Demo, pilot, deployment: how to tell where a vendor actually is

You are probably not buying a humanoid robot. You are deciding whether a $20-a-month assistant can be trusted to draft the client email without you rereading every line [8], or whether the automation you built last month can run for a week while you are on holiday. The feeling that keeps recurring is the same one either way. The demo was excellent, and the thing you actually got needs watching. That is not a failure of judgment on your part. It is the predictable output of a market where a demo costs a weekend and a deployment costs years.

Humanoid robotics is the best place to watch that gap, because the evidence is unusually public. A robot company merging into a listed shell has to put its operating claims into documents filed with the SEC, signed by people who carry liability for them. One selling into warehouses has to meet safety standards written by committees that publish their drafts and their deadlines. No AI writing tool is obliged to tell you how many hours it has run inside a paying customer’s building; Agility Robotics, taking its Digit robot public through a business combination with Churchill Capital Corp XI at a pre-money equity value of $2.5 billion, is [1]. What follows reads that disclosure as a template. This is not a procurement manual for industrial robots. If you are actually buying one you have a safety engineer and a process, and this is not it. It is for the operator who wants a repeatable way to tell whether any vendor, hardware or software, has shipped or is still rehearsing.

The demo is the cheapest thing a vendor makes

A demonstration video is a claim about a best case, filmed as many times as needed, in a space arranged for filming. Nothing about that is dishonest. It is simply the cheapest artifact a company can produce, which is why it is the first one produced and the one produced in the greatest volume. What it tells you is that the capability existed once, under conditions the vendor chose. That is worth something. It is not worth what most buyers pay for it.

The expensive artifact is a number that accumulates. By June 2026 Agility had disclosed more than 65,000 hours of Digit operation across 9 customer facilities [1], and in November 2025 it had announced that Digit moved over 100,000 totes at GXO’s Flowery Branch facility [2]. You cannot film 65,000 hours. You cannot stage it, and you cannot borrow it from a partner for an afternoon. It either happened inside someone’s building or it did not, and because the figure appears in an exhibit filed with the SEC, somebody with liability has signed off on it [1]. That is the difference in kind you are looking for: not a better demo, but a number that had to survive contact with ordinary weeks.

Most software vendors publish no accumulating number at all, and you should notice that rather than resent it. A private company has no filing obligation and no reason to volunteer its operational scars. But the absence changes what you can conclude. When a vendor shows you a capability and nothing else, you have learned that the capability exists, and only that.

The four disclosures that mark a real deployment

The first is a named customer. Agility states that Digit is commercially deployed with Schaeffler, GXO, Toyota Motor Manufacturing Canada and Mercado Libre, automating repetitive physical tasks across manufacturing, distribution and logistics [1]. A named customer has agreed to be named, which means someone at that customer signed off, which means the thing works well enough that being publicly associated with it is not a risk. “A Fortune 500 logistics leader” is a customer who declined.

The second is a cumulative usage figure with a unit and a date attached. Hours, totes, facilities, all as of a stated moment [1][2]. A number without a unit is marketing. A number without a date is stale by construction, because the useful thing about accumulating numbers is the rate at which they accumulate.

The third is a named constraint. As of July 2026 Digit worked in a human-free space, kept apart from people on the floor, with version 5 expected to be unveiled in the autumn able to sense humans and drop the robot-only zone [6]. That is a vendor telling you exactly where its product stops. It is the most useful sentence in the whole story, and it is the sentence most software vendors will not say out loud.

The fourth is a standard, named, with its status. Agility has published specifics rather than adjectives: a category 1 stop that holds power to the actuators through deceleration, a safety PLC providing functions that meet performance level d, an on-robot emergency stop, a wireless teach pendant with an integrated emergency stop, and FailSafe over EtherCAT for safe communication between devices [3]. Those are checkable claims against published documents. “Enterprise-grade safety” is not.

All four translate to software without much effort. The named customer becomes a logo you can email; if the vendor cannot connect you to one operator of roughly your size, the reference does not exist. The cumulative figure becomes anything the vendor is willing to quantify and date, which for most AI products means a status page, an uptime history or a published changelog with real dates rather than a marketing timeline. The named constraint becomes the answer to what the tool is not permitted to do unattended, which every serious vendor has thought about and every unserious one deflects. And the standard becomes whatever the vendor claims to comply with, checked against the issuing body’s own page rather than the vendor’s summary of it. Three of those four cost you a single email each.

Cooperative safety is where the years actually go

The reason a working robot still lives inside a cage is rarely that it cannot do the work. It is that letting it near a person is a different engineering problem with a different evidence burden, and that burden is written down. ISO 10218-1:2025, published on 5 February 2025 as the third edition and superseding the withdrawn 2011 version, sets safety requirements for industrial robots addressed as partly completed machinery [4]. Meeting it is not a sprint.

