LLMs and Scheduling: Explain the Scheduler, Don't Replace It

On 14 October 2026, IIoT World published a Q&A with Neil Smith (Segment President, Consumer Packaged Goods, Schneider Electric) and Andrew Waycott (President and Co-founder of TwinThread). The five answers stem from audience questions raised during the session "From AI Pilots to Profit: Unlocking Manufacturing's Hidden Data Advantage" at the Industrial AI Summit 2026, sponsored by TEEPTRAK and TwinThread. Since both speakers are vendors, their positions inevitably carry a commercial slant, but the picture that emerges is useful for anyone working with MES and production planning.

LLMs as an interface, not as the engine

The most relevant point for MES practitioners concerns production planning. According to Smith, large language models should not replace optimization and scheduling engines, which remain the components responsible for determining the best answer. The role of the LLM is that of an interface:

In practice, the LLM can:

• explain why a schedule has changed;
• let the planner ask "what happens if this order slips?";
• translate a recommendation into concrete actions for the shop floor.

The final decision stays with people. Smith argues that this approach matters most in high-volume, low-margin plants and in engineered-to-order plants, where choices about sequencing and changeovers must be precise and explainable.

The LLM explains and supports; the optimization engine decides what the best answer is.

Validating AI on a digital twin

Waycott describes a low-risk path before anything goes live on the line. With TwinThread's "Execute over time range" function, models are run on historical data to check anomaly scores, predictions and recommendations, catching unsafe settings early. Thresholds are then tuned, followed by time-limited controlled trials before full rollout.

Small models, at the edge, inside existing workflows

Smith argues that value comes from specialized, domain-specific models rather than large general-purpose ones: they need less data and, in his words, often show a return "in weeks, not years." Running them at the edge reduces latency, allows operation even with poor connectivity, and keeps sensitive operational data under the plant's control.

The AI that gets adopted, he adds, is embedded in existing operational workflows, narrow in scope and built on automation already in place: failure prediction, line-by-line energy optimization, routine quality checks. He also cites Industrial Copilot, which he says could save up to 50% of engineering time, and ProLeiT's CIPOptimize for Clean-in-Place and recipes. These are vendor claims, not independently verified.

The real obstacle is readiness

The Schneider Electric 2026 Industrial AI in CPG study, covering 1,453 manufacturers, indicates that the barriers are mostly about readiness, not technology:

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

For this reason, Smith stresses data contextualization, citing AVEVA PI System for process, batch and material data in audit-ready form, and AVEVA Connect to extend it to suppliers.

Key takeaways

For those running MES and planning, there are three practical takeaways:

• use generative AI as an explanation and what-if layer on top of the scheduler, without replacing it;
• test models and thresholds on historical data or a digital twin, then in controlled trials, before touching a real line;
• invest first in data quality and context: without them, no model will deliver reliable results.

Note on method: this article is based on an interview with interested parties and on data from a vendor study; the results cited should be assessed against your own production context.