Your product team just announced they are building an agentic system. You feel that familiar unease — the one you felt during the Agile roll-out that never really stuck, the data lake that became a data swamp, the chatbot that answered every question with an apology. This time feels different because the technology is more impressive. It is not.
I have watched three waves of enterprise technology promise transformation, only to see the same failures dressed in new syntax. Every time, the story was the same: the models got better, the results did not. Agentic AI will repeat that pattern unless we stop pretending that technology can outrun dysfunction.
The failure rate for enterprise AI pilots sits somewhere between 40 and 88 percent, depending on who counts and how they define failure. Every study I have read lands on the same conclusion: the reasons are rarely technical. Misaligned incentives, unclear objectives, leaders who treat AI as a plug-in fix. None of those are things AI is good at. They are product leadership problems. And agentic AI, with its autonomy and complexity, will not fix them. It will magnify them.
Agentic AI is not a fix for broken product leadership. It is a microscope. It will magnify every decision making vacuum, every misaligned incentive, every team that celebrates demos instead of outcomes. The question is not whether your agent can reason. It is whether your organisation can absorb what the agent reveals.
The déjà vu of enterprise AI
Remember the Agile transformation that took three years, two consulting firms, and a cultural collapse? The Agile advocates said it would make teams faster. Instead, it made chaos repeatable. Stand-ups became status meetings, retrospectives became blame sessions, and every team did their own thing because no one had the authority to say no.
AI transformations are following the same script. Leaders are enthusiastic but unprepared for the structural and strategic shifts required. They buy tools before identifying needs, celebrate pilots that cannot scale, and measure activity instead of outcomes. None of this is new. None of it is a technology problem. It is a leadership failure that agentic AI will accelerate, not resolve.
An autonomous agent operating inside a dysfunctional team does not become wise. It becomes a faster, louder generator of the same broken behaviour. It will ship the wrong features faster, escalate the wrong conflicts more frequently, and optimise for the wrong metrics more efficiently. If your decision making process is unclear, the agent will not clarify it. It will execute the ambiguity and call it progress.
Where the failure lives
The research is clear: barriers to enterprise AI success are never about intelligence in the models. They are about incentives, governance, and the discipline to define clear outcomes. I have been in enough product reviews to know that most teams cannot articulate what success looks like for a simple feature. Now they are supposed to define goals for an autonomous system that learns and adapts.
The pattern is predictable. Someone reads an article about agentic AI, buys a platform, assigns a junior product manager to “figure it out,” and waits for magic. Six months later, the agent has been decommissioned or ignored, and the team blames the model. The real blame belongs to the leadership that thought technology would replace thinking.
Product leaders who lack strategic depth will be exposed by agentic AI. The tool-users will survive — they will prompt their way through the day. But the strategists who can design goal-oriented systems that reason, take action, and adapt over time are rare. Agentic AI will separate them cleanly. That separation hurts. It should hurt.
The readiness trap
Here is the counter-cycle argument: instead of rushing to build agents, product leaders should first diagnose their organisational readiness. Not a readiness for technology — a readiness for autonomy. Can your teams operate without micromanagement? Do they have clear, measurable goals that align across departments? Is your data governed well enough that an agent can trust it? If the answer to any of those is no, you are not ready.
I am not arguing for waiting. I am arguing for preparation. The difference matters. Waiting is paralysis. Preparation is deliberate work: fixing decision making processes, aligning incentives, practising the behaviour of delegating authority. An agent that operates in a prepared organisation becomes a force multiplier. An agent that operates in a chaotic organisation becomes a chaos amplifier.
The most dangerous phrase I hear is: “Let’s just get something in front of users and iterate.” That works for a button. It does not work for an autonomous system that makes decisions on behalf of the business. The iteration cycle for agentic AI is not hours or days. It is weeks, and the cost of a wrong decision at scale is not a bug fix. It is a business disruption.
The harder part
What does readiness look like? It is not a checklist or a maturity model. It is a few hard questions that most leaders avoid because the answers expose their own gaps:
- Do your product managers own outcomes, or do they just manage backlogs?
- Can your teams say no to a stakeholder without escalation?
- Is your data trustworthy enough that you would bet a quarter’s revenue on a decision it supports?
If your answer to any of those is a wince, agentic AI will hurt you. Not because the technology is dangerous. Because the technology will reveal how little you have invested in the human systems that make autonomy safe.
I agree with the practitioners who say that with rigorous engineering (guardrails, observability, human-in-the-loop design) the failure rate can drop. But that rigour demands product leadership that already knows how to ship, how to measure, and how to kill a bad idea before it costs a fortune. Most product leaders are not there. And agentic AI will not take them there. It will show them the gap and ask them to cross it.
The rush to agentic AI is repeating the same mistake: assuming the technology is the transformation. It is not. The transformation is the discipline to design systems — human and automated — that reinforce each other. If your product org is dysfunctional, agents will make it worse. Not because they are broken. Because they will follow the commands you gave them, and those commands were always the problem.
And that is a problem no model can solve.
Leave a Reply