Mitos Robotics
← Writing13 August 2026

Selling Robots with a Deadline

The irresistible opportunity

Every robotics pitch deck opens with the same slide: global labour is worth tens of trillions a year and most of it could be done by robots. The number is real, but it measures what the work costs today, not what anyone will pay you tomorrow. Nobody buys labour for its own sake; they buy a problem being dealt with, and hiring someone is just the arrangement they landed on, usually because it was the best option at the time.

Which means you are not replacing a wage but competing with every other way that problem could go away: a machine, a change to the process, a supplier who delivers the thing pre-sorted, or simply living with it. And the person you are displacing was never doing only one task: she also spotted the damaged pallet, told someone the line was about to jam, and covered for the machine that broke down.

Overcoming these challenges is not unique to robotics. In the startup world, the received wisdom is to solve a specific, painful, expensive problem, so founders pick a task: folding towels, unloading a trailer, cooking a meal.

I worry the advice is poor in robotics, because anything with volume, repetition and a stable specification already belongs to a purpose-built machine, and what is poorly specified and varied is done by people. That leaves the robot in an awkward position.

So the robot arrives to find itself squeezed from both sides: the purpose-built machine below it takes everything stable and repetitive and wins on speed, cost, footprint, reliability and serviceability, and the person above it is cheapest on everything variable, unstructured, low-volume or short-run. What is left in the middle is the residual: too varied to justify tooling, too small to justify installation, and priced against a worker who needs no capital commitment, no service contract and no floor space beyond where they stand.

The difference that decides it is which costs repeat: a person costs you money every hour they work, whereas a machine costs a great deal once and then relatively little, because installation, supervision, floor space, downtime and the capital itself are all paid up front and spread across everything the machine ever does. Run it hard for ten years and those five costs diminish greatly per unit while the wage bill keeps climbing, which is why the repetitive work was automated first and why it is not coming back.

Now take what's left: short runs, where those same costs are spread across a few hundred units and the wage bill is small because a short run does not take many hours. A robot that matched human dexterity tomorrow would still lose these tasks on those five costs, almost none of which a person incurs.

Labour scarcity widens the gap: where workers cannot be hired, the wage the robot competes against rises. But a higher wage raises the return on every substitute at once, the purpose-built machine and the process change included, and the work that stays with people mostly stays because of what it demands rather than what it costs.

That is the squeeze, and it is the nature of the market rather than a stage in its development. The gap it leaves in the middle is a great deal narrower than the trillions on the slide.

In theory the robot fills that gap

The formal definition of a robot is perhaps useful here: it is a reprogrammable, multipurpose manipulator. A machine's purpose was fixed by a designer years before it shipped and it will do that one thing until it is scrapped. A robot turns up able to be told what to do.

In theory then it is the natural candidate for work that keeps changing. In practice two things get in the way:

The first is that installation is expensive enough to decide who gets to automate at all. The US Census has numbers on this: manufacturing plants with fewer than twenty employees use robots at about three percent, while plants with more than two hundred and fifty run at nearly thirty. The gap survives controlling for industry, so it isn't that the big plants are doing different work. The small ones cannot carry the cost of the project.

The same study turned up something blunter: a region with at least one robot integrator in it is twenty to twenty-five percentage points more likely to be a robot hub than a region with none, and the effect is not a proxy for a skilled workforce: add education and STEM employment to the regression and it is those that drop out, not the integrator. Whether a factory has robots depends less on the factory than on whether there is somebody nearby who can install them.

The second problem is that the flexibility barely gets used. An arm is bought flexible and then programmed to perform a single trajectory it will repeat for a decade, a career trajectory many of us can relate to. The generality is killed as soon as it is installed. Forty years of industrial robotics exploited our ability to specialise a general machine, not a general machine that can specialise itself.

The bet the money is making

There has been a recent revival of interest in robotics because of breakthroughs in LLMs. Robotics startups raised more in the first half of this year than in the whole of any previous year. The surge is concentrated in embodied AI, and the money has gone in two directions.

The first is the model: well-funded labs building the general system, on the view that this is a research problem and it will break the way language did. Get the thing working and the rest follows, because a machine you can instruct in plain English does not need an integrator. That is a coherent position and it is why those companies are shaped like labs.

The second is hardware and data, from a camp that thinks the models are only half of it, and that the missing half is the millions of hours of real robots doing real work that nobody has collected yet. So you build the machine, get it onto sites, and let the data accumulate until the model catches up. They deploy far more than the labs do, which makes them look like the commercial ones.

