Stalled AI Pilots: A Three-Question Data Readiness Test
An IIoT World article dated 2 November 2026, based on a panel at the Industrial AI Summit 2026, explains why AI pilots fail to reach production and proposes a data maturity test. For plants old or new, the starting point is machine connectivity.
On 2 November 2026, IIoT World published a summary of a panel from the Industrial AI Summit 2026, featuring Benson Chan (Strategy of Things), Flemming Kongsberg (Siemens), Maddie Zeng (NeoAI) and Jonathan Weiss (Cumulocity). It is not new market research but a report of practitioners' opinions. Still, the guidance is concrete and useful for anyone deciding whether and where to fund an artificial intelligence project on the shop floor.
The Pilot Iceberg
Chan compares a pilot to an iceberg: the visible part is small, while the operational requirements stay beneath the surface. A pilot costing around $20,000 (the article cites a range of $10,000-30,000) works with already clean data, minimal integration and a narrow scope. Moving to production, however, requires:
- cybersecurity;
- handling machines running different software versions;
- integration with MES, ERP and planning systems;
- data and model governance;
- training and organizational change for staff.
According to the panel, production typically costs "multiples" of the pilot and takes much longer. The article gives no precise multiplier, so this should be read as a qualitative order of magnitude, not a figure to put in a budget.
Three Questions Before You Invest
The most practical contribution is a data maturity test to apply to the candidate use case:
- Is the data available?
- Can it be trusted?
- Is someone already using it, even manually?
If all three answers are yes, the foundation for AI exists, and it makes sense to start with the use case where data already flows. If instead the correctness of the numbers is still debated, or the data does not exist, the first step is a modernization project. The example cited is plants with 16-20 unconnected machines, where scrap is assessed by looking at the waste bin.
Why a Successful Demo Is Not Enough
The panel also identifies three structural causes of stalling after a positive test:
- unclear ownership: no business owner is accountable for the metric, and the project stops at the boundary between departments;
- generic models: they give no competitive advantage until combined with the company's own data and process knowledge;
- prototype mistaken for product: a demo is not enterprise-grade, governed software.
Old and New Plants: Data Matters, Not Age
An interesting point concerns the distinction between brownfield and greenfield. An older plant is not necessarily "too old": sensors are cheap today, and adapters for serial, USB and other legacy interfaces can extract data even from machines 40-50 years old. Conversely, a new plant is not ready by definition: even modern sites mix different PLC brands and MES software, and integration demands significant engineering effort.
What counts is whether the data exists and is used, not how old the machine is.
What to Do in Practice
For plant managers and IT/OT leads, the test is a simple screening tool to use before approving a budget. It can be applied to each candidate: where all three answers are yes, the risk of stalling after the pilot is lower; where they are no, spending should go first to connectivity and reliable production data, for example through MES and IIoT. A business owner for the metric should also be assigned from the start, and the cost of industrialization should be budgeted, not just that of the pilot.
The underlying message, consistent with the panel, is that AI does not replace foundational work: it presupposes it. Low-cost retrofit makes this path viable even for the oldest lines.