The standard that would actually cover a walking robot is not finished. ISO/CD 25785-1, covering safety requirements for dynamically stable industrial mobile robots including bipedal, quadrupedal and wheeled balancing forms, was still a committee draft at stage 30.60, close of comment period, dated 8 July 2026 [5]. So the robots are working in customer facilities while the standard describing how to judge them is still in draft [1][5]. Agility co-founder Jonathan Hurst put the commercial stakes plainly in the merger announcement, saying the company believes cooperative safety is “the critical unlock for scaled humanoid adoption” [1].

Read that as the general case. The gap between “it can do the task” and “it can do the task next to a person, unattended, in a way somebody will insure” is where product timelines actually go. It is measured in standards committees and incident logs, not in model capability. The software version is identical in shape. Your AI tool can draft the invoice today. Whether it can send the invoice is a question about supervision, liability and reversibility, and it will be answered years after the drafting was solved.

Contracted is not delivered

Agility disclosed more than $300 million of multi-year orders for Digit v5, subject to the realization of certain contractual milestones, alongside a growing pipeline of over 30 customers [1]. Both numbers are real and both are forward-looking. An order subject to milestones is a promise to pay if a thing arrives and performs; it is not the same object as revenue, and the version under order in June was the version described in July as expected in the autumn [1][6].

None of that is a criticism of the company. It is a description of how order books work, and it is exactly the distinction most buyers collapse. The headline figure a vendor leads with is usually the largest defensible number available, which means it is usually the one furthest from cash. Your job is to walk it backwards: from booked orders to shipped units, from shipped units to units running unattended, from units running unattended to hours accumulated. Each step down is a smaller number and a stronger claim.

Software has the same ladder with different words. Announced becomes waitlisted, becomes in preview, becomes generally available, becomes generally available without a human approving each action. Vendors publish the top of that ladder and let you assume the bottom. The specific question that cuts through it is not “is this available” but “which rung is the thing I would receive on, today, on my plan”.

That last clause carries most of the weight. Capability demonstrated on a vendor’s internal build, on an enterprise tier, or in a region you are not in, is capability you cannot buy. The cheapest way to settle it is to stop reading the announcement and read the pricing page, because pricing pages are maintained by people who get complaints when they are wrong. If a feature is real on your plan it is usually named there, and if it is named only in a blog post it is usually a rung above you.

The same reading, applied to the software you actually buy

Pricing structure is a disclosure, and it is the one software vendors cannot avoid making. Zapier never charges a task to check for new data, and charges only when a Zap successfully completes an action [7]. That is a vendor staking its revenue on completion rather than on attempts, which is a statement of confidence you can read directly off the pricing page. Watch the overage terms too: past your plan’s task limit Zapier switches you to pay-per-task billing, charged at a higher per-task rate than your base subscription tasks [7], so the week the automation works better than expected is also the week it costs more than expected.

Seat pricing carries the opposite signal, and neither signal is bad. Claude sells Pro at $17 a month billed annually or $20 monthly, and a Team standard seat at $20 a month billed annually or $25 monthly [8]. Per-seat pricing tells you the vendor expects a person present. Per-outcome pricing tells you the vendor expects to be left alone. If you are buying a tool to remove yourself from a loop, and it is priced by the seat, you are buying against the grain of the product and you should expect to supply the supervision the price assumes.

The last move is to price your own attention, because that is where the demo-to-deployment gap actually shows up on your books. A tool that costs $20 a month and needs 5 minutes of checking per output is not a $20 tool. Run the arithmetic before you renew, not after.

What still goes wrong

The disclosure asymmetry cuts both ways. Agility published its operating hours in an exhibit filed with the SEC as part of a merger with a listed shell [1]; a private software vendor with a genuinely excellent deployment record faces no equivalent filing and may publish nothing. Silence is therefore weak evidence, not proof. What you are reading with this method is a vendor’s willingness to be held to a number, which correlates with maturity but does not equal it.

Cumulative figures also hide their own distribution. More than 65,000 hours across 9 facilities [1] tells you nothing about whether one site carries most of the total, how often a technician intervened, or what the failure modes were. Intervention rate is the number you actually want from any autonomous system, robot or agent, and almost nobody publishes it. When you can ask a vendor directly, ask for that instead of uptime.

And the ladder moves underneath you. Standards arrive after the products they govern, as the still-draft status of ISO/CD 25785-1 shows while robots are already working in customer facilities [1][5]. A tool that needed watching in March may not need it in September, and a claim that was true at purchase may be conservative by renewal. This reading is worth about 20 minutes and it expires. Redo it when the invoice arrives, not when the announcement does.