But the deployments are hardware and pilots going out below cost, because the vendor is buying data rather than margin. That is not a criticism, it is the strategy working as intended, and it does put robots in buildings. It tells you nothing about whether anyone would buy the thing at a price that sustains a company, and the buyer who took a subsidised pilot has often not had the conversation that decides it.

So both camps are deferring the same question: solve the model, with or without a data-gathering detour first, and selling gets easier. Distribution is a later problem either way.

Deferring it is reasonable; the question is how long the deferral runs. Software gives the first measurement: language models spread very fast on the consumer side, because one thing served everyone, copying it cost nothing, and the buyer was a person with a browser. Inside companies it has gone slowly, and not because the models are bad. Someone has to find the work worth doing, change the process around it, sort out who owns the data, get legal to sign off, and train people to use it. Most firms bring in outside help for that, and a consulting industry has grown up in three years to bridge the gap. That is not what you would expect from a technology with no installation cost.

Robotics has all of that plus the parts software does not have. The machine has to be bought, which means a capex budget; it has to fit in the building, satisfy an insurer that it is safe around people, and be fixed when it breaks by someone who can get there. And the setup costs four to six times the machine, at every site, every time. If the software version is still taking years with no hardware involved, it is hard to see why the version with steel in it goes faster.

Enterprise software hit the same wall twenty years ago, with almost exactly the same cost ratio. A study comparing on-premise and hosted systems found that the software itself came to 17% of the total cost of an on-premise deployment, with the rest going on infrastructure, implementation, support and training, and it is still true of the big systems: implementation runs to between half and seventy percent of first-year ERP spend, and the licence to somewhere between a fifth and a third. Firms that budget from the software quote find this out during the project.

What fixed it was not better software but the same product on a different balance sheet. SaaS moved the setup, infrastructure and upgrade costs onto the supplier as a cost of doing business, so the buyer stopped needing a capital committee and a market that had been restricted to firms large enough to run an implementation project opened up to everyone else. It took a decade and it worked.

Industrial IoT is the one that did not, and it happened in factories within the last ten years. The pitch was familiar: sensors on everything, data off the floor, savings to follow. What happened was that 84% of companies got stuck in pilot for more than a year and 28% for more than two, held up by thin resources, missing expertise and a business case nobody could make stand up at scale. It did not improve with time: by 2020, 74% of manufacturers said they had failed to scale some or many of their use cases, eighteen points worse than the same survey a year earlier.

The sensors delivered data but not a case: the savings at any one site rarely covered the integration, expertise and process change needed to collect them, so the projects died between the pilot and the second site, a distance measured in budget approvals rather than technical progress. Somebody has already written down what that does to the vendors: pilot purgatory slows large companies down and kills the startups selling to them. The polite name for the resulting activity is industrial tourism.

The obvious lesson is to do what software did, and the industry noticed: Seegrid put its warehouse robots on a subscription in 2021, and Locus Robotics was selling the same model before that, with a setup fee and a contract bundling support, maintenance and upgrades. It is the right answer for the plant that cannot carry a capital project, and the awkward part is who is shipping it: vendors with service teams and an installed base, none of whom had to solve general-purpose manipulation to get there.

The money also picks the target. A fund writing a nine-figure cheque needs a market worth trillions, so the pitch has to be general labour rather than a job shop running four hundred part numbers. But the work that survives the squeeze is small, specific and dull, and a company built to do it does not raise at those numbers. Which is not only a constraint on the companies that took the money; it sets the price of entry for everyone, because the alternative, a company charging what the work is worth and growing on what customers pay, is bidding for the same buyers against vendors who do not need the money.

The road and the money

None of which means either bet is wrong; it means only one shape of robotics company gets funded right now, and that company needs several years of money before it needs a customer. Which is fine, as long as the money lasts as long as the problem does.

Self-driving ran this bet once before, all the way to the end, with both camps at the table: it had its lab and its fleet. Waymo bet the system would come together first, Tesla bet the missing half was data from cars already on the road, and the predictions from the middle of the last decade gave both of them two or three years.

What actually happened took most of two decades and north of a hundred billion dollars across the sector. Uber sold its self-driving unit in 2020 after a fatality and years of spending. Argo shut down in 2022 with several billion of Ford and Volkswagen's money in it, not because the technology stalled but because its owners stopped believing what remained was shorter than their patience. Cruise took more than ten billion from General Motors and was wound down anyway. The company that made it through started in 2009, ran eleven years before its first paid driverless ride, and could afford to because its parent prints money. The fleet launched its first small robotaxi pilot last year with a safety monitor in the passenger seat. Capability arrived and the rollout still goes city by city, with depots, remote operators, mapping and a regulator to arrange in each one. Distribution did not get easier when the model got good; it became the next problem.

