Article contribution from Ryan Gordon of foundrynet.io

Walk any modern plant floor and you will find a KUKA arm, a Fanuc CNC, an ABB palletizer, and a Universal Robots cobot working within a few meters of each other. Each was commissioned by a different integrator, in a different year, speaking a different dialect. None of them agree on what a word like temperature or cycle time means, and most were never designed to coordinate with anything outside their own cell.

That is the floor AI agents are walking onto. And it explains a number that should bother anyone budgeting automation this year.

80% of enterprise applications now embed at least one AI agent (Gartner, Q1 2026). 11% of organizations have gotten them to production at scale, per McKinsey, Deloitte, and Gartner independently. Almost nobody is talking about the 68% in between.

The 68% are not stuck on model quality
The reflex is to blame the AI. The data does not support it.
86% of enterprises require infrastructure upgrades before agents can deploy. 82% have discovered AI agents running in their environments that nobody authorized (Cloud Security Alliance, 2026). Integration with existing systems is the number one deployment challenge for 46% of organizations (2026 State of AI Agents Report). Not reasoning. Not accuracy. These are not model failures. They are plumbing failures. The AI works. The pipes underneath it do not.
For anyone who has commissioned a cell, this will read as obvious. Decades of buildup in proprietary systems became the normal way to solve specific problems in specific plants. A Fanuc controller exposes data through FOCAS. A Siemens 840D does not. An older Haas may give you something over MTConnect if someone enabled it, and a 2009 machine may give you a serial port and a manual the vendor no longer publishes. All of them report spindle load. None of them report it the same way, in the same units, at the same rate.

Those silos were never designed to talk to each other. Now they have to, and the translation layer does not exist.

Why picking a better vendor does not fix it
Gartner estimates that only about 130 of the thousands of agentic AI vendors are real. The rest are engaged in what Gartner calls "agent washing," rebranding existing chatbots, RPA tools, and virtual assistants with new marketing language and no meaningful agentic capability.
The obvious conclusion is to shortlist carefully. But that is not actually where the 68% lost. They did not all pick washed vendors. Plenty picked real ones and stalled anyway, because the infrastructure layer underneath every vendor, real or rebranded, does not exist inside their building.

This is the part that gets missed in evaluations. Feature matrices and model benchmarks measure the demo. The demo was never the hard part. Day one on a mixed floor is.

A more useful question for a shortlist: what does this do when it meets a control system nobody documented, and how do I prove afterward what it read and what it changed?

The governance half nobody scoped
Most companies do not possess identity verification for each agent they deploy within their own walls, let alone scoped permissions with controls that prevent catastrophic actions from being carried out at machine speed.

An agent that drafts a summary is one risk category. An agent that can adjust a setpoint, override a feed rate, or release a hold is another entirely, and it acts faster than anyone can read the alert it generated. Ask who authorized the agent's last action and the honest answer, in most deployments, is a log line if you are lucky.

This is why governance and security remain the second most cited cause of failed deployments. It is rarely obstruction from the security team. It is that there is genuinely nothing to review. No per-agent identity, no scoped permission model, no durable record after the fact.

The practical fix is unglamorous: make the record a property of the read itself, not a reporting feature bolted on later. If every call an agent makes against a machine is hashed and anchored, then "the agent caught a thermal event at 847 degrees and paused the run" becomes a claim a third party can verify independently. Not a screenshot and not a vendor's word.

The problem is semantic, not statistical
The core issue across all of this is semantic, and it is not model specific. What is lacking is the infrastructure underneath the agents being deployed.

A missing horizontal layer means data that was never structured for machine ingestion gets misunderstood, confidently, by the systems trying to act on it. That confidence is the dangerous part. An agent reading an unlabeled PWM value as a percentage when the underlying scale runs 0 to 127 will not hesitate or flag uncertainty. It will report a number that is wrong by a third and act on it. Multiply that across a mixed cell and the failure mode is not an agent that refuses to work. It is an agent that works incorrectly, quietly, at speed.

For reference on scale: doing this properly across a mixed floor currently means on the order of 14 protocol adapters, 18 OEM families, and 16,908 curated mappings. That is the size of the problem the 68% are being asked to solve internally, as a side project, with the same overloaded controls engineer who keeps the line running.

40% of agentic AI projects are projected to be canceled by the end of 2027 (Gartner), driven by escalating costs, unclear business value, and inadequate risk controls. The pattern is consistent: organizations invest in the intelligence layer while neglecting the infrastructure that makes intelligence operational.

The 68% do not need better models. They need better plumbing. The layer that closes this gap, from data normalization to agent identity to verifiable settlement, is the difference between AI that demos well and AI that runs in production.