The pitch is seductive: autonomous agents that plan, buy and replan while you sleep. Then an operator asks the one question that ends the demo, and the agent cannot answer it.
After two years of copilots, the vendor story moved on to agents: software that does not just suggest but acts, chaining decisions and taking steps on its own. In planning that means an agent that reads demand, places orders, adjusts buffers and reacts to disruptions without a human in the loop. It demos beautifully. It is also, increasingly, being switched off.
Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. It also flagged how much of the market is “agent washing,” older chatbots and automation relabelled as agents; by its estimate only around 130 of the thousands of self-described agentic vendors are the real thing. In 2026 the firm added a sharper warning: enterprises are finding the governance gaps in autonomous agents only after a production incident, and then demoting them.
We sit across the table from the people who run mid-market manufacturers and distributors, and we have watched more than one of them sit politely through an agent-driven pitch and leave unconvinced. The reason is always the same, and it is not fear of technology. These are operators who would love to spend less time firefighting. What stops them is a single question: show me how it decided, and show me it again.
An autonomous agent that places a purchase order has made a decision with money and shelf space attached. When the owner asks why it bought what it bought, and whether the same situation would produce the same move, the honest answer from most agents is that it depends on the model's state at the time. That is disqualifying. You cannot run an operation, brief a board or pass an audit on a decision that will not reproduce and cannot be read.
The backlash is often described as being about cost or hype, and those play a part. But in planning the binding constraint is narrower and more stubborn: reproducibility. A forecast or a reorder is not an interesting opinion; it triggers commitments, ties up cash and sets service levels, and everything downstream inherits whatever doubt sits in the original number. If the number cannot be reproduced on the same inputs, the doubt never clears.
Regulators are formalising exactly this. The EU AI Act requires high-risk systems to keep automatic logs so their behaviour can be traced and reconstructed. In regulated sectors, data-integrity rules already demand that any record be attributable, consistent and reproducible. An agent whose decisions cannot be replayed is not merely risky; it is, in those settings, non-compliant by construction.
None of this means the learning is worthless. It means the architecture has to change. Separate the part of the problem that benefits from learning from the part that must be deterministic. Let models find patterns and sharpen the demand signal, where more data genuinely helps. But let the decisions that commit money, how much to hold, when to reorder, how much to free, run on transparent math that returns the same answer every time and that a human can read and overrule.
That is not a retreat from AI; it is what makes AI usable in a business that is audited and cash-constrained. The agent can still draft the plan and flag the exceptions; it just no longer gets to make an unexplainable decision in the dark. The companies that come through the agentic backlash with something working will be the ones that kept the intelligence and threw out the opacity. The model proposes; a method you can reproduce disposes.
Gartner expects over 40% of agentic AI projects to be scrapped by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. It also notes heavy “agent washing,” with older tools relabelled as agents, and warns that enterprises often find governance gaps in autonomous agents only after a production incident.
An autonomous agent that places orders or adjusts buffers is making decisions with money attached. If the same inputs can produce a different decision, and no one can reconstruct how it decided, the output cannot be reproduced, explained, audited or defended. In planning, where every downstream commitment inherits that doubt, this is disqualifying.
No. The workable architecture separates learning from deciding: machine learning sharpens the demand signal, while the decisions that commit money run on deterministic, transparent math that returns the same answer every time and that a human can read and overrule. You keep the intelligence and remove the opacity.
Yes. The EU AI Act requires high-risk systems to keep automatic logs so their behaviour can be traced and reconstructed, and data-integrity rules in regulated sectors require records to be attributable, consistent and reproducible. An agent whose decisions cannot be replayed is non-compliant by construction in those settings.
The companies that come through the agentic backlash with something working will be the ones that kept the intelligence and threw out the opacity.
Picture your next S&OP meeting: one number on screen that nobody argues with, and a trail behind it that an auditor could follow. See it free, on your own data, in 45 minutes; and if it does not hold up, we part as friends.