And self-driving was the friendly version of the bet in every respect but one: a mistake on the road kills someone, and a real share of those two decades went on buying that risk down to a standard a regulator would accept. Manipulation in a plant escapes most of that bill. But the safety case is also what let the vendor own the cars and skip selling into anyone else's building, so the plant robot escapes one burden and inherits the other: thousands of buyers, each with a building, an insurer and an integration to arrange, paid retail every time. Self-driving had one task, one embodiment and a product every buyer already understood. General work in a plant has none of that, and the fund holding the position has ten years.

If the money runs out first, the failure probably won't be the technology stalling but a few hundred plants running a pilot, watching it end when the vendor runs out of road, and writing the category off. Those plants are not neutral afterwards; they have a specific answer ready for the next three vendors who call. I saw the exact same thing happen when selling software into last mile logistics providers.

Deploying too early, the move fast answer, will not solve the problem. Below a certain level of capability a pilot tells you nothing, because the customer has not been asked to depend on it, but it still uses up a building: there are only so many, each says yes to a category like this roughly once, and one that has watched a robot nearly work is much harder to get back into than one that has never seen it. The sensors worked, and it made no difference.

So the industry gets one pass at each building and about ten years of patience, and it is currently spending both on demonstrations. The work that decides success is in making the model better and in finding the equivalent of software's move before the people paying for the research stop waiting.

References

The comparative-cost argument

Adoption tracks establishment size

Integration and cell economics

Enterprise software cost structure

  • Hurwitz & Associates (2010). The TCO Advantages of SaaS-Based Budgeting, Forecasting, and Reporting. Software costs at 17% of total on-premise solution cost. Summary via GlobeNewswire. globenewswire.com.
  • ERP Research. ERP Implementation Cost Breakdown. Implementation at 50 to 70 percent of first-year ERP spend, licence at 20 to 30 percent. erpresearch.com/en-us/erp-implementation-cost-breakdown.

Industrial IoT pilot purgatory

  • McKinsey & Company (2018), with the World Economic Forum. IoT survey: 84% of companies stuck in pilot for more than a year, 28% for more than two. mckinsey.com.
  • McKinsey & Company. Industry 4.0 global survey of 400 manufacturers, late 2020: 74% had not yet scaled some or many use cases, eighteen points worse than the 2019 survey. Reported via IndustryWeek. industryweek.com.
  • Vause, C. (2020). What is "Pilot Purgatory" and Why Every Start-Up Should Know This Term. Source of the pilot-purgatory-kills-startups observation and the phrase "industrial tourism". carl-vause.medium.com.

Robots as a service

  • Seegrid Corporation (June 2021). Addition of a Robots as a Service subscription model alongside purchase and leasing options. businesswire.com.
  • FreightWaves (2022). Seegrid RaaS launch, with Locus Robotics on setup fees and one- to three-year contracts bundling support, maintenance and upgrades. freightwaves.com.

Self-driving

  • Chafkin, M. (2022). Even After $100 Billion, Self-Driving Cars Are Going Nowhere. Bloomberg Businessweek, October 2022. Source for the sector investment total. bloomberg.com.
  • CNN Business (December 2020). Uber sells its self-driving division to Aurora after a five-year effort marred by litigation and a fatal crash. cnn.com/2020/12/07/cars/uber-sells-self-driving.
  • TechCrunch (October 2022). Argo AI shuts down; Ford records a $2.7 billion impairment, Volkswagen had invested $2.6 billion in 2019. techcrunch.com.
  • CNBC (December 2024). GM exits the robotaxi market after spending more than $10 billion on Cruise since 2016. cnbc.com.
  • Waymo (October 2020). Fully driverless service opened to the public in Phoenix, eleven years after the project began as Google's self-driving car effort in 2009. waymo.com/blog/2020/10.
  • Wikipedia. Tesla Robotaxi: limited launch in Austin, June 2025, with a safety monitor in the front passenger seat. en.wikipedia.org/wiki/Tesla_Robotaxi.

Capital

  • Crunchbase data on robotics venture funding: $18.8 billion raised globally in the first half of 2026, exceeding the $15 billion full-year 2025 total and the $14.1 billion 2021 peak, concentrated in embodied AI. Via Value Add Pulse. valueaddvc.com/pulse.
Selling Robots with a Deadline · Mitos Robotics