Humanoid robots need a labor ledger, not just an uptime score

GXO and Agility announced further Digit deployments on October 26, 2026, planned for early 2027. Partnership announcement. The commercial question is not simply whether more machines can work. It is whether the labor needed around each machine becomes easier to organize as deployment grows.

This is an accounting question before it is a verdict on any product. A robot can be available while the surrounding process waits. A completed transfer can require preparation or recovery that appears in another team's schedule. Conversely, a slower machine can be valuable if it makes a shift more predictable. The following is OMN's analytical framework, not a claim about undisclosed GXO performance.

Decide what belongs inside the account

Agility's current FAQ says production operation is autonomous, while monitoring and maintenance involve Agility and operating partners. Manufacturer FAQ. Autonomous execution and absence of supporting labor are therefore different propositions. This distinction does not imply that a person secretly drives the robot. It means a task-level description cannot, by itself, answer a whole-process staffing question.

An operator evaluating deployment should compare accepted work with all labor required to deliver it: ordinary preparation, exception diagnosis, physical recovery, maintenance and coordination. Supplier-side effort matters even when it is bundled into a service fee rather than billed as customer labor. A service arrangement changes who pays and manages the work; it does not make the work disappear.

The boundary must also stay consistent. If a manual baseline includes handling awkward items but a robotic trial excludes them, the comparison measures a changed workload as well as automation. Excluding difficult cases can be a sensible deployment choice. The honest conclusion is then about the selected workload, not the entire job.

This also explains why procurement and floor supervisors might judge the same project differently: one sees a bundled service price, while the other absorbs preparation and coordination. Combining the accounts does not presume higher costs. It prevents departmental boundaries from standing in for production boundaries.

Recovery is a workflow, not a reset button

Consider a hypothetical tote that cannot proceed to its next station. Someone may need to identify the cause, decide whether the item can be moved, perform the authorized action and confirm that the downstream process has resumed. These are illustrative steps, not a reported incident. Timing only the final reset would omit the wait for an authorized person and the effort of reconstructing what happened.

A 2020 NIST interface paper links useful diagnostics with operators' ability to understand failures and respond. Research reference. Our practical inference is to measure the handoff as work: what information accompanies an alert, who owns the next action, and what establishes closure? Counting alerts without their duration and consequences can confuse an improved detector with a worsening operation.

There is also a distinction between time spent and time interrupted. A brief intervention can fragment a worker's main assignment; a longer scheduled maintenance task may be easier to absorb. Neither should be reduced to a simple intervention count. A good labor ledger records active support time and the interruption's operational effect separately, without double-counting the same minutes as both wages and lost output.

When a worker must remain available for support while holding another assignment that is hard to interrupt, the absence of a dedicated support role does not mean no capacity is reserved. Reviews should test whether that arrangement survives a busy shift rather than treat demonstration staffing as permanent. This is a question to investigate, not evidence of hidden understaffing at a particular warehouse.

The fastest movement need not produce the best shift

A 2022 NIST assembly study found a faster method needed more manual intervention under degradation than a slower, more robust alternative. Study. This was not a Digit warehouse trial. It supports a narrow methodological warning, not a transferable performance number.

For a warehouse decision, the relevant comparison is the completed shift under its actual task mix and support constraints. A faster local cycle may increase queues elsewhere. A more cautious cycle may reduce recoveries but still fail a dispatch deadline. Throughput, workload and service commitments must be evaluated together; none automatically substitutes for the others.

Scaling creates another possible trade-off. Shared support can spread expertise across sites, but simultaneous requests can also compete for the same specialist. That is a scenario to test, not evidence of a current bottleneck. A deployment review should ask whether service assumptions remain credible during overlapping exceptions, rather than infer staffing needs from a quiet demonstration.

What stronger evidence would look like

A useful comparison would preserve task definitions, quality requirements and support boundaries between the baseline and the robotic workflow. It would distinguish initial training from recurring work, show variation across shifts rather than only an average, and disclose which difficult cases remain outside the automated scope. Early learning costs should not be mistaken for permanent inefficiency, but neither should they silently vanish from an expansion plan.

The optimistic interpretation is that experience makes recovery routines simpler and shared expertise cheaper. The competing interpretation is that apparent labor savings partly move into preparation or support accounts. Both remain hypotheses until comparable operating evidence separates them.

An expansion announcement cannot resolve that dispute. The more useful next disclosure is not another isolated uptime percentage, but a consistent account of accepted work, supporting labor and delivery reliability. That would make it possible to distinguish a robot that performs a task from an operating model that genuinely releases capacity.