Agentic AI, software that plans and takes action on its own instead of waiting for a person to approve each step, is the technology every company is being pitched right now. A new test led by researchers connected to MIT suggests a lot of that money is about to be wasted, and not for the reason most people assume.
The researchers built an AI system to handle a real and boring business problem: catching errors in supplier invoices for a company that processes over 75 billion dollars worth of them every year. They ran four different versions of the system against the same 44 invoices. One version used the strongest, most expensive AI models on the market. Another used cheap ones. A third mixed price tiers in the middle.
The expensive version came in last. It correctly resolved 91 percent of the invoices without needing a person to step in. A cheaper setup handled 95 percent. The mid-priced version handled all 44, using models that cost a fraction of the premium ones.
Every mistake happened on one task: spotting duplicate invoices, which requires checking five different rules at once to decide if two invoices actually match. The expensive model overthought it. It hedged on close calls that a simpler, more rigid model just decided on and moved past. More brainpower turned into more doubt, not more accuracy.
There was a second surprise buried in the cost numbers. When the researchers kept the expensive reasoning model but swapped in a cheaper model to read the invoice documents, the whole system got more expensive to run, not less. The reasoning model had to spend extra effort cleaning up messier text before it could even start its actual job. Saving money on one piece of the system pushed the cost onto another piece.
None of this is an isolated finding. Gartner has predicted that more than 40 percent of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear payoff, or weak controls. A separate study out of MIT looked at 300 real company AI rollouts and found that only around one in 20 produced a fast, measurable financial return. In both cases, the researchers point to the same root cause: companies are matching the problem to a model's reputation instead of testing it against the actual job.
There is a deadline attached to this for a lot of companies too. The European Union is phasing in mandatory electronic invoicing over the next couple of years, which means more businesses will need automated systems that can actually read and cross-check invoices correctly, not just quickly.
The takeaway is not that agentic AI is a bad bet. It is that picking the most capable model on a leaderboard and calling it a strategy is a guess, not a decision. The businesses that get real value from this technology will be the ones that test the specific task on the specific data first, watch the whole chain of tools work together, and only then decide what to buy.