06-reference/research

service as software physical verticals

2026-09-01·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)

Four real fits, three boundary cases, four mislabels: what Service-as-a-Software actually looks like when the work is made of atoms

The question

"What are 5-10 documented examples of Service-as-a-Software (per Rico/Foundation Capital framing) being successfully applied to PHYSICAL verticals — not consumer commerce (UCP/Walmart/Tatcha) and not pure knowledge work (Harvey/Sierra/Decagon) — but instrumentation, robotics, manufacturing, or industrial processes?"

Context: the founder's 2026-05-03 mission reframe recast Service-as-a-Software as "agents replace coordinators in physical-world ops." The vault has the framing but no physical-vertical evidence base under it. This brief supplies one, and pushes back on the framing.

What we already know (from the vault)

What the web says

I scored every candidate against four tests: (1) buyer purchases work output, not a tool or asset; (2) price is denominated in the unit of work; (3) the money comes from the labor/services line, not the software or capex line; (4) the vendor owns execution, so the customer's operator headcount actually goes away.

Passes all four — real fits (4):

Boundary cases — labor-benchmarked subscriptions, not outcome pricing (3):

Mislabels — vertical AI or defense hardware wearing the label (4):

One category-level claim worth flagging as weakly sourced: several 2026 RaaS pricing guides quote pay-per-pick at $0.03-0.06 per item and pay-per-delivery at $0.50-2.00. The category is real (Locus and inVia have priced per-pick for years), but those specific figures come from SEO-grade aggregator pages with the texture of generated affiliate content. Do not put those numbers in front of a client.

Convergences and contradictions

Synthesis for RDCO

The single discriminating variable across all eleven candidates is not AI capability, sensor cost, or autonomy level. It is whether the vendor took the labor onto its own balance sheet. Gecko sends its own crews. Path runs its own welding cells. Built operates a fleet. Aurora is the driver. Every mislabel (Skydio, Saronic) and every boundary case (Formic, Chef, Deere) leaves the asset and the operator with the customer, and correspondingly leaves the revenue in the software or capex line. That is a cleaner, more falsifiable test than the five diagnostic signals for physical verticals, and it is the thing to add to the per-bet evaluator: whose payroll does the work sit on after the sale?

This has a direct, uncomfortable consequence for opportunity #2 in [[2026-05-03-opportunity-map]]. That bet was specified as $400-1,500/mo per monitored asset. On the test above, per-monitored-asset-per-month is the Formic/Chef shape: a labor-benchmarked subscription on customer-owned equipment. It is a real and financeable business, but it is not service-as-a-software and it will not earn service-as-a-software multiples. The Gecko shape for the same vertical would be "we own your reliability outcome" — we send the crew, we run the inspection cadence, we generate and close the work orders, and the SMB manufacturer's maintenance planner role goes away. That is a harder business with atoms-work and liability the map already flagged. It is also the one with the $1.25B comparable at the top of it. The founder should make that choice deliberately rather than inherit the subscription framing by default.

Second implication, on pricing: the map's economics for #2 assumed compression against a $5-20k/mo incumbent integrator retainer down to $4-15k/mo. Against the physical-vertical evidence that is correctly calibrated — near parity, modest discount — and the map should be read as already having got this right, while the Rico note's 1:10 claim should carry a knowledge-work-only caveat wherever it appears in RDCO material. Anyone pitching a physical-vertical bet on 90% price compression is importing a knowledge-work assumption.

Third, on competitive position: Gecko occupies the top of the instrumentation vertical with a $71M Navy contract and its own field organization. It does not serve the 200,000+ US SMB manufacturers running reactive maintenance, because sending crews does not scale down to a $500/mo account. That gap is real and is the honest version of the #2 thesis. But the lesson to carry from Gecko is not "sell sensors cheaper." It is that the margin lives in the engineering-judgment layer sold alongside the robotic data collection, and the data collection is the wedge that earns the right to sell it.

Why this is in the vault

This is the evidence base under opportunity #2 of [[2026-05-03-opportunity-map]] (retrofit predictive-maintenance for SMB industrial), and it changes a specific unmade decision inside it: whether to price per-monitored-asset-per-month (Formic shape, subscription multiples) or per-reliability-outcome with RDCO owning execution (Gecko shape, services-absorption multiples). It also supplies the original re-frame a Sanity Check piece on this thesis would need in order to clear the no-derivative rule: in physical verticals the agent eats the crew, not the coordinator, and coordination gets repriced upward on the vendor's side.

Open follow-ups

Related

Sources

Vault

Web