Bad Data Kills Good AI

An AI pilot project runs for a few months, and the results we expect don’t show up. The conclusion is that AI just didn’t work in our business. But what went wrong?

It’s easy to draw the conclusion that AI is not ready for a company like ours. That conclusion is usually inaccurate. The technology was capable. It was just pointed at information it could not use.

Harvard Business Review research on companies deploying AI agents in procurement identifies that treating data as an afterthought as one of the three problems that most reliably stalls them.¹ These systems are built on customer, supplier and product records that disagree from one department to the next, and on categories that were never finished. The results come back inconsistent, and the company decides the technology failed when the data never gave it a chance.

The gap is close to universal. In a 2026 survey, 94 percent of executives agreed that having well connected data, processes and applications is important for successful AI adoption. Only 27 percent said their own were connected adequately.² A different, but more difficult, aspect of the problem unstructured data. Structured data is the tidy material in rows and columns, such as your accounting system or customer list. Unstructured data is everything else: PDFs, emails, images, contracts and call recordings. It is an estimated 80 to 90 percent of what a company holds, and companies rate themselves twice as ready to use their structured data (67 percent) as their unstructured data (33 percent).²

Properly preparing your data upfront will delay the launch of your first AI project. That is a real trade-off, but there can be a significant benefit to doing so because it removes a significant number of future bottlenecks and headaches related to any technology development work, including the adoption of AI.

Before you approve the next project, name the information that the system will depend on, where it is kept, and the person who can confirm it is accurate and complete. If you are not yet able to answer these questions, you likely need to undertake a data project as part of your AI adoption process.

Sources

¹ Heiner Himmelreich, Ilan Oshri, Paolo Scala and Anas Zaidani, “Why Agentic AI Could Transform Procurement,” Harvard Business Review, August 13, 2026.

² “Bridging the Gap to the Content-Powered Agentic Enterprise and Elevating Workforce Productivity,” Harvard Business Review Analytic Services webinar summary, June 25, 2026.