On 26 October 2026 GXO and Agility Robotics announced that Digit, Agility’s humanoid, will go to two more GXO warehouses — Statesville, North Carolina and Lebanon, Tennessee — with deployments expected to begin in early 2027 on Digit 4 and migrate to Digit 5 once it is commercially available. The evidence offered is the first site, Flowery Branch, Georgia: a proof of concept in late 2023, a multi-year Robots-as-a-Service (RaaS) agreement in June 2024, and more than 100,000 totes moved onto conveyors alongside existing automation. Agility adds more than 65,000 hours of real-world operation across all of its customers, three unnamed manufacturers in paid programs, more than 30 prospects, and more than $300 million in committed multi-year Digit 5 orders that it says are not current-period revenue. All of this lands weeks before Agility expects to close its SPAC merger and list as AGLT.

The same customer’s CEO, Patrick Kelleher, said on GXO’s 5 August earnings call: “We have done 45 pilots on humanoids so far. We have not achieved ROI on humanoids yet. I think we are a couple of years away from that.” He also said that of the roughly 20,000 robots GXO expected to deploy this year, “in production, zero will be humanoids.”

Both statements can be true at once. That is the interesting part.

What I concluded

The expansion proves that a sophisticated buyer will keep paying for an option on humanoids. It does not prove that a humanoid pays for itself on a tote-moving job — and nothing in the public record lets an outsider check.

Agility’s own chief business officer, Daniel Diez, gave the right yardstick at the company’s 6 October investor day: customers decide on “five critical factors. Safety, accuracy, uptime, throughput, and ROI … these are the metrics that you should be paying attention to.” Score the public record against that list. Accuracy: published — about 98% at GXO and Schaeffler. Safety: described qualitatively; Digit 4 has worked inside fenced cells. Uptime: not published per site; the closest number is Digit 4’s two-to-one run-to-charge ratio. Throughput: not published per robot-hour. ROI: the customer says not yet.

What is published instead are cumulative totals — totes, hours, pounds, order value. Those are numerators. A business case needs the denominators: how many robots, over how many scheduled hours, at what rate against a human baseline, and at what cost per tote. The product lesson is not “be sceptical of humanoids”. It is that the acceptance gate has to be written before the robot arrives, and for a company asking public investors to fund the scale-up, the gate should be public.

Constraints

Four constraints explain why the disclosure looks like this. None of them requires bad faith.

Customer confidentiality. Agility says the KPIs — throughput, accuracy, uptime — are “mutually agreed upon” with each customer and “stipulated in the agreement”. Three of its new manufacturing customers chose not to be named. Per-site rates are commercially sensitive for the 3PL as well as the vendor.

Throughput depends on layout. Diez’s own example: if a customer asks Digit to walk through a maze before the task, “throughput will be a little bit different than if it’s just walking in a straight line.” A per-robot-hour number from one building does not transfer cleanly to another, which is a real reason not to headline it — and also the reason a single cumulative total says even less.

Battery duty cycle. Jonathan Hurst, Agility’s co-founder, said Digit 4 was “gated by a two-to-one run-to-charge ratio” — about an hour on the charger for every two hours of work — and that Digit 5 targets roughly 90 minutes of run time with an under-nine-minute charge, or more than 20 productive hours a day. The new GXO sites start on Digit 4.

The vendor’s economics assume the next robot. CFO Michael Beer’s illustration prices RaaS at about $100,000 per robot per year and compares it with a $30-an-hour fully burdened rate that “would equate to roughly $180,000 per year”, leaning on three-shift operation. My arithmetic: $180,000 at $30 an hour is 6,000 labour-hours a year, about 16.4 hours every day. A Digit 4 capped at two-thirds of the clock by charging tops out near 5,840 hours a year before any other downtime. The illustration is a Digit 5 argument, and Digit 5 general availability is targeted for the end of 2027.

The options, and what I traded away

The decision here is the one a product manager on either side of this contract has to make: what evidence counts as “it works”, and who gets to see it.

