You're arguing about the wrong layer

Businesses are running bake-offs about which AI chat assistant is best. But it's the tip of the iceberg.

Copilot vs. Claude vs. ChatGPT — which drafts the better email, has the better Cowork assistant, etc. Real questions. But they're about the smallest layer.

Listen to this post on Substack.

Earlier this week I posted — when a company calls enterprise AI "a risk," they usually mean "we haven't done the paperwork yet" — the boring procurement stuff that's been around for decades.

So to minimize "risk," Microsoft shops reach for Copilot, Google shops for Gemini.

But whichever one wins, it's still just the part above the water.

The 90% below the waterline is your data — trapped in a dozen systems that don't talk to each other. And the tool for THAT layer isn't a chat assistant at all.

Which brings me to a tool most teams haven't heard enough about: Databricks.

Here's the part that goes right back to "risk." Pick your walled garden — Microsoft and Azure, Google Workspace and GCP, AWS. Whichever one you've standardized on, Databricks already runs inside it: on your existing cloud bill, under paper you've already signed. Nothing new to onboard.

But the real reason I keep pointing teams there: it makes two feuding camps happy at once.

Business users get AI with the context of your real, joined-up business — not a chatbot guessing. IT and InfoSec get it governed: inside your cloud, with the same entitlements your source systems enforce — Johnny in sales still sees only Johnny's accounts, right down to the row.

Powerful AND governed.

And it's a no-regret move, because whatever chat assistant you pick still needs a data platform minding that 90%.

Picture an account manager prepping a client's quarterly review. Today: log into the web app, pull four separate reports (hopefully the right four, filtered right), paste them into a Copilot chat — all before building anything client-facing.

With the governed layer underneath, it's one sentence: "Copilot — use Databricks to prep the quarterly review for the XYZ account." Copilot reaches the clean, joined-up data directly, over the official Microsoft–Databricks connector.

Switch to Claude next year? Same sentence, same governed data underneath. Those same MCPs serve Claude. Nothing to rip out, nothing to rebuild.

So there's almost no throwaway work here. You're not betting on a chatbot. You're building the layer underneath that every chatbot plugs into — and it's already sitting in your cloud, waiting to be switched on.

If you're trying to figure out where the chat assistant ends and the data platform begins, I'm happy to sketch it out with you.