John Davie, Founder & CEO of Buyers Edge Platform and CEO of CollectivIQ, leverages data and AI to drive smarter enterprise decision-making.

getty

I've spent more than 25 years building and scaling a billion-dollar food procurement business. In doing so, I’ve navigated acquisitions, managed billions in procurement volume and led the company through some of the most disruptive shifts the industry has ever seen. But through all of that, what really keeps me up at night is the cost of making avoidable bad decisions.​​

I’ve been a vocal advocate for scaling AI adoption across Buyers Edge Platform’s 1,200-plus employees. I understand the massive potential it has to solve real operational problems across our organization, like empowering our team to analyze faster, identify pricing discrepancies and make smarter procurement decisions across more than 300,000 restaurant locations.

However, as we began to integrate AI models into our workstreams, the productivity gains were real, but so were the errors. And as I spoke with my peers, I realized I wasn’t alone. These mistakes—caused by AI hallucinations—were costing all of us significant money.​​

In fact, AI hallucinations—instances where models generate confident but factually wrong outputs—are estimated to cost businesses $67 billion globally. The average single AI error carries a price tag of $4.4 million, and nearly half of all organizations have already taken that hit.

Employees are now losing the equivalent of 51 workdays a year to technology friction, up 42% from 2025, with more than four hours a week spent just verifying AI outputs. But perhaps the most telling signal is that only 9% of workers, the people closest to the decisions, data and consequences, trust AI for complex, business-critical decisions. And yet, most organizations are pushing forward anyway, because the alternative feels even more costly.​

In reality, we are deploying technology at a breakneck speed. In the race to implement AI, however, enterprises have not comprehensively reckoned with what happens when that technology is wrong. The root of the problem goes deeper than the models themselves and extends into the architecture. In an era where leaders are turning to AI for answers to their most important business questions, relying on a single model locks employees into one set of responses, biases and blind spots. When the model produces a wrong answer, and it will, there is no mechanism to catch it. There is only an output that moves downstream and gets acted on.​

We wouldn't accept frequent errors in any other high-stakes context. We build redundancy into supply chains, require second opinions before major financial decisions and pressure-test assumptions before committing to strategy. But with AI, enterprises have defaulted to trusting a single source of truth, one they didn't build, can't fully audit and don't entirely understand.

Microsoft's own terms of service recently described Copilot as "for entertainment purposes only," warning users not to rely on it for important decisions. ​

The consequences of overreliance on these tools are already showing up in boardrooms and balance sheets. Enterprises are seeing flawed market analyses, compliance documents built on incorrect premises, strategic recommendations that sounded airtight until they weren't and countless hours spent by employees quietly double-checking AI outputs.​

For AI to be effective for businesses, we need a fundamentally different philosophy built around transparency, rather than false confidence. Companies should apply the same rigor to evaluating AI outputs as they would for any other decision-making process.​

To make that idea a reality, business leaders can build practical AI evaluation steps into everyday workflows. Instead of taking AI outputs at face value or treating any single model as an unquestioned source of truth, employees should be encouraged to double-check responses against internal data, third-party sources or subject-matter experts. Teams can also compare outputs across multiple LLMs to ensure they’re sourcing the best responses. ​

When it comes to utilizing AI, slow and steady wins the race. The companies that will succeed will be those that build the right checks and balances into LLM usage across their organization. Balancing velocity with verification will allow businesses to realize the true transformative potential of this technology.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?