Evidence

The numerators. More than 100,000 totes at GXO’s Georgia site, at 98% accuracy on Spanx orders, per Diez. More than 65,000 hours across all customers and programs. Well over a million pounds of automotive-bearing totes at Schaeffler’s Cheraw site. $300 million-plus of Digit 5 orders, which Agility described at its investor day as 1,000 robots on three-year RaaS from one of its most experienced Digit 4 customers — by my arithmetic $100,000 per robot per year, consistent with Beer’s illustration.

What they do not tell you. The robot count at Flowery Branch is not in either release, so totes per robot-hour cannot be computed. The time window is loose: the GXO release says the 100,000 totes came after the 2024 agreement. If the count starts in June 2024, that is roughly 3,600 totes a month, about 115–120 a day for the whole site, on an unknown number of robots — my arithmetic, and not a productivity claim. The 65,000 hours cover every customer; at the illustration’s 6,000 hours a year, that is about eleven robot-years of three-shift work across more than two years of deployments. For the new sites, GXO did not say what it is investing or when it expects ROI, and AGCO’s 2027 pilot disclosed no robot count, investment or benchmarks.

The buyer’s own read. GXO has run 45 humanoid pilots and has also tested Apptronik’s Apollo and Reflex Robotics’ robot. Kelleher named hand dexterity and task speed as the biggest remaining gaps, while saying established warehouse robotics already earns ROI. This is a customer that measures, and it has said publicly where the line is.

The pattern is not unique to humanoids. In the same week Teradyne and Sereact said that Universal Robots arms running Sereact’s VLA model are processing Zalando returns at CEVA sites in Germany and Poland, with “hundreds” of arms shipped — and no disclosed processing rates or returns. The one announcement in the week that stated its bar in advance was a small Korean delivery pilot.

Counter-arguments worth taking seriously. “A repeat customer is the strongest signal there is” — it is the strongest signal of willingness to keep testing; Kelleher’s own words put the payback a couple of years out, and under RaaS the vendor carries the robot’s bill of materials, which makes renewal cheap for the buyer. “Confidentiality makes per-site numbers impossible” — Agility already publishes one aggregated figure, about 98% accuracy across Schaeffler and GXO, so aggregation without naming terms is clearly possible. “Throughput is layout-specific, so any number misleads” — then publish it with the layout class; a range beats a total. “The Digit 5 economics are what matter” — agreed, which is why Digit 4 totals should not be used to sell them.

Where this stops being true

The argument holds for humanoids sold on labour-substitution economics into structured material handling — totes, bins, conveyors — by a vendor raising public money on that pitch. It weakens when the buyer is explicitly buying learning rather than payback (a 3PL with more than 1,000 facilities can afford a few experimental sites and say so); when per-site KPIs are contractually visible to the only party that bears the risk; for fixed-arm cells where ROI is already established and not in dispute; and once Digit 5 ships, because the charging constraint that drives my duty-cycle arithmetic changes by design. What I am not claiming: that Digit does not work, that 98% accuracy is unimpressive, or that GXO made a bad decision. I am claiming that none of the published numbers can answer the ROI question their own vendor says matters most.

What I would do differently

Four things, all checkable.

Publish one rate per site, not one total per company: robots deployed, scheduled hours, and units per robot-hour against the human baseline for the same move.

Publish uptime the way the vendor defined it — including charging time — so the Digit 4 versus Digit 5 duty-cycle claim can be tested when Digit 5 arrives.

State the pilot bar before the pilot. If AGCO’s 2027 start is worth a press release, so is the throughput and cost-per-unit threshold that would turn it into an order.

As a buyer, write the stop rule into the RaaS contract: the cost per tote at which you expand, the cost at which you stop, and the date you decide.

My reading is that GXO is doing exactly what a careful operator should — renting an option on a technology it expects to pay off in about two years — and that the industry’s disclosure is lagging its fundraising. A cumulative total only ever goes up, so it can never show you a failure. Safety, accuracy, uptime, throughput and ROI are the right five metrics. Until the public record has the denominators for at least three of them, 100,000 totes is a milestone, not a business case.