Humane shipped around 10,000 AI Pins. By August 2024, four months after launch, more were coming back each day than going out. The Verge put total sales at roughly $9 million against the roughly $241 million the company had raised. HP bought the assets in February 2025 for $116 million and switched off the servers ten days later. The full autopsy is elsewhere in this series; what belongs here is the first mistake rather than the last.
Everything about that product was competent except the part where somebody checked whether people wanted to wear it.
Two years of gadgets nobody asked for
Rabbit sold about 130,000 r1 units off the back of a CES demo and $10 million in pre-orders, with industrial design by Teenage Engineering. Pre-orders proved that people would buy it. They did not prove that people would use it: by September 2024 The Verge reported that of roughly 100,000 purchasers, about 5,000 were using the device daily.
Friend went further in the other direction. Avi Schiffmann paid $1.8 million for the domain and then over a million more for a New York subway campaign in September 2025, covering 11,000 rail cars and a thousand platform posters. People vandalised the posters. Heineken ran parody ads. The campaign generated enormous attention and very little evidence that anyone wanted an always-listening pendant that texts you.
At CES 2026 the pendants came back anyway. Lenovo showed a prototype for Motorola’s assistant, iBuddi showed a medallion, Looki showed continuous point-of-view capture, and Amazon and Meta arrived having bought Bee and Limitless rather than building their own. Bee is the interesting one: a $49.99 bracelet with a $19 monthly subscription, built on $7 million of funding, sold to Amazon in July 2025. Cheap hardware, expensive service, tested in market before anyone spent nine figures proving it.
The pattern across all of these is that a plausible idea shipped badly does more damage than an idea that stays in the drawer. Humane’s customers ended up with dead objects and no refund. Rabbit’s customers ended up with a device they stopped opening. Every one of those buyers is now harder to sell to, and so is everyone who read about it. Not shipping costs you time. Shipping something that fails in public costs you the audience.
Rank your evidence by what it costs the person giving it
Most founders collect the cheapest signals and then act on them as though they were expensive ones.
A like costs nothing and means nothing. An email address on a waitlist costs three seconds and a throwaway alias; treat it as interest in the topic, not in your product. A survey answer costs less than that and is actively misleading, because people are generous with hypothetical money and stingy with real money.
A refundable deposit is where useful information starts. Someone has entered a card number, accepted a charge, and let it sit on their statement. It still overstates demand, because they know they can get it back, but the drop-off between waitlist and deposit is the single most informative number you will collect early. If a thousand people gave you an email and eleven give you fifty euros, you have learned something that no amount of user interviews would have told you.
A non-refundable pre-order at close to the real price is better again. A purchase order from a business, with a delivery date and a penalty clause, is better than all of it, which is one reason B2B hardware is easier to validate than consumer. And the signal that beats everything is the second purchase: the customer who buys another one for a colleague, or renews the subscription in month thirteen.
Set your threshold before you run the test. Write down the number that would make you continue and the number that would make you stop, and write it down before you see any data, because you will otherwise rationalise whatever you get. Most people who skip this step discover that no result is capable of stopping them.
Fake the product, run the service by hand
For a connected product, the hardware is usually the least uncertain part. Your bet is that the thing the hardware enables is worth paying for, and you can test that without building anything. Most of the small pilots we set up at RMBG start with somebody else’s device in a printed enclosure.
Something close enough almost always exists: a Bluetooth recorder, an off-the-shelf tracker, an ESP32 dev board, a sensor kit from a European distributor. Buy twenty and ship them to twenty real users, and deliver the service behind it manually. Read the data yourself in a spreadsheet. Write the daily summary by hand and send it as an email at seven in the morning. Answer the alerts from your phone. Charge for it.
The manual version answers questions a prototype never will. Do people actually open it. Do they still open it in week three. Does the thing they complain about have anything to do with the electronics, or is it the enrolment flow, or the fact that the summary arrives too late to be useful. Plenty of products die here, and they die for about four thousand euros rather than four hundred thousand.
The trap is that concierge tests flatter you. Twenty users getting hand-written attention will love the service and tell you so. Watch the behaviour rather than the enthusiasm, and put a price on it early. A user who says the daily summary is brilliant and then cancels at nine euros a month has told you the real answer.
A pre-order is a loan, and in Europe it has terms
Pre-orders are the most honest funding available to a consumer hardware company and the easiest to abuse. In the most widely cited study of the platform, about 9% of successfully funded Kickstarter projects never delivered at all. An analysis by DTU Science Park of 25 failed hardware campaigns found backers had put in a little over $26 million with nothing usable to show for it, and the causes were mundane: design problems, PCB defects, quality control, products priced below what they cost to make.
The European position is stricter than founders realise. Under Article 18 of the Consumer Rights Directive, a trader must deliver “without undue delay, but not later than 30 days from the conclusion of the contract” unless the parties have agreed otherwise. That “unless agreed otherwise” is doing all the work, and it means your pre-order page has to state the shipping window in terms the buyer actively accepts. Miss it, and the customer can set a further reasonable period and then cancel for a full refund, no matter what your terms and conditions say. The 14-day right of withdrawal for distance sales sits on top of that.
So run pre-orders like a company that expects to be held to them. Quote a date you would still hit if your first tool needs a rework and your certification lab has a queue, then add a quarter. Do not spend the deposits on the deposits: keep pre-order money in a separate account and fund operations from equity or debt, because a company that is paying this month’s salaries out of next year’s deliveries has already failed and just does not know the date yet. Publish a refund policy that a normal person can read in a minute, and honour it without argument. And write to your backers monthly even when there is nothing good to report, because silence is what converts a delay into a scandal.
