An agent only knows what you give it.

Your CRM wants to build you a team of AI agents. Not someday — this week. HubSpot has rolled out Agent Hub and Agent Builder in public beta. You type what you want in plain English, and an agent gets to work on the customer data you already have. No consultants. No long rollout.

For a growing company, that's genuinely significant. Work that used to require a data team and a real budget is now a button.

And look — that's useful. We're not here to rain on it.

But it got us curious.

What "shared context" leaves out

These agents run on what HubSpot calls shared context — a tidy way of saying the stuff already in your CRM. Which skips right past the more interesting question:

Because your invoices? They live in your billing system. Your costs live in your ERP. Your shipments live in whatever runs your trucks and warehouses. Your contracts, your service history, your margins — all living in systems the CRM has never touched.

Can the agents connect to some of that? Sure. Modern platforms support third-party connections, and the integrations keep getting better. But "connected" is a choice. And somebody has to make it. Somebody has to decide which systems matter, which fields are authoritative, and what happens when two systems disagree about the same customer.

An agent only knows what you give it. So — who's deciding what that is?

Three questions to ask around your own table

Before anyone hits the shiny new button, work through these three. Out loud, with the people who'd have to live with the answers.

Three yeses? Great. The agents will speed up what's already working, and you'll see it in the numbers inside a quarter.

Anything less, and they'll just help you do the wrong things faster. That's the part nobody puts in the launch post. Automation is an amplifier — it does not care whether the thing it's amplifying is correct. A misaligned definition of a good customer, run at machine speed across your whole pipeline, doesn't stay a small problem for long.

The hard part was never the agent

Notice that none of the three questions are technical. None of them are answered by the platform, the model, or the vendor. They're questions about definition, ownership, and traceability — the foundation the deployment sits on.

That foundation is the work. Deciding what the agent should know, and who owns what it does. Building the agent was never the hard part — and it's getting easier every quarter. The gap between companies getting value from this and companies collecting pilots isn't tooling. It's whether the answers to those three questions existed before the tool showed up.

Foundation first. Then deploy.

So we're curious — which of the three would slow your table down? For most teams we sit with, it's number three.