Exploring the Three Levels of AI in Credit Unions

Most AI conversations at credit unions start with the tool. Which one should we buy? Is Copilot good enough? Should we block ChatGPT?

Those are fair questions, but they skip a more useful one: what can you safely give the tool? That decides what it can do for you.

I spoke about this at the North Idaho International Credit Union Day dinner, where the theme was how we keep connecting with and having an impact on our communities. AI might seem like an odd fit. I don’t think it is.

Alexandre Barbosa, CFO of Sicredi in Brazil, put it well at the World Credit Union Conference in Sydney in July. As Credit Union Daily reported:

“The future of credit unions won’t be determined by who uses AI first. It will be determined by who uses AI best to serve their communities.”

That should be a relief to anyone who feels behind. Nobody is too late.

Community work never has enough of two things: time and insight. AI doesn’t shake hands at the food bank or sit on the chamber board. People do. What AI can do is give those people more hours and a better read on where they’re needed.

For the work we do, I think about it in three levels, based on the information we can give it.

Level 1: Give people their time back

Meeting notes and follow-ups. Email and content drafts. Talking a problem through before the meeting so you walk in with a point of view. This is a useful place to start, using an approved tool and keeping sensitive information out.

The point isn’t the notes or the drafts. It’s where the hours go. An hour not spent on a recap email is an hour at a branch, a school or a community event.

The risk at this level is sounding like everyone else. Members can tell when a message was written by nobody. Use AI for the first draft and keep your own voice in the final one.

Level 2: Give it enough context to be useful

This is where it started to pay off for us.

The difference is that the AI knows your credit union: your plans, your products, your markets, your history. It has something to work with beyond a generic question.

Credit unions sit on information that’s disconnected by default. Survey results in one place, performance benchmarks in another, product plans in a third. Too many decisions get made from one number in its own silo while the details that could change the outcome sit nearby, unconnected.

We built one knowledge base and reused it all year. Member survey results, benchmarks, every product and what sets it apart, our strategic plans, our mission and values. When I asked it for a product marketing plan, the first draft came back informed because it already knew us. A few rounds of “change this, don’t touch that” and we had something we could act on in 20 to 30 minutes. I wrote up the full setup in Beyond Meeting Notes.

Here’s the part I most want other credit unions to hear: our data wasn’t perfect. We left identifiable member data out entirely. No names, no accounts, only aggregated, anonymous survey responses. We cleared it with data governance before loading anything. It still worked.

Clean data is nice to have at this stage. It shouldn’t be the reason you wait. Start with what you’re approved to use and improve it as you learn what matters.

Level 3: Build inside an environment you control

Private AI means running AI in an environment your credit union controls. For the work I’m describing, the requirement is that member data stays inside that approved environment. Calling a product “private” doesn’t establish that on its own.

That control is what makes deeper work possible: tools and products shaped around the people you serve, with more say over what gets built and when.

We’re not at this level yet. Getting there may involve a dedicated vendor or platform, and it requires a real conversation about data governance and AI policy before anyone touches member data. Where the system runs, who can access it and where information goes all matter.

The distinction between the levels is what you can safely bring into the work:

• Level 1: Everyday work, drafts and general information. Keep member information out of general-purpose AI tools.
• Level 2: Your own plans and documents, plus aggregated, anonymous data, approved for use in that environment.
• Level 3: Member data, but only within an approved environment with the controls to protect it.

A human reviews the work at every level. Don’t take the first draft at face value. Push back and revise.

And turning off a training toggle is not a governance answer. It’s a good first step. Treat anything sensitive as off-limits until you have a better answer than “I turned off a setting.”

Move one level up

If you haven’t started, hand one weekly task to AI. If you’re at level 1, give it approved context for one real decision. If you’re at level 2, ask what you would build if no vendor stood in the way, then work out what would make it safe to build.

The technology is the easy part. What we do with the time is the part that matters.