Factory ROI of a physical AI pilot

Why start with a tightly scoped pilot

In physical AI, the gap between a demonstration and reliable production remains the main financial risk. A scoped pilot exists precisely to measure that gap on your own ground: your product references, your cycle times, your teams, your real-world variability. The most publicised humanoid deployments of 2024-2026 all followed this path: GXO ran Digit at one site before signing a multi-year agreement, Toyota Canada piloted for about a year before its February 2026 commercial agreement, and BMW validated Figure 02 at Spartanburg before expanding the programme.

A good pilot has three traits: a narrow scope (one cell, one flow, one shift), a single success metric defined before purchase (cost per tote moved, scrap rate, uptime) and written criteria for scaling up or stopping. Without that third element, the pilot becomes a permanent showcase that never decides anything: the "pilot purgatory" familiar to every industrial leadership team.

The use cases that work in 2026

Documented results converge on five families:

• Palletizing and depalletizing: the safe bet. Mature technology, repetitive cycles, well-mapped integration; it is the most frequently cited first project.
• Vision-guided pick-and-place: sorting and packing of variable parts benefits directly from advances in AI perception.
• Vision-based quality inspection: inline defect detection, often the cheapest AI pilot because nothing is being handled.
• Intralogistics: the best-documented humanoid deployment is GXO's, where Digit (Agility Robotics) has been handling totes at the Spanx site in Georgia since June 2024 under a robots-as-a-service model, with more than 100,000 totes moved; in November 2025 Digit also passed the first OSHA-recognised (NRTL) on-site safety field inspection at a live customer site.
• Machine tending: Schaeffler has had Digit robots working 8-hour daily shifts at its Cheraw, South Carolina plant since early 2025.

In automotive logistics, BMW ran Figure 02 at Spartanburg for around ten months: more than 90,000 parts moved over roughly 1,250 operating hours in support of more than 30,000 BMW X3s, before deploying Figure 03 in June 2026 and announcing a first European pilot in Leipzig in February 2026. Keep perspective, though: no multi-site humanoid fleet has been publicly verified to date. In 2026, classic robotics (arm plus AI vision) remains the ROI baseline; humanoids earn their place where multi-task flexibility and human-designed environments dominate.

A realistic cost structure

The most common budgeting mistake is confusing the price of the robot with the price of the project. In an industrial cell, the arm typically represents 25 to 40 percent of total cost according to integrators. One telling public example: a 40,000 dollar arm becomes a palletizing system of roughly 120,000 dollars once gripper, conveyors, safety, programming and commissioning are added. Public estimates for a complete palletizing cell range from about 100,000 to more than 500,000 dollars depending on complexity. Our platforms comparator details arms, cobots and AMRs with their prices.

The full budget of a physical AI pilot therefore includes:

• Robot CAPEX: the arm or platform, sensors included.
• Integration: plan for 2 to 3 times the robot price (tooling, conveying, safety, programming). Buyers who budget only the arm routinely face 50 to 80 percent overruns at integration time.
• MLOps and software: for a learning-based system, data collection, retraining, drift monitoring and edge-case handling are recurring costs, not a one-off purchase.
• Training and change management: operators, maintenance, process engineering.
• Compliance: risk assessment, CE marking of the cell, documentation (see our AI Act and standards pillar).

An emerging alternative: robots-as-a-service turns CAPEX into OPEX through a subscription covering hardware, software and support, on the model pioneered by GXO and Agility Robotics in 2024. Useful for a pilot; re-run the economics before scaling.

This is not a footnote: in Deloitte's State of AI in the Enterprise 2026 survey, cost is the barrier to physical AI deployment most often cited by the 3,235 executives polled. The report recommends reasoning in total cost of ownership: sensors and robots, facility retrofits, integration with existing systems, maintenance and spare parts, downtime during commissioning. A warehouse automation project can require hundreds of thousands of dollars in AI development, but millions in physical infrastructure and site modifications.

A simple, honest ROI calculation

The core formula fits on one line: payback = total investment / annual net cash flow (savings minus recurring costs). All the value of the exercise lies in honest assumptions. The table below is a purely illustrative example: none of these values is market data, replace every line with your own numbers.

Two robustness tests are mandatory. First, the number of shifts: the same calculation on a single shift pushes payback beyond 5 years; utilisation is the first lever of robotics ROI. Second, the pessimistic scenario: 80 percent uptime, a 6-month ramp-up, one position saved instead of two. If the project only survives the optimistic scenario, it is not a project, it is a bet. For reference, integrators commonly advertise 18 to 36 month paybacks on mature, highly utilised applications; treat that range as a vendor-claimed order of magnitude, not a promise.

The 5 classic mistakes

1. Budgeting only the arm. Integration costs 2 to 3 times the robot; forgetting it produces the 50 to 80 percent overruns seen among buyers who discover the integration bill mid-project.
2. Picking the most spectacular use case rather than the most repetitive one. The right first pilot is boring: high volume, controlled variability, a metric you can measure today.
3. Underestimating data and MLOps. A learning-based system drifts as products, lighting or packaging change. Without monitoring, retraining and a manual fallback procedure, day 1 performance says nothing about month 6.
4. Deploying without an owner on the shop floor. A pilot carried only by management or the innovation team, with no named line supervisor and no trained, involved operators, stops at the first breakdown.
5. Computing ROI on the ideal scenario. One shift instead of two, compliance (CE, AI Act) left out of the budget, integration lead times ignored, and no written scale-up criteria: the recipe for the eternal pilot.

Decision checklist

1. The use case is repetitive, high volume, and its current performance is already being measured.
2. A single, quantified success metric is defined before signing (cost per unit, scrap rate, uptime).
3. The budget covers the full cost: integration at 2-3 times the robot price, MLOps, training, compliance.
4. ROI stays positive in the pessimistic scenario (one shift, 80 percent uptime, slow ramp-up).
5. A production owner is named and operators are involved from the pilot's design stage.
6. The compliance strategy (CE marking of the cell, AI Act, GDPR where relevant) is identified and costed.
7. Scale-up or stop criteria are written down and dated before launch.
8. The contract covers maintenance, software updates and exit terms, whether the model is a purchase or robots-as-a-service.

Keep reading

• Build the full decision case with our factory director's roadmap.
• Scope the first prototype before costing a pilot: our first prototype pillar.
• Compare the humanoid robots used in these deployments in our robots comparator.
• Check whether your pilot can be co-financed before costing its net ROI: our funding your project pillar.