Why Industrial AI Pilots Stall Before Production

On November 2, 2026, IIoT World published an article by Paul Viorel about a panel at the Industrial AI Summit 2026. The participants were Benson Chan (Strategy of Things), Flemming Kongsberg (Siemens), Maddie Zeng (NeoAI) and Jonathan Weiss (Cumulocity). The topic: why so many industrial AI pilot projects work in a demo and then never reach production.

The Pilot Iceberg

Chan compares the pilot to the visible tip of an iceberg. A pilot costing roughly $10–30K (the article's example is $20K) works with data that is already clean, one or two integrations and a narrow scope. Moving to production adds everything that sits below the surface:

• cybersecurity requirements
• legacy plants running different software versions
• integration with MES, ERP and scheduling systems
• governance and training
• change management for staff

The article states that the production phase typically costs a multiple of the pilot and takes much longer, but it gives no precise figures. The claim should therefore be read as a qualitative indication, not as a budgeting rule.

The Data Readiness Test

The panel suggests three questions to ask before starting a project:

• Is the data available?
• Can you trust that data?
• Is anyone already using it, even manually?

If the data doesn't exist, the project is not an AI project but a modernization project, and it comes before AI. The advice is to start with the use case where data is already flowing, rather than waiting until everything has been modernized.

What Kills a Project After a Good Demo

Three structural causes emerged from the discussion:

• No business owner. Projects that span design, production and operations stall at the boundaries between departments when nobody owns the metric.
• Generic models. A general-purpose model gives no competitive advantage until it is combined with company data and domain knowledge.
• Prototype treated as a product. A demo built with a coding assistant is not enterprise-grade software.

Brownfield vs. Greenfield

In brownfield sites, sensors are now cheap, and adapters for serial, USB and legacy interfaces make it possible to extract data from machines 40–50 years old. Retrofitting is easier than many teams assume. In greenfield sites the situation is not automatically better: even new plants mix PLC brands and MES software on the same site, and integration remains expensive.

In both cases, what matters is whether the data exists and whether it is being used.

Why This Matters for Automation Professionals

The takeaway is that a pilot's success is decided less in the model and more in the industrial work around it: PLC and MES integration, data standardization, assigning an owner and retrofit connectivity. These are typical skills of people who know both the shop floor and plant software.

Practical Guidance

Before approving a pilot, check that:

• the three data tests are passed for the chosen use case
• there is a single owner of the business metric
• the budget covers integration, security, training and change management, not just the demo
• the model leverages the company's own data and process knowledge
• the path from prototype to production software is defined from the start

Note: the original article is a panel report and its cost guidance is qualitative; the pilot budget range comes from the article's FAQ.