Why AI Pilots Stall Before Production: The Cost Iceberg

On November 2, 2026, IIoT World published an article by Paul Viorel summarizing a panel at the Industrial AI Summit 2026. The panelists were Benson Chan (Strategy of Things), Flemming Kongsberg (Siemens), Maddie Zeng (NeoAI) and Jonathan Weiss (Cumulocity). The topic will be familiar to anyone working on a factory floor: many AI projects perform well in a demo and then never reach production.

The pilot iceberg

Chan describes a pilot of roughly $20,000 as the visible tip of an iceberg. The pilot runs on data that is already clean, with one or two integrations and a narrow scope. The submerged part is the production environment, which includes:

• cybersecurity requirements;
• legacy machinery running different software versions;
• integration with MES, ERP and scheduling systems;
• governance and staff training;
• resistance to change from the workforce.

According to the article, production typically costs "multiples" of the pilot budget and takes much longer. Note that the text cites pilot budgets between $10,000 and $30,000 but does not quantify the cost of production: the multiplier remains qualitative.

A three-question data readiness test

To assess a use case, the panelists suggest three questions:

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

If all three answers are yes, the foundations for AI exist. If not, the work to be done is data modernization, and it comes first. The advice is to pick the use case where data already flows, rather than waiting to modernize everything. The example given is a plant with 16-20 unconnected machines where scrap is estimated by looking at the waste bin: here the problem is not the model but the absence of data.

Three structural causes after a good demo

• Unclear ownership: no business owner is accountable for the metric, and the project stalls at the boundaries between departments.
• Generic models: they provide no competitive advantage until the company adds its own data and domain knowledge.
• Prototype mistaken for product: an application built quickly with a coding assistant or a chatbot session is not production-ready software.

Brownfield vs. greenfield: what matters is whether the data exists

According to the panel, both scenarios are harder than they look. In existing plants, retrofit sensors and adapters for serial, USB and legacy interfaces make it possible to extract data from machines that are 40-50 years old. New plants still end up with a mix of PLC brands and MES software, with a similar integration effort. The deciding factor is not the age of the machines, but whether the data exists and is being used.

Putting it into practice

The article is a report of a panel, with no survey or measured data, and should be read as the experience of industry practitioners rather than quantitative evidence. It does offer a useful checklist for technical and IT leaders:

• build the production budget from the start, including OT security, integration and training, not just the pilot cost;
• apply the three-question test before approving a use case;
• appoint a business owner accountable for the metric;
• plan how to train or fine-tune models with your own data and process knowledge;
• treat every prototype as a prototype, with an explicit engineering path to production.

The underlying message is that the real work of industrial AI lies in integration, OT security and organizational accountability more than in the model. Teams that budget for this up front are more likely to get past the pilot stage.