AI Ready Is Not Decision Ready: Metrics and a Ladder of Autonomy

On 30 October 2026, IIoT World published a write-up of a panel from the Industrial AI Summit 2026, authored by Paul Viorel. A note on method: the page is labeled "Sponsored by Arch Systems", and the article states that AI tools helped summarize the discussion. It should therefore be read as sponsored content. The ideas, however, are useful to anyone who works daily with automation and supervisory systems.

The panelists were Tim Burke (Arch Systems), Abby Eon (Velotic), Bryce Barnes (LG CNS) and Sahil Yadav (AOI), moderated by Jeff Winter of IIoT World.

The gap between adoption and scale

The article cites two surveys; we report their figures as quoted, without independent verification. According to Parsec's 2026 manufacturing report, 72% of manufacturers have adopted AI in some form, but only 10% have deployed it in production at scale. According to Grant Thornton's 2026 AI Impact Survey, 62% focus AI on operations, but only 7% have tested an incident response plan for AI-related incidents and only 12% believe they could pass an independent governance audit.

"AI ready" versus "decision ready"

For the panelists, a system is AI ready when the model can produce an answer. It is decision ready when someone, based on that answer, changes a schedule, dispatches a crew or adjusts a quality process. Technically correct outputs often go unused because nobody trusts them enough to act.

Before AI influences operations, five conditions must hold. It must be:

• grounded in real data;
• contextualized;
• explainable to people who did not build it;
• governed, with clear accountability;
• limited in the scope of its actions.

Eon singled out the ability to flag uncertainty as the most important. A sensor reading of zero is not the same as a failed sensor, and a model that cannot tell the two apart will confidently give the operator wrong instructions.

"Operators do not trust systems that are always sure. They trust systems that can say 'I don't know right now.'"

Three metrics instead of accuracy

Yadav proposes measuring AI with operational indicators rather than model accuracy alone:

• Mean time to explanation: how long it takes a person to understand why the AI made a recommendation. If it exceeds the decision window, the recommendation is operationally irrelevant.
• Trust-adjusted risk: the monetary cost of an AI error, including reputation, supply chain and regulatory and legal exposure.
• Adoption rate: the share of recommendations operators actually act on. A highly accurate but rarely adopted model signals a trust deficit.

All three should be measured before introducing AI, and the test cases should be chosen by the operations lead.

A four-level ladder of autonomy

Each level requires specific evidence:

• Observe: the data picture is accurate.
• Recommend: precision against real historical incidents already labeled by people.
• Act with supervision (human confirmation): circuit breakers, complete audit trails, explainability and a limited blast radius.
• Self-heal: an extended track record, fully reversible actions and trust-adjusted risk within accepted limits.

Outlook and advice for scaling

The panelists expect that within about three years generative AI will change how operational decisions are selected and presented, but the decisions will remain human. LLMs predict the next token, whereas control logic requires deterministic outcomes: AI will therefore choose which deterministic workflow to run and propose it to an operator.

To scale, the advice is to attach AI to a shared enterprise architecture, unify one or two existing systems and deliver one use case from that foundation. The anti-pattern is a three-year, IT-only data integration project that delivers nothing along the way. At AOI, the most experienced engineers were the most skeptical; starting with a boring, non-critical task won them over.

On individual answers, the main barrier to scale is organization (Burke), context (Eon), data (Barnes) and trust (Yadav). Among the first use cases to scale, the panel cited quality root cause analysis, downtime attribution, workflow assistance and predictive maintenance. Considered overhyped: humanoid robots, copilots on messy data, LLMs themselves and autonomous agents on the shop floor.

Why it matters for automation professionals

The value lies in concrete acceptance criteria and a gradual rollout, which also helps justify projects to management. The link to SCADA and PLC work is direct: level 1 depends on data quality and on handling the "failed sensor versus zero value" case; level 3 depends on audit trails and circuit breakers. The practical takeaway: before a pilot, define the three baseline metrics, start at Observe and promote a use case to the next level only with the required evidence.