A great model isn't a governed system

Every few months a new AI product lands for a regulated industry, and the demo is genuinely impressive. Then someone asks who's allowed to see this client's record, and the room goes quiet.

The model was never the hard part.

In September, Anthropic launched Claude for Financial Advisors, which connects to the CRMs, custodians and planning tools advisors already use. I read its product page (opens in a new tab) closely, and the most interesting thing on it isn't a feature. It's how plainly the vendor describes what it leaves to you.

The vendor is telling you the foundation is yours

"Claude only reaches the systems you connect, with the access you already have in them," the page says. "Each platform stays the system of record for its own data."

That's a sensible design, and an honest one. It also means the AI inherits whatever is already true about your systems. If the access people have today is wrong, the AI's access is wrong too. If one client shows up as three different people across three systems, the AI sees three people.

The page goes further. Notes, emails and CRM updates are "drafted and held until you approve them." The Enterprise plan is "the only plan that includes audit logs and compliance controls for a regulated firm." And then: "The advice and the decision stay with you."

I agree with every word of that last line. Decisions belong to people. But read together, these lines describe a clear division of labor. The model drafts. Almost everything that makes the work trustworthy in a regulated firm stays with the firm.

Five foundations the model sits on

When I look at any AI tool for regulated work, I look past the model to five things underneath it.

1. Resolved identities. Is this the same client in the CRM, the custodian, the planning tool and the email archive? If people can't answer that with confidence, the model can't either.
2. Enforced permissions. The AI should see only what the person using it is allowed to see. That depends on entitlements being right today, not the way they were set up years ago.
3. Retention. What gets kept, for how long and where, including the drafts, prompts and outputs the AI produces along the way.
4. Auditable actions. Who did what, on whose approval, and when. The record should show that a person reviewed and approved a draft, not just that a draft went out.
5. Location. Where the data sits while the AI works on it, and who else holds a copy.

None of these are new. Regulated firms have worked on them for decades. What's new is that an AI tool exposes the gaps faster, and dresses them up in very fluent prose.

Same old foundations, smarter software

I've spent about 12 years in AI and enterprise data, much of it in-house at Credit Suisse, JPMorgan and BNY. Every enterprise I've worked in had more than one version of the truth. That's not a scandal. It's what happens when large organizations grow, merge, and build systems for different purposes at different times.

The mistake is assuming a new tool makes that go away. It doesn't. Most AI projects don't fail on the math. They fail because the organization hasn't done the human work: deciding who is who, who may see what, and who signs off.

Here's a simple hypothetical. An advisor asks an assistant to summarize a client's recent activity. The client sits under two IDs because of an old account migration, and one account's permissions were never updated after a team change. The summary comes back clean, confident and well written. It's also incomplete, and part of it was drawn from records that advisor shouldn't have been able to see.

No model upgrade fixes that. A better model just gets you to the wrong answer faster, and more convincingly.

What this means for buyers

None of this is an argument against the tool, or against AI for advisors. It's an argument for doing the evaluation in the right order.

Before comparing models, ask your own people a few questions. Can we resolve a client across our systems today, without someone doing it by hand? Do our entitlements reflect who actually does what now? Do we know what we keep, and for how long? If a regulator asked who approved a particular client email, could we show them in minutes?

If the answers are mostly yes, a good AI tool will make your people faster. If they're mostly no, the tool will make the gaps faster too. That's worth knowing before the contract is signed, not after.

The good news is that this work pays off well beyond AI. Clean identities, accurate permissions and real audit trails make every system you run more trustworthy, including the ones you'll buy next year.

Who did what

The tool drafts. People approve. The records prove who did.

Buy the best model you can find. Just don't expect it to know who your clients are.