Glenn Hopper isn’t trying to pick a fight with IT. But the former CFO, author and advisor to hundreds of companies on AI in finance thinks most leadership teams have the ownership question wrong.
Handing employees a chatbot license isn’t an AI strategy, Hopper told the AI Value Creation Forum at Chief Executive Group’s Finance and Accounting Technology Expo. The next wave is agents that run long, multistep tasks on their own, and that puts AI squarely in the CFO’s lane. Finance decides what gets funded, owns the controls and signs the filings. “So if you have accountability for it, it’s an argument for having ownership for it.”
His advice for finance leaders:
- Your controls are your agent framework. Finance long assumed gen AI was a poor fit for work where the trial balance has to balance. The reverse turned out to be true. “As generative AI has gotten better, it turns out these very controls that we use for our humans are the framework that we can use to set up our agents.”
- Document before you automate. Hopper makes teams write standard operating procedures for everything they do, then look at the workflow and see what can be done to improve it. These improved SOPs can be then used as a foundation to train agents. “You can’t automate chaos, and if you take AI and put it into a broken process, it’s just throwing gasoline on the fire.”
- Set a baseline to determine ROI. Hopper treats AI as a portfolio of discrete projects, each with its own return. “If every project you do is its own project with its own ROI, then it’s much easier to measure what ROI is because you’re not just saying, ‘Well, we did AI and by doing AI, then we ended up getting these savings.’”
- If you can’t reproduce it, you can’t use it. Keep the math in code and give agents a bounded scope inside the workflow. “You can’t tell an auditor when they ask you where the number came from, ‘Oh, I gave it to the AI and it went and did some stuff and it came back with the number.’ You have to be able to reproduce it later.”
- Put named humans at the gates. Say September comes in soft and you’re reforecasting. An agent might spot October rebounds in the two prior years and call it seasonality, with no idea you just lost a big customer. Hopper puts an analyst and account owner at a first gate and a named finance leader at a second, before any money moves. “AI is very good at following the map that was laid by what came before. And just because orders rebounded in October the two years prior doesn’t necessarily mean that that’s going to happen.”
- Give every pilot an exit. Set a baseline, an owner and stop rules, then run the pilot for 90 days, or three closes. At the end, scale it, kill it or, only if it’s merited, refine and retest. And make sure that a human is still responsible for the results. “There always has to be a named human at the end that is, ‘I’m the one who is signing off on this.’”
Where to begin? “The way to start with this is pick one workflow to do this quarter and know its value,” he says. “Know what it’s taking [to do this work] today, know its limit, give it an owner and you can move forward with that.”





