Every company buying an AI agent this year is really buying a mirror. The agent will act exactly as well as the data and rules you hand it, and a new survey from Collibra and The Harris Poll shows most companies are handing it a mess.
The numbers are blunt. Nine in ten decision-makers said their teams still regularly re-check whether the information an AI agent is using is accurate and current. More than half said staff spend hours every week reviewing and fixing what the agent produced. Seventy two percent said that when their AI projects fall short, the reason almost always traces back to data that was never properly organized in the first place.
This is not one vendor's marketing claim. It lines up with separate research from Gartner, which expects more than 40 percent of agentic AI projects to be canceled by 2027 due to cost overruns and unclear payoff. It also lines up with a widely cited MIT study which found that 95 percent of generative AI pilots produced no measurable financial return. Different researchers, different methods, same conclusion: the bottleneck is not the model, it is what surrounds it.
Here is the part that matters for anyone running a business. For years, company data was built to be read by people. If a spreadsheet looked wrong, an employee would ask a colleague, check another system, or simply use judgment. That informal fixing is exactly what an agent cannot do. When an agent hits a gap or an unclear definition, it does not stop and ask. It fills the gap with a guess and acts on it, with full confidence, often invisibly. A separate Google Cloud and MIT report found that the average company only gives its AI tools access to 45 percent of its own data, and only half of companies actually trust the answers their agents give.
So the real project behind "adding AI agents" is not a software purchase. It is agreeing, across departments, on what words like "customer," "active order," or "approved vendor" actually mean, and writing down who is allowed to change what. That is unglamorous, slow, back-office work, and it is now the single biggest factor separating companies that get value from AI agents from companies that just get a new source of errors to clean up.
There is also a regulatory angle worth watching. California just signed laws creating a formal system for independent audits of AI systems. The EU's AI Act already requires companies to prove people understand how their AI tools work. Senator Bernie Sanders has proposed pausing advanced AI development entirely. Companies that build clean records now, showing where their data came from and who approved what, are also building the paperwork regulators will eventually demand.
The uncomfortable truth is that most businesses skipped this step because model demos are exciting and data governance is not. Expect that gap to keep punishing early adopters through next year, while the companies that quietly fixed their data first pull ahead.