AI in production · 2026

Agent systems on the factory floor: where they actually work today

Everyone is piloting AI agents. Very few are running them in production. The data — and early deployments like Bosch’s — show where agents actually pay off, and what separates the plants that capture value from the ones that stay stuck in pilots.

The gap between pilots and production

McKinsey’s State of AI survey (2025) captures the gap in two numbers: 79% of organizations say they are using generative AI, yet fewer than 10% report scaling AI agents in any function. The bottleneck is not the technology — the same survey finds that high performers are nearly three times more likely to have fundamentally redesigned their workflows around AI, rather than layering it on top of existing processes.

That distinction matters most in manufacturing, where an agent that merely summarizes data changes nothing, but an agent wired into the flow of work — reading machine states, triggering actions, escalating to people — changes the numbers.

Where agents pay off first

The strongest early results come from narrow, high-frequency problems — not from “AI everywhere” programs. The reference case is Bosch’s Shopfloor Agent: deployed in its own plants and offered to external customers since late 2025, it helps line staff identify and fix machine errors faster, without deep technical expertise. Bosch reports annual savings of roughly €850,000 per plant from reduced machine downtime.

The size of the prize is set by the size of the problem: in 2025, 55% of US manufacturers experienced unplanned downtime. Every hour an agent shaves off error analysis is an hour of production recovered.

The anatomy of a deployment that survives

What separates Bosch-class deployments from stalled pilots is rarely the model. Working deployments share four traits. They live inside the operators’ existing screens, not in a separate chat window. They read live system state — MES order status, machine signals, quality results — instead of stale exports. They act through the same interfaces people use, under explicit guardrails: opening work orders, adjusting schedules, flagging batches. And they escalate to a named human the moment confidence drops. An agent with those four properties earns trust shift by shift; one without them is a demo.

This is also why McKinsey’s finding that high performers are ~3x more likely to redesign workflows matters more than any model benchmark. If the workflow still assumes someone will read a report and decide tomorrow, the agent’s output has nowhere to land. The flow has to be rebuilt around the loop: detect, decide, act, escalate, log.

A realistic first 90 days

Weeks 1–3: pick the process, agree the metric, capture the baseline — downtime hours, scrap rate, changeover time. Without a baseline there is no honest success. Weeks 4–8: wire the agent into MES and ERP in shadow mode — it recommends, people decide, and every recommendation is scored against what actually happened. Weeks 9–12: switch on narrow autonomy where the shadow record is strong, keep humans on the rest, and review the metric weekly. Scale only what pays.

What this means if you run a plant

Three practical conclusions follow from the evidence. First, pick one process with a measurable cost — downtime, scrap, changeover — rather than a broad “AI transformation.” Second, wire the agent into the systems that already run your operation (MES, ERP, quality), because an agent without system access is a chatbot. Third, agree the target metric before the work starts, and hold the deployment to it.