Abroad, physical AI arrives as robots. In South Africa, the part that pays for itself today is the software underneath them, and it runs on hardware most plants already own.
Somewhere in a factory near Durban, a small robot is doing the stocktake, and nobody is watching it. This is, I will argue, the most important detail in the whole story of robots and South Africa, and I will get to it. But first you have to sit through the videos, because everyone else has.
You know the ones. A warehouse in Georgia where humanoid machines lope between shelves carrying plastic bins. A production line in Shenzhen where the arms never pause for lunch. They arrive on your phone with a swell of music, and they leave you with a feeling somewhere between wonder and dread. What they do not leave you with is an accurate picture of where this technology is useful, or where it is going to be useful in a country where labour is cheap and capital is not.
Take the arm out of the picture. What remains is less cinematic and more valuable: software that looks through the cameras a plant already owns, works out what it is seeing, and tells someone. The man on the floor without his hard hat. The line that stopped eleven minutes ago. Bay four, which has been empty since the morning shift and which nobody has thought about since. That software is being written now, some of it in South Africa, and it does not need a single new machine to run. The machines will come when the sums allow. In the meantime, the software is the whole show.
Three layers, not one
The phrase "physical AI" does a lot of damage, because it makes people think of a body. A more useful mental picture is a stack of three layers, like a sandwich nobody would want to eat.
The bottom layer senses, and it is much broader than cameras. A sensor in a skip that knows how full it is. A small Bluetooth tag on a trolley that pings when it passes a doorway. A tracker that says exactly where a vehicle or a pallet is standing right now. Weighbridges, the scanner at the gate, the controllers wired into your machines. People call this the internet of things, which is a grand name for something simple: ordinary equipment that can report on itself. Most plants have far more of it than they realise, and very little of it is watched by a human being for more than a few minutes a day.
The middle layer understands. This is software that reads what the sensors produce and says, in ordinary words, what is happening: that pallet is in the wrong place; that person should not be in that zone; that line has been quiet for longer than a changeover takes.
The top layer acts. A machine that moves, grips, sorts or drives.
Nearly all the attention goes to the top layer, which is also the least finished. Amazon's fleet passed a million robots last year, and almost none of them resemble a person; they are fixed, single-purpose machines that shuffle shelves. The walking kind exist, and they do jobs, but narrow and supervised ones. The most concrete public figure is one model clearing a hundred thousand bin moves inside one warehouse in Georgia. In the same year, the two best-funded companies in the field each stumbled in public: one missed its own production target by more than a year, the other cancelled a new warehouse robot four months after unveiling it. The people who have spent careers on this will tell you the hard part is no longer the motors. It is that a robot has nothing like the internet's worth of examples that taught language models to write, a gap the Berkeley roboticist Ken Goldberg has put at about 100,000 years of human reading time.
Sean Kelly, the chief technology officer of Dragonfly, the platform I will come to shortly, puts it in a way a plant manager might recognise. "Less sci-fi than people think," he said. "The next few years are mostly about plants finally being able to read themselves. What's currently underestimated is how much industrial data is sitting right there and is completely unreadable. Not missing. Just in forty different formats with nobody to translate."
The money, for what it is worth, is real. Venture investment in American hardware start-ups is on course for a record this year, according to Silicon Valley Bank. The same report finds that fewer than half of large American warehouses use AI in an ordinary working day. The layer that is getting better, and paying for itself, is the one in the middle.
What the middle layer looks like in practice
A disclosure before we go further: I have spent the past year working, as a developer, on the Dragonfly platform, which lives in exactly that middle layer. It is used, the company says, by Fortune 500 companies and the largest consulting firms in the world, and it runs in both the United States and South Africa. Here is the short version, without the jargon.
A working site produces a flood of information it never reads. Camera feeds, yes, but also the fill level in a skip, a tag pinging as a trolley leaves the yard, a tracker following a load across town, the controllers on the machines. In the office, the systems that hold orders, stock, staff and maintenance. All of it lives in separate rooms, and most of it is consulted only after something has gone wrong, which is the one moment it is no longer useful. Dragonfly connects to what is already there, pulls it into a single shared picture of the operation, and sets AI agents to read that picture. They notice what matters and tell the right person. Or, within strict limits, they act. The result is that a supervisor who has never used the words "internet of things" in their life ends up with the whole operation on one screen. Mr. Kelly's version, in one sentence: "A camera sees something, the platform turns it into a record an agent has full context over so it can reason, recommend, act."
The agent proposes, a person decides, and the decision is logged. Illustrative scenario, not a real site.
Belt 3 stopped eleven minutes ago. Nobody has logged a reason, and the camera that watches it is the only thing that has noticed.
The strict limits are the part I would want a manufacturer to remember if they remember nothing else. Any AI can be allowed to think about the site. There is only one safe path to act, and it is checked, limited and recorded. By default it does not act at all. It recommends. The agent proposes, a person decides, and the decision goes in the log either way. The alarms are wired so that they keep working if the AI is switched off. The device on the wall can see everything and change nothing. That caution is the reason you can put the thing in a plant where a wrong move costs money or sends someone to hospital.
