Industrial AI: Operators in the Loop and Shared Accountability When the System Gets It Wrong

On 15 October 2026, IIoT World published "The Human Side of Industrial AI", an article of about six minutes' reading time. In it, Neil Smith, Segment President for Consumer Packaged Goods (CPG) at Schneider Electric, and Andrew Waycott, President and Co-founder of TwinThread, answer six audience questions raised during the "From AI Pilots to Profit" session at the Industrial AI Summit 2026.

The common thread is clear: industrial AI delivers more when it works alongside people rather than replacing them. It is worth remembering that both speakers represent vendors, so their views are positions, not independent evidence.

Return on investment comes through people

According to Waycott, the highest return comes from augmenting the capabilities of process experts. Smith adds that organizations with the best ROI from AI also report higher workforce engagement and retention, and argues that training should be considered part of the return equation.

Operator in the loop: recommend, don't act

For system designers, the most interesting point is how TwinThread handles process drift. When the system detects it, it does not act autonomously: it sends a recommendation to the operator.

The operator has three options, and the model learns from the choice made:

• approve the recommendation;
• modify it;
• reject it.

Waycott observes that operators already "train" the AI this way: their historical actions define the initial setpoints and operating limits, and engineers then use their feedback to refine the models. Both speakers also believe that embedding veterans' experience in models and workflows helps pass it on to younger staff. For Smith, AI is decision support, while strategic and ethical oversight remains human.

Who is accountable for a mistake?

The most uncomfortable question came from the audience: who is responsible if an AI acting without a human in the loop gets it wrong? Waycott answers that there is no single owner: it is a failure of the workflow as it was built. Responsibilities are distributed:

• data engineers must ensure clean, reliable data;
• model vendors must set thresholds that alert the team when a failure is imminent;
• engineers and operators must have clear procedures to spot anomalies, diagnose them and intervene before an asset or process is compromised.

Workforce resistance

Smith cites Schneider's 2026 study on AI in CPG: 25.7% of manufacturers name workforce resistance as a risk to scaling AI. His advice is to involve operations and maintenance from the selection of use cases and keep them involved throughout rollout, investing as much in change management and skills as in platforms. The final part of his answer is not reproduced here in full.

What engineers and integrators can take away

Beyond the commercial context, a reusable design pattern emerges for AI-based recommendations in the plant: explicit approval points, structured collection of operator feedback, and shared safeguards with defined roles along the data-model-procedure chain.

In practice, it makes sense to log every operator decision along with its rationale, define in advance who is accountable for data quality and alarm thresholds, and involve line personnel early.

Limitations to keep in mind

The article provides no ROI figures beyond the 25.7% cited, so the claimed benefits remain qualitative. The topic also comes from the same summit and the same Schneider CPG context already covered in earlier articles on data and governance of AI agents in OT; the angle here, however, is different: people, accountability and adoption.