What I heard at Made in Connecticut: manufacturers need an AI operating model, not another demo.

What I heard

The conference conversation was about operating pressure, not AI theater.

At the 2026 Made in Connecticut summit, I heard AI discussed beside workforce pressure, rising operating costs, technical skill needs, data quality, cybersecurity, and the practical difficulty of changing how work gets done. My takeaway was not that every manufacturer needs an AI project. It was that experimentation without an operating model will leave teams with scattered tools, inconsistent rules, and no dependable way to decide what should continue.

The official manufacturing report provides the wider context. It describes a state with a highly technical manufacturing base, long hiring cycles, small and midsize suppliers, capital-investment pressure, and a policy recommendation to support AI adoption alongside equipment, technical assistance, cybersecurity, and workforce training. Those findings are statewide evidence, not a diagnosis of any particular company.

What it means

Awareness is useful. The missing layer is repeatable responsibility.

One report-panel discussion moved from early emphasis on prohibited behavior toward safe places for employees to practice, verify results, and learn appropriate uses. The AI panel returned to repetitive work, measurable progress, data quality, subject-matter expertise, training, human review, and change management. I am paraphrasing the sessions rather than attributing an undiarized recording to an individual speaker.

Taken together, those themes point to a system problem. A policy without practice leaves employees uncertain. A workshop without workflow ownership fades. A pilot without a baseline cannot establish improvement. A tool without approved sources and named review authority creates new ambiguity. A successful demonstration without an operating owner becomes a stranded experiment.

A practical framework

A six-part operating model for manufacturing AI adoption.

The model is intentionally cyclical. A source change, tool release, policy update, recurring error, or new use can send the workflow back to an earlier decision. “Deployed” is not the same as “done,” and stopping a weak workflow is a valid result.

Applied-physics lens

Define the boundary before optimizing the system.

My applied-physics and engineering background makes me skeptical of changes that cannot name their boundary, inputs, observable output, uncertainty, and failure conditions. That mindset is useful here. It does not make an AI workflow automatically safe or correct, but it provides a disciplined way to ask what is actually changing and how we would know.

• State the system boundary: one workflow, user group, approved information set, and decision handoff.
• Measure the current state before introducing a tool, including time, rework, queues, interruptions, and common exceptions.
• Change a coherent set of variables and test representative normal, difficult, and failure cases.
• Keep consequential authority with a person who has the competence, evidence, time, and organizational authority to change the outcome.
• Record what failed and what changed instead of treating polished output as proof.

Practical next decision

Begin with one workflow description, not a tool purchase.

A useful first step is to write a one-page current-state description: who performs the work, what triggers it, which sources govern it, where time or errors accumulate, what output moves forward, who reviews it, and what measure would matter. Then ask whether the information is permitted, the errors are detectable, the output is reversible, and an approved tool can be tested without expanding into an unauthorized decision.

If those answers are unclear, the next step is education or assessment. If they are clear, a bounded implementation may be justified. Either way, the operating model should make the reason visible. Connecticut manufacturers do not need to imitate the largest enterprise program; they need a scale-appropriate way to learn, choose, test, and remain accountable.