Dragonfly does not make robots. It makes software that sees and understands. So why build the safe path to act now, years before there is a robot arm on the other end of it? Mr. Kelly answered without much pause. "Because you can't bolt it on later. If we let models talk to machines first and then try to add safety, it becomes a patch instead of the gate."
Why the software comes first here
Now the South African part, which is mostly arithmetic and can be done on a napkin.
A worker on an American recycling line earns a median of $22 an hour, according to the Bureau of Labor Statistics. South Africa's national minimum wage is R30.23, about $1.87 at this month's exchange rate. The best public estimate for the parts in a general-purpose humanoid robot is around $55,000, from a component teardown whose method has not been independently checked, so treat it as a rough figure. In America, that is about a year of one worker's pay. Here, it is more than a decade. You can see the consequence in the count: South Africa has about 12 industrial robots for every 10,000 manufacturing workers. The global average, according to the International Federation of Robotics, is 162.
US median $22.00/hour (BLS) against South Africa's R30.23/hour minimum wage, at R16.17 to the dollar.
But look where that arithmetic actually lands. Only on the top layer. A camera costs the same in Pinetown as in Pittsburgh, and more often than not it is already screwed to the wall. The intelligence is software, and software has never asked what the minimum wage is. A South African operator can have the first two layers today, for a fraction of what the third would cost, and get most of the benefit: knowing what is happening, and what to do about it.
The proof on a factory floor in Pinetown
Which brings us back to the small robot near Durban.
Morgancoat is Southern Africa's largest manufacturer of self-adhesive labelstock, with three facilities in KwaZulu-Natal. A full stocktake there used to take four to six hours, before anyone began to chase the numbers that did not add up. It now takes five to ten minutes to count more than 1,500 pallet positions.
The visible part of that is a robot named Pegasus. It does a circuit of the factory floor, on its own or when asked, carrying an RFID reader that picks up the chips built into the labels on every pallet from a distance. It was not bought from a catalogue. It was built by ProtoGarden Robotics, Morgancoat's own in-house robotics and software company, championed by Tom Van Den Bon, the firm's lead software developer since 2019. Hiring him, says Marc Black, Morgancoat's managing director, was "one of the best technology decisions we made." A label manufacturer in Pinetown, in other words, has its own robotics shop.
And the robot is still the least interesting thing in the building.
Depending where in each range the count lands. The same work, on the same floor.
before anyone chased the discrepancies
with the same day's discrepancies flagged
What turned hours into minutes was the software behind it: a system that tells the team where any batch is without anyone walking the warehouse, and that flags a discrepancy the same day rather than weeks later. It came from the same in-house team, and it was written around the way this particular business actually runs. Marc Black is blunt about why. "Most ERP systems are fundamentally accounting systems with fairly rigid stock and manufacturing functionality bolted onto them." His own was "designed around exception reporting rather than data dumping. The system tries to tell us what is wrong or requires attention, rather than simply showing us everything that is happening."
That is the same core idea as Dragonfly, arrived at independently by a company that makes labels. And for anyone who has been worrying about jobs since the second paragraph: "Since we started seriously down this road in 2020, our headcount has actually increased by over 20." Automation at Morgancoat, Mr. Black said, "hasn't been about replacing people, but about increasing the amount of useful work each person can do."
Custom by nature
Every plant is different, so the software has to fit the plant rather than the reverse. Morgancoat did it the hard way, by building its own robotics and software team under its own roof, which most plants cannot do.
What makes a platform like Dragonfly custom is something more prosaic: it connects to whatever a site already runs. The SCADA system on the floor. The cameras. The ERP and the CRM in the office. And the places where people actually talk, Teams or Telegram, so that an alert turns up in front of a human being rather than on a dashboard that nobody has opened since it was installed. Nothing gets replaced. Everything ends up in one place, running on the plant's own data, with agents reading it.
Mr. Kelly has a test for whether a platform deserves the name, and it is worth borrowing whoever you buy from. "The most important test is: is the second site cheaper than the first?" Part of Dragonfly is built from South Africa; riivo builds the screens its partners and customers use and is helping to build the safe path to act. So this kind of software is already being made in this country, for plants that are running now.
Mr. Black's advice to a fellow manufacturer is the sensible place to stop. "Start somewhere. Find a genuine pain point, then find or build a system that solves it properly. You don't need to automate the entire factory on day one."
The machines that move will arrive when the arithmetic changes. The software that sees and understands is here already, quietly doing the stocktake while nobody watches. When the machines finally show up, it will be that software telling them where to go.
FAQs
What is "physical AI", in one sentence?
Software that takes in what sensors and cameras see in the real world, works out what is happening, and either tells a person or, within set limits, acts on it. The robot is only the last of those three steps, and the least mature.
