As CFOs Press For AI Returns, CIOs Prove Efficiency And Let The Business Own The Dollars

Enterprise AI spending is climbing faster than the methods to measure what it returns, and finance teams keep asking technology leaders for a single number that a new deployment can't yet produce. It's become the broader challenge of proving AI's value, and the pressure lands on the CIO. A more durable answer starts before anything ships, with a clear division of labor: technology proves the efficiency it can actually measure, the business validates the financial impact, and governance decides where AI belongs in the first place.

Efe Miller is the Chief Information and AI Officer at Phipps, the nation's oldest not-for-profit organization dedicated to developing, acquiring, operating, and managing affordable housing, where he's building the company's enterprise AI strategy. Across 25 years in technology and cybersecurity, he created and led Astro, the multimillion-dollar automation program at PwC that connected major enterprise systems like Workday, ServiceNow, SAP, and Oracle and is now used daily by roughly 75,000 employees, and he's held senior technology roles at Infosys Consulting, Lightsource BP, and SEI Investments. His approach to AI economics starts with a boundary he draws around his own job.

"I can calculate the efficiency of a process and the time we save down to the minute. But the financial gains on that, I ask the business to validate. As a technologist, I couldn't do that myself," Miller said. The split is deliberate. Boards still want one ROI figure, and Miller’s position is that the team closest to the build should hand the financial call to the people who own the P&L, while it holds itself to metrics it can defend.

• The budget no one can forecast: "When we get asked what the AI budget is, there are gray areas we can't calculate, because we haven't implemented it yet, and we don't know what our users will consume in tokens or tool sets until we do. I can put the tool costs, the resourcing for the AI team, and the governance in front of the board. The token spend per employee is impossible to pin down until you actually start running it." The honest gap sits between fixed costs, which are knowable, and consumption, which isn't until real users meet the system. That uncertainty is pushing CIOs into a role that increasingly looks like advising the CFO on where technology spend pays off, rather than defending a figure invented for a slide.

• What the technologist can prove: "We're not here to fire people and say, 'here's how much we saved on their salary.' That's not the business we're in," Miller said. Stripping the dollar claim out of the technology team's remit is partly a credibility move and partly a practical one, since AI tends to add cost before it returns. What stays on the technologist's side is concrete: minutes saved per process, tooling, resourcing, consumption once it's visible. The financial translation belongs to whoever owns the P&L.

The economics only work if the process underneath is worth automating. Miller spends much of his time in workshops with department heads and business leaders, mapping what a process actually looks like before any model touches it. That upfront work changes the token math, because a leaner process asks less of the system. It also changes the investment case, by exposing steps that don't need AI at all.

• Clean the process first: "You have to look at the process itself first. There's no point automating a bad process." The data backs the instinct. McKinsey’s latest survey found that the AI high performers reporting real value were far more likely to fundamentally redesign their workflows than to insert AI into the ones they already ran. Miller takes it a step further: "I've guided teams to change a process with other technologies, not AI, where it became so clean and small that we didn't need AI at all," he said. The freedom to say no is part of the discipline, and it saves money twice.

Once a process earns automation, the question shifts from economics to control. Miller's governance model runs on hard checkpoints rather than good intentions. A QA team tests any strategically important system and gets business sign-off before it reaches production. After launch, authority to decide where a human intervenes sits with the department, not the technology group.

• Where humans stay in the loop: "Some processes don't need human interaction," Miller said, "but the ones that matter strategically have to pass a human before they execute. The AI takes its steps, then it comes to the governance group for a double-check, and that person clicks the button." Who holds that button, and when, is a business call, which is why so much of the design effort goes into deciding when an agent has earned autonomy. Miller is direct about the ceiling on that trust: strategic business processes should never get 100 percent of the decision.

• Monitoring agents like security: "Just as we monitor cybersecurity, there has to be an AI security team watching those logs and behaviors. Tools that give you insights into each agent, and tool sets that monitor what every agent does." Treating agent behavior as a security discipline gives the checkpoints something to act on, since live systems drift after launch in ways a sandbox never shows. When something breaks, it routes back to the technology team to diagnose from the logs and fix, then back through QA before it relaunches.

All of this scales only if the right people can build. Miller empowers employees to automate low-impact work while reserving anything with corporate-level consequences for the technology and governance groups, the discipline behind sustainable adoption once a program grows past its first pilots. The pipeline feeding that model is enormous. His workshops surface "hundreds, if not thousands" of processes per department, each one evaluated for whether technology, automation, or nothing at all is the right answer. Managing AI as an estate, not a set of experiments, is what keeps the spend tied to outcomes.

Miller saves his sharpest warning for the rollout itself. "You can implement AI all day long. But if you get the change communication wrong, the adoption won't be there, and it will be seen as a failed project."