A working prototype stopped being evidence
A previous post in this series covered how cheap a first build has become: two or three weeks and a few thousand euros now gets you a looks-like, works-like unit.
That is excellent news for building and terrible news for validating, because it is now true for everyone. The demo that got a standing ovation in 2018 is table stakes. When you show a room your prototype and they say it looks amazing, they are responding to a printed enclosure and an animation, and so is every other founder in the same category that week. Rabbit had a spectacular demo. It sold 130,000 units on the strength of it and retained roughly one user in twenty.
Assume your prototype tests desirability of the concept, and nothing else. It cannot tell you what someone will pay, whether they will keep using it, or whether the promise survives contact with a normal week. Only money and elapsed time tell you that, which is why the deposit and the twenty-unit pilot matter more than the render.
Search 1688 before you write the pitch deck
Before anything else, spend an afternoon on Alibaba and 1688 searching for your product in the plainest possible language. Not your positioning, the object: “wearable audio recorder”, “smart pill dispenser”, “LoRa soil sensor”. Search in Chinese via translation, because 1688 listings are aimed at the domestic wholesale trade and are usually a year ahead of what appears in English.
One of two things happens. Either nothing close exists, in which case find out why, because the absence of a white-label version of an obvious product usually means a certification barrier, a patent thicket, or a category that has been tried and does not sell. Or you find fourteen suppliers offering something within 20% of your concept at a third of your intended price, which is the more common outcome and is not necessarily bad news. It just tells you what business you are in. If the object already exists as a commodity, your company is a brand, a software layer and a distribution channel, and every euro of your budget should go there rather than into industrial design.
Plaud is the example worth studying. The founders were an operator and a Shenzhen factory owner who had already sold an app-controlled recorder called Izyrec before they pivoted to AI notetaking after ChatGPT launched. They ran a Kickstarter in 2023 that raised over $1 million in pre-orders despite tripling the price during the campaign, which is about as clean a demand test as exists. Forbes reported the company on track for around $250 million in annualised revenue in 2025, over a million devices sold, profitable, and bootstrapped, with roughly half of revenue coming from subscriptions. The hardware is a recorder. Anyone can buy a recorder. The company is the service and the distribution, and they knew that from the start.
Patents rarely arrive in time to matter
First-time founders spend disproportionate energy here. In the US, practitioner summaries of USPTO data put average total pendency at roughly 26 months for FY2022, with current estimates of 24 to 32 months depending on technology area, and more than 90% of applications receive an initial rejection. All-in cost for a complex utility application commonly runs $15,000 to $30,000 or more. Your consumer product’s commercial life may be shorter than that timeline.
None of which means file nothing. It means file cheaply and quickly to hold a date, then spend your money on shipping. A provisional or a first national filing establishes priority and gives you twelve months to decide whether the thing is worth the European and US costs, by which point you will have real sales data to decide with. Registered design rights are the underrated instrument for consumer hardware: in the EU they are fast, cost a few hundred euros, and protect the thing a copyist actually copies, which is the shape. Keep your calibration routines, your data pipeline and your manufacturing process as trade secrets rather than publishing them in a specification for competitors to read.
What defends a small hardware company in 2026 is usually the accumulated customer data, a service that improves every month, the retail and distribution relationships you have built, and being two product generations ahead by the time anyone bothers to clone you. Patents do matter enormously when you sell the company, and HP’s purchase of Humane’s roughly 300 patents is a reminder of that, though an asset sale is not the outcome you are planning for.
What honest validation looks like on a calendar
reMarkable is a useful shape to copy, and it is European. Magnus Haug Wanberg founded it in Oslo in 2013 with an unfashionable idea, a tablet that only does paper. Pre-orders opened in November 2016 and took about $12 million. First units shipped on 29 August 2017, roughly nine months later, and the company did around 45,000 units and NOK 184 million of revenue that year. In 2024 it sold about 791,000 devices for $434.3 million in revenue, $24.7 million of net profit, and 568,000 paying subscribers. The pre-order was a demand test the company then honoured on a date it had published, and the compounding happened afterwards.
A version of that fits in a quarter. Week one, the 1688 search and a serious look at who has already tried this and how they did. Weeks two and three, a landing page with a real price and a deposit button, driven by a few hundred euros of ads pointed at the narrowest audience you can define. Weeks four to eight, twenty concierge units with off-the-shelf hardware, charged for, watched daily. Weeks nine to twelve, either a pre-order campaign against a schedule you would bet your own money on, or a decision to stop.
The stop decision is the one that pays for the whole exercise, and it is the one most founders build their process to avoid ever reaching. Humane raised roughly $241 million to learn something a hundred paid deposits would have suggested in a month, and the ten thousand people holding a dead pin paid for that lesson as well.
Sources
- Returns of Humane AI Pin outpacing sales, August 2024 — MacRumors
- Humane’s AI Pin is dead as HP buys the startup’s assets for $116M — TechCrunch
- Rabbit r1 — Wikipedia
- Friend (product) — Wikipedia
- AI pendants and wearables return to CES, promising personal assistants around your neck — Malay Mail
- Amazon acquires Bee, the AI wearable that records everything you say — TechCrunch
- How an AI notetaker became one of the few profitable AI startups — Forbes
- $26 million lost: why crowdfunded hardware projects fail — DTU Science Park
- Directive 2011/83/EU on consumer rights, Article 18 (delivery)
- How long does a patent take from filing to grant — The Rapacke Law Group
- reMarkable (manufacturer) — Wikipedia
- reMarkable closed funding at a $1B valuation after selling 1M devices — TechCrunch