Validating AI on a Digital Twin Before Going into Production

On October 14, 2026, IIoT World published a Q&A drawn from the session "From AI Pilots to Profit: Unlocking Manufacturing's Hidden Data Advantage" at the Industrial AI Summit 2026. Five audience questions are answered by Neil Smith (Schneider Electric, Segment President for Consumer Packaged Goods) and Andrew Waycott (TwinThread, co-founder and president).

A necessary caveat: the session was sponsored by TwinThread and TEEPTRAK, and both interviewees represent vendors. Statements about products are vendor claims, not independent evidence. It remains useful as a description of a method.

The validation workflow on the digital twin

According to Waycott, before go-live the recommendations of an AI model can be simulated and validated on a digital twin. In the TwinThread platform, the "Execute over time range" function runs models on historical data, letting engineers check anomaly scores, predicted values and recommendations.

This, he argues, surfaces unintended consequences or unsafe settings early. The team can then adjust model configurations and acceptable ranges or thresholds. Only afterwards does it move to controlled, time-limited trials, and then to full rollout. Waycott says this path reduces operational risk, but he provides no quantified results.

In practice, the process can be summarized as follows:

• replay the model on the digital twin's historical data;
• review anomalies, predictions and recommendations, looking for abnormal or unsafe settings;
• tune configurations, ranges and acceptable thresholds;
• start controlled, time-limited trials;
• extend the rollout only after these steps.

LLMs as an interface, not a solver

On production planning, Smith states that language models should not replace optimizers and scheduling solvers. The solver finds the best solution; the LLM acts as the interface: it explains why a plan changed, answers questions such as "what happens if this order slips?" and translates recommendations into actions for the shop floor. The final decision stays with people.

It is an interesting design pattern: optimization remains deterministic and verifiable, while natural language makes the results easier to access.

Small models, close to the process

Smith argues that value comes from domain- and task-specific models rather than large general-purpose ones, that they require less data, and that they can show a return within weeks. Running on the edge keeps decisions close to the process, reduces latency, works with limited connectivity and leaves sensitive data under the manufacturer's control. He cites Schneider's EcoStruxure Automation Expert and Foxboro Software-Defined Automation.

Barriers to adoption

Schneider's 2026 study of industrial AI in CPG, which surveyed 1,453 manufacturers, identifies these barriers:

• shortage of AI or data science skills: 43.0%;
• legacy automation and infrastructure: 37.5%;
• lack of contextualized operational data: 36.3%;
• workforce resistance: 25.7%;
• cybersecurity or compliance concerns: 21.7%.

The study is Schneider's and covers the CPG sector, so its figures should be generalized with care. Smith adds that the AI uses with the most traction are embedded in existing workflows and narrowly scoped, such as predictive maintenance, line-level energy optimization and automated quality checks. He also states that Schneider's Industrial Copilot can save up to 50% of engineering time: a vendor-declared figure.

Why it matters for decision-makers

The most useful message for technical and IT leaders is the sequence: replay on historical data, tuning, limited trials, then production. It does not depend on a specific product and applies to any model that influences process parameters.

First replay on a digital twin, then controlled trials, finally full rollout.

A practical takeaway follows: before evaluating new models, check that you have historical data complete and contextualized enough for replay, and define acceptable ranges and thresholds in advance. The study's data suggest that skills, legacy systems and data context weigh more than the choice of technology.