Agents Became Infrastructure While You Were Running Pilots
The pilot era is over. Agents now sit in operating budgets, not innovation budgets. The question is no longer whether to adopt but whether your…
The pilot era is over, and the companies still scheduling steering committees to evaluate it missed the transition. This year JPMorgan set a technology budget near twenty billion dollars with two thousand people assigned to AI. Enterprise now accounts for more than forty percent of OpenAI's revenue. Agents are no longer a line item in an innovation budget waiting for proof. They are core infrastructure inside the operating budget. The question on the table is no longer whether to adopt. It is whether your architecture can absorb systems that are already running at the companies you compete with. Where a capability sits in the budget reveals how an organization actually thinks about it. Innovation budgets fund experiments. They are sized for learning, insulated from the core business, and judged by whether they produce a convincing result. Operating budgets fund the things the business cannot run without. They are sized for scale, wired into the core, and judged by uptime and outcome. For three years, AI agents lived in innovation budgets. The signal in this year's numbers is that, at the companies setting the pace, they crossed into operating budgets. Crossing into the operating budget is not a louder version of the pilot. It is a different commitment. When a function moves from innovation to operations, the conversation changes from whether it works to whether it is running. Steering committees that are still debating the first question are answering a question the market already closed. Reading the macro numbers as permission to start adopting is tempting. That is the wrong lesson, because adoption is no longer where the difficulty lives. Buying access to capable models is easy and cheap. Your competitors have all done it. The hard part, the part that actually separates outcomes, is absorption: getting an agent into production inside your operation and having it improve the business rather than sit in a corner producing demos. Absorption requires three things that pilots never forced you to build. A data foundation clean enough that an agent operating on it produces correct results. An integration surface that lets the agent read and write to the systems where work actually happens. And an operational model with confidence thresholds, escalation paths, and feedback loops, so the agent degrades gracefully and improves over time. A company can adopt in a week. Absorbing takes the architecture most pilots deliberately skipped. Here is the uncomfortable implication for the teams that spent two years on pilots. A pilot is optimized to prove a point in a controlled setting. Architecture is optimized to run in production at scale. They are different work, and time spent perfecting the first does not accumulate toward the second. The companies that ran pilot after pilot have a portfolio of proofs and very little production capacity. The companies that spent the same period building data foundations and integration surfaces can now put agents into operation quickly, because the hard part was already done. This is why absorption capacity, not model access, is the moat. Models are converging and commoditizing. The ability to put them into production inside a specific operation is not. The question to ask is not what should we pilot next. Ask instead whether your data, integrations, and operating model could absorb a production agent this quarter if you decided to run one. If the honest answer is no, that gap is the work, and it is the work we do.