A new study paints a clear picture of where companies actually are with AI agents, the software that can complete tasks on its own instead of just answering questions. Businesses are buying more of them, fast, but they still cannot get them to work together.
The study comes from IDC, which surveyed more than 400 enterprise decision makers on behalf of an AI company called Leah. Two out of three organizations already run AI agents in real operations, not just tests. They expect to have six times as many agents running by the start of 2027, and they plan to double their budget for agentic AI over the next year.
Here is the catch. Only 29% of those agents currently interact with each other. Just 7% of companies have what the report calls advanced coordination, where multiple agents work as a team on a task. More than half are stuck with the simplest setup: one agent finishes its piece and hands it off to the next, with no shared understanding of the bigger picture.
This is not a small technical detail. A contract, an insurance claim, or a supply order usually moves through several departments: legal, finance, operations. If the agent handling procurement cannot share what it knows with the agent handling finance, you get the same old departmental bottlenecks, just with a computer doing the paperwork instead of a person. The report found companies losing four to seven extra days on contract handoffs specifically because of this kind of disconnect.
Oversight is the other missing piece. Separate research from New Relic, which surveyed thousands of IT leaders, found that one in four AI agents run in production with nobody watching what they are doing. That means a quarter of the software making decisions inside these companies could be making mistakes, wasting money, or exposing data, and no one would notice until something breaks.
This fits a wider pattern showing up across nearly every recent study on enterprise AI. A widely cited MIT report found that 95% of company AI projects show no measurable financial return, largely because tools get bolted onto existing workflows instead of being built into how the business actually runs. Deloitte found a similar gap: most leaders expect AI agents to reshape half their business processes within the next few years, yet only one in five feel ready to redesign anything for it.
The pattern across all of this research is consistent. Companies are treating AI agents as a shopping list, one tool per department, rather than as a system that needs a shared set of rules and a single view of what every agent is doing. The businesses getting real value are the ones building that shared structure first: standard ways to build and check agents, oversight that covers the whole company, and measuring results by what actually improved, not by how many agents got deployed.
The money is already moving fast. The discipline to manage it is not keeping pace, and that gap is where the risk sits.