I’ve watched this pattern repeat across every transformation programs I’ve led: the model is never the first thing to fail. The cracks appear months earlier, inside the workflows, data pipelines, and governance structures that practitioners see everyday and executives rarely do. Once you’ve lived inside real operational machinery, you recognize the signs long before AI ever touches the process.
Most AI failures don’t happen when the model runs. They happen months earlier, inside workflows, data pipelines, and governance structures executives never see. Practitioners see the failure long before AI touches the process. And now, with multi-agent systems, the next failure mode is already forming.
Executives often talk about artificial intelligence as if it arrives in a vacuum, a clean, neutral technology dropped into the business to accelerate outcomes. Practitioners know better. They know AI never enters a vacuum. It enters an organization with a history, a workflow, a culture, a set of habits, and a thousand small operational decisions that have accumulated over years. And those decisions determine whether AI succeeds or fails long before the first model is deployed.
Practitioners site inside the machinery or business. They see how work actually moves. Now how it is described in strategy decks. Not how leaders assume it moves. How it really moves. They see the undocumented exceptions, the tribal knowledge, the inconsistent handoffs, the data that changes meaning as it travels across teams. They see the places where ownership is unclear, where governance is missing, where decisions are made informally because the formal process broke years ago. They see the drift, the slow, quiet erosion of operational clarity that accumulates until the organization is held together by habit rather than design. AI does not fix this drift. AI exposes it.
Inside cross-functional meetings, practitioners watch leaders debate priorities that do not align with the workflows required to deliver them. They watch engineering teams build around assumptions that operations teams know are no longer true. They watch finance teams model scenarios using data that practitioners know is incomplete, inconsistent, or interpreted differently across business units. They watch commercial teams make commitments the underlying systems cannot support. None of these is visible from the executive balcony. All of them is visible from the practitioner floor.
This is why practitioners react differently when AI enters the conversation. They do not ask, “What can the model do?” They ask, “What environment are we placing it into?” Because they know the environment is the real constraint.
When practitioners map workflows, they do not start wit the official process. They start with the lived process: the one people actually follow. And what they find is rarely deterministic. It is conditional. It is exception-driven. It is dependent driven on individuals who carry institutional memory that has never been documented. It is shaped by decisions made years ago that no one remembers but everyone still follows. AI cannot operate inside of this ambiguity. It collapses into it.
The same is true for data. Executives assume data reflects reality. Practitioners know it reflects the workflow that produced it, and if the workflow is inconsistent, the data will be too. They trace data lineage and find that definitions drift across teams, metrics mutate as they move through systems, and meaning is lost in translation long before AI ever sees the input. Deloitte found that only 32% of organizations trust their operational data. Practitioners could have told you that without a study.
Governance is no different. AI requires clear decision pathways, escalation routines, and ownership structures. Practitioners know where those structures are strong and where they have eroded. They know which teams follow the operating rhythm and which teams bypass it. They know where accountability is crisp and where it is blurred. AI cannot compensate for blurred accountability. It amplifies it.
I saw this play out years ago when I warned upper management and the board that the direction they were taking, and the guidance they were receiving from their engineers, was fundamentally misaligned with how the work actually moved. I showed them the results were never consistent, that the system behaved differently depending on who touched it, and that the workflow itself was the source of the instability. But the executives were focused on becoming “leaders in AI,” convinced the company was already behind the trend. Instead of stepping back to understand the operational reality, they pushed forward and deployed whatever was ready. The outcome was predictable long before the model ever ran.
This is why practitioners are the first to predict AI failure. They see the modes before the model is even trained. They know when success criteria are ambiguous. They know when workflows are unstable. They know when data is untrustworthy. They know when governance is missing. They know when expectations are unrealistic. They know when the organization is asking AI to solve a problem that is not an AI problem.
And when the model finally fails, practitioners are the ones who diagnose the truth: the model did not fail. The environment failed. The workflow failed. The data failed. The governance failed. The success criteria failed. The organization failed long before the model ever touched the process.
And now, a new failure mode is emerging, none that executives are being sold aggressively, and one practitioners already distrust; multi-agent AI architectures.
Executives are being promised a future where AI agents behave like autonomous digital employees, independent, collaborative, handing work off to one another without human involvement. It’s a compelling story. It’s also the wrong one.
Practitioners see something very different. They see fragile sub-processes chained together with assumptions that break the moment they touch real operational complexity. Each agent interprets context differently. Each agent transforms data in ways the next agent doesn’t expect. Each agent introduces drift. And in many organizations, each agent is written by a different person, a different team, or even a different vendor, each with their own assumptions, coding style, data expectations, and interpretation of the workflow. The result is not a coordinated system. It is patchwork of mismatched logic stitched together by hope.
I saw this dynamic firsthand when I tried to warn leadership that the multi‑agent architecture they were pursuing was fundamentally unsafe. The agents were communicating through MCP, which meant STDIO commands were being executed without proper isolation or sanitization. There were no governance boundaries, no guardrails, no deterministic rules. Agent‑to‑agent communication was effectively a free‑for‑all — each agent could create, modify, or trigger whatever it wanted as long as another agent accepted the message. I demonstrated how inconsistent the outputs were, how the agents drifted apart in behavior, and how the architecture itself created security exposure. But the executives were focused on “catching up” to the AI trend, convinced that speed mattered more than stability. Instead of stepping back to examine the operational reality, they pushed forward and deployed the system as it was. The failure that followed wasn’t surprising. It was embedded in the architecture long before the agents ever ran.
When you string these agents together, the drift doesn’t just accumulate, it compounds. A small misinterpretation in one agent becomes a larger misalignment in the next, and by the time the output reaches the end of the chain, the original intent has been distorted beyond recognition. Executives see a seamless pipeline. Practitioners see a Rube Goldberg machine.
The moment something goes wrong; a missing field, an unexpected exception, a misinterpreted instruction, the agents don’t fail gracefully. They fall silently. They pass corrupted context forward. They amplify the error. They create a cascade.
This is why practitioners don’t trust multi-agent architectures. Not because the models are weak, but because the architecture is wrong.
The agent shouldn’t be a worker in the chain. The agent should be the chain.
It should be the platform, the orchestrator, the interpreter, the validator, the governor, the memory, the rules engine, the stabilizer. A multi-agent pipeline is brittle. A single agentic platform is stable. A change amplifies drift. A platform absorbs it. A chain depends on perfect handoffs. A platform eliminates them.
This is the practitioner lens applied to agentic systems: AI doesn’t need more agents. AI needs a deterministic agent platform.
Executives often interpret AI failure as a technology problem. Practitioners interpret it as an operational truth. AI is not revealing something new. It is revealing what was already there.
This is the practitioner lens; the quite, unglamorous, deeply operational perspective that determines whether AI becomes a strategic asset or a strategic embarrassment. It is the lens that sees the system clearly enough to prepare it. It is the lens that understands AI does not replace operations; it depends on them. It is the lens that knows AI is not the worker.
AI is the system builder.
And until the organization adopts this lens, AI will continue to expose the failures they never knew they had.