A pilot that worked at one site says little about the other twelve.

• Whether the workforce accepts it: crews, shifts, union agreement, and whether people see it as a threat to their jobs.
• Whether the maintenance and reliability team will run the system after the vendor leaves, and which integrator or distributor supports it in each region.
• What each site's legacy control systems and brownfield network can support, and whether the OT security owner at each site, not only at the pilot plant, has approved the connection. A security review that approves the pilot does not approve the network.
• Whether each plant controller signs the savings (downtime, scrap, payback) against a baseline measured the way the pilot's was, and from which budget: a capex request in this year's capital cycle, or a subscription carried in the plant's operating budget.

Why pilots stall between the first site and the second: → Too successful to stop, too unproven to scale

Offered an outcome-based service contract by an equipment maker? The same test applies from your side: who sets the uptime baseline, who measures it, and what the penalty pays. → The servitization bet

→ Through the integrator, on what to ask the partner who will support you

Thing Company's founder works today inside an industrial IT and automation group that integrates control systems, machine vision and robots and runs managed services for plants. Before that, as a partner at an industrial venture capital firm, he did commercial and technical diligence on products that run in plants and in the field. Current roles are disclosed in writing before any engagement they could affect. → The record behind it

How do you know if an industrial AI pilot will scale to other plants?

Test what changes from site to site before the rollout: whether the workforce accepts it, whether maintenance will run it, what each plant's legacy controls and network can support, and whether each plant controller signs the savings. A pilot that worked at one site is evidence about that site.

Do manufacturers buy industrial AI and automation as capex or as a subscription?

Mostly they buy the equipment as capex and buy the service as opex. A capital purchase is depreciated over years and sits below EBITDA, and in the US the July 2025 tax law made 100% bonus depreciation permanent for qualifying equipment acquired after January 19, 2025, which improves the after-tax payback of owning; your tax team confirms what qualifies. A subscription that replaces a cost the plant already carries, such as labor, maintenance, spare parts or downtime, sells more easily than one that asks finance to move an asset off the capital budget. Which form each plant will approve is part of the test.

Will an outside review sink a rollout I am sponsoring?

Three of the four verdicts keep it going, and the verdict is a recommendation: you and your funder decide. What changes is what you take to the capital board: a case graded plant by plant, so the plants that are ready can go ahead while the others get a dated test. The brief is yours.

Will plant managers speak candidly if the sponsor reads the brief?

The brief cites role and site, not names, unless the person agrees to be named. A plant manager can say what will not carry over to their site without it being read as a position against the sponsor.

What does a rollout Sprint need from our team?

The names of the plant managers, plant controllers, maintenance leads and OT security owners at the sites in scope; a lead on your side for four to six hours a week; and a sponsor who can convene the decision. A Sprint needs no access to plant systems or networks; it works from interviews and the documents you choose to share.

Why do industrial AI rollouts stall at security review?

Usually because the review approved one pilot, often on an isolated network, and the rollout needs a pattern every site can accept. In Cisco's 2026 research with more than 1,000 OT decision-makers, 40% named cybersecurity as the biggest obstacle to scaling AI. The test is whether each site's OT security owner has reviewed the design against your own standard, such as NIST SP 800-82 or ISA/IEC 62443, before the capital request rather than after it.

What payback do manufacturers need to approve an automation project?

Each company sets its own hurdle. One accounting firm's automation guide puts typical payback at 12 to 18 months without naming its data; other practitioner guidance runs from under a year to more than two, longer for complex lines. An AI project at a single site often does not pay back on its own until its cost is spread across sites. The test is whether each plant controller will sign the savings against your hurdle, not whether the vendor's model shows them.

Who has to approve an industrial AI project in a plant?

More people than the champion. Operations, IT and operational technology (OT), environment, health and safety (EHS), procurement, engineering and finance can each stop a project, and a plant manager often holds a site-level veto. In our experience, deals rarely die at evaluation. They die in procurement, when a specific budget line has to produce the money, or in integration, when the OT team is consulted for the first time. The paper The committee nobody mapped sets out each seat.

How do I judge whether a vendor's pilot results are real?

Ask what at the pilot site made each benefit possible, and whether the next site has it. The result is real but describes one site: the vendor's engineers are on site, a motivated sponsor clears obstacles, and the line was often chosen because it was likely to succeed. Sort each benefit as one that travels, one that depends on a condition, or one that belonged to the pilot, and check that every site measures the baseline the same way.

Why should frontline interviews happen without a manager in the room?

In our operator engagements, a technician's or line operator's unprompted reaction is one of the strongest predictors of adoption at network scale. It is most telling before they hear the management framing or sense what answer is expected, so a Sprint interviews frontline workers and line supervisors before management, at more than one site. Management enthusiasm without frontline buy-in predicts a failed rollout just as reliably.

Who owns the return on investment (ROI) number for a rollout?

A named person at each plant has to, and a vendor's forecast cannot stand in for them. If the forecast is the only number, the capital request is unlikely to survive the finance review. A site sheet asks, plant by plant, who holds the budget, whether they have agreed to fund the running cost, and what payback the plant controller needs to sign. Where nobody owns the decision, pilots drift: in pilot purgatory, no one convenes the call to scale or stop.

What should I ask before approving capital for an AI rollout?

Ask for a site sheet for each plant before the decision. It covers who holds the budget and will fund the running cost, which costs are capital and which are operating, the payback the plant controller needs, the integration work this site needs that the pilot site did not, who maintains the system after the vendor leaves, and whether OT security has reviewed the design. Write unknown where an answer is missing, because that is a finding.