A new survey from Kyriba, a finance AI platform, interviewed 1,000 CFOs and found 96% are prioritizing integrating AI—even though many still have major concerns about doing so.

AI often functions like a “black box,” creating uncertainty about how it arrives at its final outputs. Additionally, there are concerns around data privacy and security as well as whether AI compliance.

But even with these risks, it seems like the promise of improved efficiency is winning out;according to the study, 86% of CFOs are already using AI in some or most aspects of their job.

So how should companies and CFOs prepare for AI adoption to mitigate risk?

Black box. According to Bob Stark, global head of enablement at Kyriba, there are ways to combat these concerns that could ease CFOs’ comfort with integrating AI.

“Every CFO that we talk to, they say the same thing,” he said. “‘It needs to be our data. [We] need to understand how it works, and we need to ensure that the outputs are our own and only our own, and that they can work within [our] own organization’s policy.’”

While even some software engineers building AI may not entirely understand how it works, the AI products could at least be more open about their work so CFOs could independently validate the outputs, according to Stark. On the security side, there should be guardrails in paid AI products that can be turned on to prevent data from being saved, used to train models, or exposed to others, he added.

According to Glenn Hopper, head of AI research and design at Eventus Advisory Group, the same security rules for services like Google, Snowflake, or AWS should apply to the enterprise versions of AI products.

“The security concerns have gotten overblown a little bit,” he said. “The danger with data being uploaded into the models is very easy to overcome.”

As for compliance, an industry that isn’t known for moving quickly might have to if it wants to keep up with the proliferation of AI and get in front of these risks, according to Hopper.

Know your goals. Before deploying AI, Stark recommends that CFOs really understand what their aims are. .For example, AI could help with exposure management, hedging, and accounting processes, according to Stark.

After identifying the goals for AI use, its accuracy should be tested. Stark suggests starting by comparing previous methods of forecasting with the new AI-powered results.

“That’s the kind of journey that can help build trust,” he said.

Create a policy and train employees. Once organizations are clear on the scope of the AI’s work, it can roll it out to employees with clear policies and thorough training.

Hopper advises CFOs to work with the senior management to create AI usage policies that everyone can agree on, which should state which AI systems employees can use, what they’re allowed to upload, how to use it, and when to integrate a human into the process. Stark also urges companies to explain what compliance under the policy actually would look like.

AI is more flexible than traditional tools, according to Hopper. Managers tell employees how to use traditional software packages, whereas with AI, employees will shape how the tools are used in the workplace.

“They’re going to figure out on their own how to automate parts of their job,” he said. “And you want it to be out in the open. Because if it’s out in the open and someone’s using it improperly, you can figure it out. You don’t want people to be scared of it, but you also don’t want them to be reckless.”

Hopper calls for basic training on prompt engineering, outlining specific tasks best suited for AI, explaining how to verify data, and checking for hallucinations with a structured employee rollout.

“In finance, we don’t anticipate roles being replaced, but we do recognize that people with AI may replace people that don’t have AI,” Stark said.

This report was originally published by CFO Brew.