Nobody Warns You About Authentication: Rolling Out AI Coding Tools Inside a Bank

Everyone demos AI coding tools in five minutes. Nobody warns you about authentication.

I'm rolling out AI-assisted development tooling to thousands of engineers inside a major financial institution. The tools themselves are the easy part. The CLIs work. The models are good enough. If this were a laptop on a home network, we'd be done by lunch.

It is not a laptop on a home network. It is a bank.

"Log in" is a different sentence inside an enterprise

In a consumer setup, the flow is trivial. You install the tool, you authenticate, you're generating code two minutes later. The friction is close to zero, and that frictionlessness is exactly what makes these tools feel magical in a demo.

Inside a large financial institution, "log in" unpacks into something much heavier. Identity is federated through a gateway. That gateway needs specific proxy configuration to let anything talk to the outside world. And making the whole chain work reaches down into individual developers' environment variables.

Read that again. To let one engineer use an AI assistant, you potentially touch their local environment, in a place where "change nothing without a ticket" is not a joke. It's the operating system of the entire organization.

The part the keynotes skip

Every conference talk about enterprise AI shows you the output. The generated function, the passing test, the productivity chart going up and to the right. What they never show you is the six weeks before any of that, spent making authentication survive contact with a security perimeter that was designed long before anyone imagined an AI writing code.

Here's the reframe that took me a while to reach:

• The model isn't the hard part.
• The distribution isn't the hard part.
• The plumbing that connects a modern AI tool to a decades-old enterprise identity perimeter, that's the hard part.

And it's worth being precise about something: this isn't a flaw. It's the tax you pay for operating somewhere that takes security seriously. A bank that let an unmanaged tool reach into developer environments without friction would be a bank with a problem. The friction is the feature, from their point of view.

The job isn't to resent the wall

When you're the person responsible for adoption, it's tempting to treat the security perimeter as the enemy. It slows you down. It turns a two-minute setup into a multi-week negotiation. It makes you explain, repeatedly, why the thing that works instantly everywhere else needs special handling here.

But resenting the wall is a losing game. The wall isn't going anywhere, and it shouldn't. The job is to find the door: the configuration that satisfies both the tool and the perimeter, the path that gets thousands of engineers productive without asking the institution to lower its guard.

That's the actual work of enterprise AI. Not the prompt engineering. Not the model selection. The unglamorous, deeply specific work of making powerful new tools coexist with the constraints that keep a serious organization serious.

More field notes as the rollout continues.