change is the one that matters, because it costs nothing and it works on the tier you are probably already on.

What it is good for on day one

Read-mostly questions, asked by a person, answered from your own data. These are the ones we would switch on first:

• "Which deals over five thousand dollars have had no activity in 14 days?"
• "Summarise everything we know about this company before my 2pm call."
• "Draft a follow-up to every contact from last week's webinar who has not replied, in our usual tone." (Draft. Not send.)
• "Which open tickets mention the same product issue?"

Each of these used to require a report, a filter, or a person who knew where to look. Now they are a sentence. That alone is worth an afternoon of setup.

Where it goes wrong

Write access is the new part, and it is where the mistakes will happen. Some are technical: the assistant creates a second record for a company that already exists with a slightly different name, or moves a deal to "Closed Won" because someone said "we won that one" in passing. Some are organisational: the assistant acts with the permissions of whoever is typing, so a junior team member's chat can edit the pipeline. And some are about data leaving: your CRM contents are now visible inside a third-party chat product, under that product's retention and training policies, which you should read before you turn it on.

The rollout we recommend

Alongside the phases, three habits. Name one person as the connector owner. Ask that person to skim the CRM's activity log for AI-created records once a week; it takes ten minutes. And write down which chat product is allowed to see the CRM, so a well-meaning colleague does not connect a fourth one.

When to graduate to a real workflow

A connector is a person asking an assistant to do something now. That is enormously useful for discovery, because within a fortnight you will notice the same requests being typed every day. Those are the workflows. "Summarise before every call" becomes a scheduled brief that appears in the calendar invite. "Which deals went quiet" becomes a Monday morning list with a suggested action beside each. "Draft the follow-up" becomes a queue that a person approves in batches. Each of those needs a trigger, an approval gate and a log, which is exactly what tools like n8n or Make provide and a chat window does not. We wrote a comparison of the three platforms if you want to go deeper.

The honest summary

Switch the connectors on. Keep them read-only for a fortnight. Watch what people ask. Then automate the three questions that come up every day, with a person between the agent and anything a customer sees. That sequence is boring, and it is the one that works.