Do I need to buy a robot to benefit from any of this?
No. The two layers that pay for themselves in South Africa today, sensing and understanding, run on cameras, sensors and machine controllers most plants already own. A robot is the third layer, and at current prices it is a long way off for most operations here.
Will I have to replace my cameras, my ERP or my SCADA system?
Not for a platform built the way Dragonfly is. The point is to connect to what is already there and bring it into one picture, then have agents read that picture. Replacing systems is the expensive route, and it is not the one described in this piece.
How safe is it to let an AI "act" in a plant?
Safe, because the AI does not get to act on its own. Every action that reaches the physical world is approved by a human first: there is one checked, limited, recorded path for action, and its default setting is to recommend rather than act. The agent proposes, a person decides, and the decision is logged either way. Alarms keep working if the AI is switched off, and the device on the wall can see but not change anything.
What does this do to jobs?
The one South African case in this piece went the other way. Morgancoat's headcount has risen by more than 20 since it began automating seriously in 2020, and its managing director describes the aim as increasing the useful work each person can do. That is one company, not a law of nature, but it is a real number from a real factory.
Why is South Africa different from the countries in the videos?
Wages. A general-purpose robot's parts cost about a year of one worker's pay in America and more than a decade of pay here, which is why South Africa runs about 12 industrial robots per 10,000 manufacturing workers against a global average of 162. That gap applies to the robot layer only. Software costs the same everywhere.
Where would a manufacturer start?
Marc Black's answer is the practical one: find a real pain point, then find or build a system that solves it properly, and do not try to automate the whole factory on day one. The stocktake was Morgancoat's pain point. Yours will be different.
Who is behind Dragonfly, and what is riivo's part in it?
Sean Kelly is Dragonfly's chief technology officer. riivo, where the author works, builds the screens Dragonfly's partners and their customers use and is helping to build the safe path to act. Part of the platform is built from South Africa, and it runs in both South Africa and the United States. The full story is in our Dragonfly case study.
Sources and notes
- Amazon's robot fleet passing one million: GCN, "Amazon's warehouse fleet crossed a million," July 2025. gcn.com
- One humanoid model clearing 100,000 tote moves at a GXO warehouse in Flowery Branch, Georgia: Robotics & Automation News, 24 November 2025. roboticsandautomationnews.com
- Tesla's Optimus production target missed by more than a year: Electrek, 22 April 2026. electrek.co
- Amazon's Blue Jay warehouse robot shut down about four months after its October 2025 launch: Yahoo Tech, February 2026. tech.yahoo.com
- Ken Goldberg's "100,000-year data gap": Berkeley News, 27 August 2025. news.berkeley.edu
- Record venture investment in US hardware and fewer than half of warehouses over 250,000 sq ft using AI day to day: Silicon Valley Bank, "Physical AI and Robotics" report, 2026. svb.com
- US median wage for refuse and recyclable material collectors, $22.00 an hour ($45,760 a year): US Bureau of Labor Statistics, Occupational Employment and Wage Statistics, May 2023, code 53-7081. bls.gov
- South Africa's national minimum wage of R30.23 an hour from 1 March 2026: Labour Guide. labourguide.co.za
- Exchange rate used, about R16.17 to the US dollar: XE, September 2026. xe.com
- Humanoid robot component cost of roughly $55,000: GrabaRobot bill-of-materials teardown, 2026. A secondary source whose method is unaudited; used as an order-of-magnitude figure only. grabarobot.com
- Global robot density of 162 per 10,000 manufacturing employees: International Federation of Robotics, press release. South Africa's figure of about 12: RoboRadar24 density rankings compiled from IFR data. ifr.org and roboradar24.com
- The napkin arithmetic: author's calculation. US worker-year taken as the BLS annual figure of $45,760, giving 1.2 years for $55,000. South African worker-year taken as R30.23 x 45 hours x 52 weeks (the maximum ordinary hours under the Basic Conditions of Employment Act), or R70,738, converted at R16.17 to the dollar to $4,375, giving 12.6 years.
- Morgancoat: company description, stocktake times, pallet positions, headcount and all quotations from Marc Black are from his written answers to the author's questions, 10 September 2026. Background on the company: Packaging Magazine SA, "Morgancoat unboxed." packagingmag.co.za
- Dragonfly: all quotations from Sean Kelly are from his written answers to the author's questions, 14 September 2026. The description of the platform's customers and operating regions is the company's own.
- Disclosure: the author is a developer at riivo and has worked on the Dragonfly platform for the past year.
- Opening footage: "Industrial Robot Arm in High-Tech Factory" by Usman AbdulrasheedGambo, used under the Pexels free licence, which permits commercial use without attribution. pexels.com It is illustrative stock footage, not a riivo or Dragonfly site.
- Cover photograph: "Ceiling-Mounted Security Camera in Indoor Setting" by Eray Karataş, used under the Pexels free licence, which permits commercial use without attribution. pexels.com