Sixty Picks an Hour, and the 88 Percent That Never Graduate

Sixty Picks an Hour, and the 88 Percent That Never Graduate

In January 2026, a humanoid robot stood in front of a tote rack at Siemens' electronics factory in Erlangen, Germany, and worked for eight hours without a supervisor's hand on the controls. It moved roughly sixty containers an hour and got the pick right more than nine times out of ten. Siemens described the trial in its April announcement with NVIDIA as proof that the factory floor was ready for something it had been promised for a decade: adaptive manufacturing run by an "AI brain" that reconfigures itself in real time. CES headlines obliged. So did the press picks.

I read the same numbers and reach a different conclusion. The Erlangen trial is real, the throughput figure is real, and the partnership math at the CES announcement — nine new copilots across Teamcenter, Polarion and Opcenter, all wired into NVIDIA's industrial stack — is also real. But the number that should sit on the same page as "60 picks an hour" is "88 percent." That is the share of AI agent pilots across enterprise — including manufacturing — that never reach production at all, per a March 2026 survey of 650 enterprise technology leaders. For every thirty-three pilots a company launches, four make it. The rest stop somewhere in the gap between a vendor demo and a maintained system that runs on a Tuesday in August when the data pipeline is broken and nobody can find the eval suite.

The marquee deployments are not the market

Let me list the manufacturing AI stories the trade press has run hot on in the past few weeks, because they share a pattern.

Figure AI's first-generation humanoid, Figure 02, contributed to the production of more than 30,000 BMW X3s at the Spartanburg plant — over 90,000 sheet metal components moved, about 1,250 operating hours across roughly 10-hour weekday shifts (Figure AI). BMW will move to its Figure 03 successor at Plant Leipzig, with a pilot phase planned for summer 2026 for high-voltage battery assembly. On 13 May 2026, a younger British company called Humanoid signed a binding agreement with Schaeffler for the deployment of 1,000 to 2,000 wheeled humanoid robots across Schaeffler's global plants by 2032, with the first units arriving at Herzogenaurach and Schweinfurt between December 2026 and June 2027 under a robot-as-a-service model. NVIDIA, in October's GTC and again at Computex, made its Omniverse DSX blueprint generally available, with Siemens, Schneider Electric, AVEVA, Vertiv and Cadence among the partners contributing assets and platforms to a digital-twin layer being marketed as the substrate for the next generation of factories.

This is a real, expensive, technically impressive set of deployments. It is also, almost entirely, the activity of perhaps three dozen industrial buyers — global Tier-1 automotive, top-quartile aerospace, the big electronics OEMs, a handful of energy majors. The reason these stories all read alike is that they come from the same address book.

The rest of manufacturing — the German Mittelstand, the Italian engineering shops, the family-owned American precision machining houses, the Vietnamese contract assemblers, the Mexican near-shore plants whose order books just exploded — is somewhere else in the journey. They are running pilots. They are watching pilots die. The reasons their pilots die are not glamorous.

Where pilots die

A pilot dies when the integration cost into the legacy MES turns out to be eight times the model cost. It dies when the system that hit 98 percent accuracy in the vendor sandbox produces 75 percent on the floor, because the light is wrong and the parts are dirty and the camera vibrates. It dies when the team that ran the proof of value gets pulled onto an SAP S/4 migration and the vendor's customer success contact takes a job at a competitor. It dies, most often, because no one in the plant owns the model after week six.

The five repeating causes pull together from across the recent literature: integration complexity with legacy OT, inconsistent output quality at production volume, missing monitoring, unclear organizational ownership, and not enough domain data (Wizr.ai). None of these are model capability. The frontier model is not the bottleneck. The plumbing is.

The funnel is rough but not dishonest. The 88 percent figure is the pilot-to-production survival rate at the entire industry level; the long-tail attrition after production is a number I have stitched together from operations conversations and vendor renewal data rather than a single published source, so treat the right-hand side of the chart as my read rather than a citation.

The honest measurement problem reaches the factory floor

The vendors will tell you AI predictive maintenance returns ten to thirty times its cost within 12 to 18 months, that companies are halving unplanned downtime, that 75 percent of adopters see positive ROI in under six months (IIoT World). I have spent thirty years watching this category. Predictive maintenance is the one part of industrial AI where the math has been defensible for the longest — Kaeser Kompressoren was selling availability as a service before the deep-learning revival began. The numbers above are not lies. They are the numbers from the projects that worked. The denominator hides off-screen. The 88 percent failure rate is the part of the denominator that does not get reported by the vendor, because the vendor charges by the deployed asset and stops counting on the ones that never deployed.

This is the honest measurement problem in industrial form. Self-reported ROI from companies who finished the deployment will always outperform instrumented results across the whole population. The two numbers are not contradictory; they describe different bases. The buyer responsible for AI investment in a plant should know which one is being quoted before signing.

The unglamorous architecture that actually ships

Here is where it gets uncomfortable, and I mean uncomfortable for the vendors selling the cinematic version of industrial AI. The systems that ship and stay shipped in 2026 are not gigantic frontier models talking to OPC-UA through a copilot pane. They are small, specialised language models running on edge hardware at the cell or line level, with cloud LLMs reserved for the long-tail of ambiguous queries.

A three-billion parameter model identifying bearing wear from vibration patterns in eight milliseconds on a two-thousand-dollar edge device with no internet connection is unsexy and works (iFactory). Inline vision-language inspection at 500 units per minute, with the heavy reasoning model on standby for cases the small model flags as uncertain, is unsexy and works. Edge-deployed small models give you the data-sovereignty answer the EU AI Act high-risk pathway is going to start asking for; they give you the cost answer your CFO is going to start asking for the moment the inference bill on a vendor copilot crosses a meaningful threshold; they give you the determinism answer your safety engineer is already asking for.

The shape of the winning industrial AI architecture in 2026 is therefore a small model at the edge, a model registry that knows which version of which model is running on which line, an eval suite that runs on every shift change, and a clear OPC-UA gateway between the deterministic control layer and the probabilistic intelligence layer. That stack, written down on one page, is the actual industrial AI operating system. It does not photograph as well as a humanoid robot.

My stake

I would bet against the timelines on every marquee humanoid deployment slipping by less than nine months — the Schaeffler December 2026 date, the Leipzig summer 2026 date, the Tesla Optimus internal targets. Industrial timelines slip; integration timelines slip more; safety certification timelines slip the most. The deployments will happen. They will be late.

I would bet, harder, that the manufacturers who get useful AI into production over the next eighteen months will not be the ones who buy the most copilots. They will be the ones who do three boring things: assign a permanent line-of-business owner to every model that goes onto a line, instrument their pilots with the same telemetry they use for OEE, and refuse to start any new pilot until the previous one has either reached the floor or been formally shut down.

If I sat on a CapEx committee for a mid-sized manufacturer this quarter, I would not approve any new vendor pilot without three things written down: an integration cost ceiling, a named internal owner who reports to operations rather than to the CIO, and a kill date. Without those three lines on the slide, the pilot is going to become slop debt — the kind of half-finished AI initiative that consumes budget, generates dashboards, and never reaches the floor.

The factory floor is finally getting useful AI. It is arriving the unglamorous way.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.