CIOs Confront The Cost Of AI Overconfidence With Stronger Review And Intake Rules

Most enterprise AI failures start with a person. Someone trusts a confident answer without checking it, acts on it, and the error shows up later, after the work has already gone out. With more of the workforce using these tools every month, the guardrails that catch a wrong answer before it does damage have become part of the CIO's core job.

Dell Joshi is the Chief Information Officer at Envu, the environmental science company spun out of Bayer in 2022, where he leads the buildout of the organization's IT and data infrastructure. Across two decades in manufacturing and industrials, he handled M&A integrations and divestitures at DuPont and at dss+, the operational consulting firm formerly known as DuPont Sustainable Solutions. He holds master's degrees in chemistry and business and carries a Six Sigma Champion certification. He approaches AI as an operator first.

"You have to have solid governance and clear guidelines, so you don't run into a situation where somebody got lazy," Joshi said. By lazy, he means the pull to accept a fluent answer at face value and move on. AI sharpens that pull because its output reads as settled, whether or not anyone has checked it.

• Google doctor: Cheap access to answers makes expertise feel optional, and people stop consulting the specialist who once stood between a question and a decision. "When Google first came out, everyone became a Google doctor, and AI carries that same danger; CIOs have to be careful of it," Joshi said. Inside a company, the stakes climb because one person's unverified guess can flow into a contract or a filing that others treat as fact.

• Confident and wrong: Joshi pointed to a law firm that filed a brief riddled with errors, AI output that reached a court because no one checked it. The same risk runs through finance, where a bad number can travel straight into a decision with money behind it. "You don't want a situation where an analyst doing research on companies gives their senior leader a paper with incorrect information, and that firm makes investment decisions based on it," Joshi said.

• What goes in: The bigger risk sits upstream of the answer, in what employees put into the tool. Joshi keeps company data out of any system he cannot control, even while he uses public AI for everyday questions. Many skip that discipline, and sensitive material can slip through an unmonitored browser tab. It reaches the highest levels. Last year the acting head of the federal cybersecurity agency uploaded sensitive documents to a public version of ChatGPT, tripping security alerts and prompting an internal review. The solution is a clear line on what data belongs where, drawn before people set their own.

Checking the output and controlling the input both matter once a project is running. But the efforts that stall usually went wrong earlier, at the planning stage, where a fuzzy goal or an unproven return should have stopped them. So Joshi puts his weight at the front, on whether a project should exist at all, which is where he turns most requests away.

• Return to sender: Every AI request goes through an intake conversation before any work begins. "If they come and say, I want to use AI, then I send them back," Joshi said. From there, he pushes for the specifics that make a request worth funding. He wants the task the tool would take on, the business problem beneath it, the return the spend has to clear, and a grounded estimate of how many people will use what gets built. He watched the same confusion with big data years earlier, when teams fixated on the technology while the business question went unasked.

• Already in the stack: Whether a tool earns its keep comes down to how ready the ground beneath it is, and Joshi checks that before he weighs a purchase. The major enterprise platforms carry AI inside them, so he starts with the capability the company owns, switching it on and adding outside tools only where the work calls for them. "The question becomes how well you're utilizing it, and how ready your processes and your data are," Joshi said.

• Count the full cost: The build estimate is only the start of what a tool costs. Joshi weighs savings against everything it takes to stand up and keep running, and he wants that math done before anyone commits. "If you use AI for a chatbot and reduce vendor tickets to save $100,000, but you spent $300,000 to build it, you have to look at the ongoing costs. Are you continuing to spend that money, and when will you get your return on investment?" Joshi said.

Where to point AI depends on the industry and the company's data, and picking well takes judgment. That's the part of the job Joshi keeps for people, however much of the work AI takes on. He sees it in the requests he sends back and the answers he checks before anyone acts on them. "There can be a million use cases, but you've got to pick the right ones. It's not one size fits all."