The word I keep hearing in every boardroom now is agents. Not pilots, not copilots, agents. As if a wave of intelligent software is about to land on top of an existing machine and quietly fix it overnight. But the leaders searching for AI agents for business rarely agree on what an agent actually is, let alone what it needs to succeed. So here is the truth I keep telling boards: you do not build a house through the roof. You start with the foundations, clean data, a real operating model, processes wired for how work actually happens. Agents without foundations are nothing. And most companies are nowhere near ready.
Everyone is searching. Nobody quite knows what an agent is.
Every week I meet leaders who have heard the phrase a hundred times and still cannot tell you what an agent actually does, where its boundaries are, or what it needs to succeed. That is the honest state of the market. Interest has raced ahead of understanding, which is exactly the moment an opinion is worth more than a product pitch.
An agent is not a chatbot with better manners. It is software that takes a goal, plans a sequence of steps, uses the tools you give it, and keeps working until the job is done, monitoring its own output along the way. That is powerful, and that is precisely why it is dangerous to bolt it onto a house of cards.
Why most will fail: the roof before the foundations
The failure I keep watching is the same one I described in why most AI transformations fail. Companies buy the shiniest agent platform, plug it into a fragmented mess of systems, and expect transformation. Then they are surprised when the agent answers questions with the wrong data, quietly executes on permissions it should never have had, or produces output nobody trusts. The technology is not the hard part. The base underneath it is.
Compare how this plays out elsewhere. In coding, agents work because the context they need lives inside the codebase, it can be tested quickly, and the technical users know what to ask for. Most business workflows enjoy more of none of that. The knowledge is scattered across stale documents, legacy systems, and the heads of people who never wrote it down. The data is fragmented. The permissions are a mess. Nothing is ready for an agent to walk in and act.
The foundation that makes agents work
So what is the foundation? Three things, and they are not glamorous. First, clean, connected data, in a form an agent can actually use, owned by the right people. Second, an operating model where processes cut across the silos that have always separated them, because agents are most useful when they follow a process from end to end. Third, governance: knowing exactly what an agent can see, what it can do, and how you audit the work it produces. This is the same discipline I cover in the AI operating model beyond pilots.
Do the unglamorous work first. Connect the data. Document the process. Set the boundaries. Then put the agents on top of something stable enough to hold them. The companies that win do not start with the pilot count, they start with the plumbing nobody wants to pay for, and that is why they are the ones still standing two years from now. The answer is never the agent. It is the ground you build it on.