Not since the term “data warehouse” took the industry by storm several years ago has a single technology generated so much enthusiasm from financial institution managers. “Artificial intelligence” headlines are everywhere, and it’s hard to overstate the promises being promulgated by an industry desperate to cash in on the Next Big Thing.

Indeed, if we believe the prognosticators, artificial intelligence (AI) will supplant humans in everything from writing loans to making dinner to flying planes. It’ll write the software we need in seconds, block 100% of all fraud, enable self-service everywhere, and file our taxes correctly and on time.

But that reality doesn’t exist yet. We remain on the steep side of the AI evolutionary curve. While AI can improve certain functions and interactions, most solutions fall well short of promises. Under the covers, most commercial AI software remains nothing more than glorified “if, then” statements and lists of keywords.

Don’t get distracted by promises. Before embarking on an AI project, consider the following:

Your credit union’s goals

Before diving into research on vendors and prices, make sure your credit union knows what it’s going to do with AI and what it will gain from it. Investing in technology for technology’s sake is never a great idea. Make sure you know the use case.

1. Define what success looks like before making a purchase. What answers do you want out of the system? What will you do with those answers? Can you monetize or otherwise act on those answers?
2. Demand a proof-of-concept test on actual data before buying. Is the tool capable of delivering desired outcomes reliably and repeatedly?
3. Assess your staff. Your team and tool operators will need to understand expected output and conduct sanity checks. It is not uncommon for AI tools to spit out incorrect or incomplete information that could be costly if acted upon.

Know the product

Now that you know what the credit union’s end goals are, determine if the AI will have the ability to get you there and what tangible data and responses from the technology you will receive.

1. Understand tool limitations and make sure the tool has the required capacity. Ask how much data it can consume when running its computations. What happens when it reaches the limit? Is there any feedback to the user to indicate it ran its analysis on partial data instead of the full data set?
2. Are you expecting repeatable outcomes from one day to the next? Often, AI engines will provide different answers each time they’re asked the same question, because it assumes the previous answer was found unacceptable. This can be undesirable behavior in some use cases.

Establish guardrails

Having guardrails on AI is critical in order to safely implement it. Here are some things to know when looking at a product:

1. Understand available guardrails for keeping data private and out of public large language models. Some products, such as Microsoft Copilot require additional investments in control panel software to classify and protect data. Will your budget support such additional tool layers, including experts to define, deploy, and maintain data privacy controls? Once the controls are in place, can they be audited for effectiveness?
2. Understand the guardrails for controlling AI software in the network. Some operate under the context of named users, and existing network controls can provide the required sandbox. However, others may run under service accounts with broader access. Ask the vendor how they will set up guardrails for the AI using the concept of least privileged access.

Choose a vendor

As with any technology in your credit union, choosing the right vendor can make or break the experience. Here’s the approach we recommend:

1. Shop multiple vendors before putting members’ money on the line, and make sure the company has a solid business model. Many AI firms will be gobbled up or disappear over the next few years. Count on that happening and do your due diligence.
2. Understand the long-term costs for running the software. Technology always requires reinvestment down the road to keep it going. Explore the vendor’s roadmap and understand where they plan to take the solution in the next several years, and what’s included today that may become “extra cost” down the road.
3. Have an exit strategy for switching to another vendor or getting the information needed an alternate way in the event the vendor goes out of business or eliminates the solution.

While these recommendations won’t guarantee a successful AI project, they can help improve the chance the tool will be a net benefit to the business instead of another write-off that didn’t quite live up to the hype.