Industrial AI: SLMs at the Edge, LLMs on Top of Solvers, Testing on the Digital Twin

On 14 October 2026, IIoT World published a Q&A drawn from the session "From AI Pilots to Profit: Unlocking Manufacturing's Hidden Data Advantage", held at the Industrial AI Summit 2026. The respondents are Neil Smith (Segment President, CPG, at Schneider Electric) and Andrew Waycott (President and co-founder of TwinThread). It is worth stating up front that both companies sponsored the session, together with TEEPTRAK, so the answers are vendor viewpoints and partly promotional.

Small models, close to the process

According to Smith, value rarely comes from huge general-purpose models. It comes from focused models trained on domain data and deployed close to the process: so-called small language models (SLMs) at the edge. They need less data and, he says, often deliver a return "in weeks, not years". The benefit of local execution is concrete: lower latency, operation even with limited connectivity, and sensitive operational data that stays under the manufacturer's control. As platforms, he cites EcoStruxure Automation Expert and Foxboro Software-Defined Automation.

The barriers are foundational, not technological

The Schneider 2026 Industrial AI in CPG study, covering 1,453 manufacturers, indicates that the obstacles are mostly foundational:

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

The data is useful because it shifts the focus: before the model, people and data preparation matter. Smith also cites AVEVA PI System and AVEVA Connect as sources of the context (process, batch, materials) that AI needs, a topic already covered in earlier articles on this blog.

AI embedded in existing workflows

The applications gaining traction, according to Smith, are narrow capabilities built into existing workflows and automation: failure prediction, line-by-line energy optimization, and automated quality checks. Schneider cites its own Industrial Copilot, which it claims saves up to 50% of engineering time. This is a company claim and has not been independently verified.

LLMs on top of solvers, not instead of them

On production planning, the message is clear: an LLM should not replace optimization or 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 shop-floor actions. The final decision stays with people.

The solver finds the best answer; the LLM explains it and makes it usable.

Validate on the digital twin before production

Waycott describes TwinThread's "Execute over time range" function: it lets you run models on historical data to test anomaly scores, forecasts and recommendations, identifying unsafe settings or unintended consequences before release. Thresholds are then tuned and time-boxed controlled trials follow, before full adoption.

Key takeaways

Setting aside the commercial slant, the pattern is reasonable and repeatable:

• small, focused models close to the process, for latency, continuity and data control
• LLMs as an interaction layer on top of deterministic optimizers
• historical replay on a twin, then controlled trials, before rollout
• priority investment in skills and contextualized data

Figures on payback times and savings should be treated with caution until confirmed on your own plant. The phased approach, by contrast, reduces risk regardless of the vendor you choose.