In Kenya, a clinic tried to send one-way appointment reminders. Somebody left replies switched on.
Pregnant women started texting back. Soon the team was getting more than 1,000 messages a day. They had to set up a helpdesk. The AI came later, to help staff work out which messages needed a human first. Jacaranda Health’s technology director tells the story.
Imagine doing customer research the day before that mistake. How many women used the clinic’s text service to ask for help? Zero. How many wanted to? Easy to confuse those two numbers.
That gap is latent demand. People want to do something, but the cost, the effort or a missing reply button keeps them out. Once they can do it, an old sales chart can look like a spectacularly bad forecast.
1 person rides.
99 want to but don't. That is latent demand.
rides at this price wants to, priced out
Every hollow dot wanted a ride all along. A survey of today's riders would never find them, because they aren't riders yet.
What was the market for light?
Between 1800 and 2000, the real price of light in Britain fell roughly 3,000-fold. The average person used about 13,000 times as much, measured in lumen-hours. Think of lighting another room, leaving a lamp on to read, then lighting a street all night. You could have surveyed every candle buyer in 1800 and still missed most of what people would eventually do with light. Incomes rose and the technology improved over those two centuries, too. Fouquet and Pearson’s estimates.
What used to be too much trouble to try?
Vermont’s federal court recorded 45 self-represented cases in fiscal 2021. In fiscal 2024 it recorded 1,121. Immigration applicants with stalled cases were sharing instructions online for filing in Vermont, where the court moved quickly. Shah and Levy, Appendix A.
In a later Reddit guide, the writer described using Copilot to draft a complaint and paying a lawyer on Fiverr $150 to check it. The court records can’t tell us how many filers used AI. That one account does tell us what became feasible: a person who knew very little law could show up with a draft and buy just the review. The guide and its second step.
This is where I’d look for AI demand. Count the people who can now attempt the job at all. Then count what happens on the other end. A court still has to read every complaint. A clinic still needs someone to take an urgent call.
How big could Uber get?
In 2014, you could make a case for Uber with a pencil and the size of the taxi market. Try it, then pick any of the other 21 cases. Skip to the AI test.
Artificial light
In Britain, light got about 3,000x cheaper over two centuries. The average person used roughly 13,000x more.
How much did total UK use of light grow?
About 40,000x, with population growth on top of the per-person jump. Incomes also grew about 10x over the period, so price is not the whole story, but no income rise explains a 13,000x jump. William Stanley Jevons saw the same thing with coal in 1865. Better engines made Britain burn more coal.
When something gets thousands of times cheaper, people stop rationing it and start finding uses nobody had priced in.
The Penny Post
In 1839 Britain sent about 76 million letters. Postage was charged by distance, and the person receiving the letter often paid. In 1840 it became one flat, prepaid penny.
How many letters did Britain send in 1840?
Letters more than doubled in a year and reached 347 million by 1850. The Post Office paid for it. Its net revenue fell by about 73% in 1840.
A simpler price can unlock as much demand as a cheaper one. Volume and profit are separate questions.
Kodak Brownie
Kodak's 1888 camera cost $25. In 1900 it launched the Brownie for $1. Kodak's best year before that was 55,000 cameras.
How many Brownies shipped in the first year?
More than 150,000 shipped in year one. The price fell 25x, and the skill barrier fell with it, because the Brownie needed no training to use.
Skill is a barrier too. Remove it alongside price and a hobby becomes something families and children do.
Source ASME History of Information
Ford Model T
The Model T launched at $850. After the moving assembly line arrived in 1913, Ford kept cutting the price, eventually to $260.
How many Model Ts were built by 1927?
More than 15 million. The cut was only 3.3x in dollars of the day, but it crossed the line into what a working family could pay.
The size of the cut matters less than whether it brings in a group that has never bought before.
Source The Henry Ford
The shipping container
In 1956, loading loose cargo onto a ship cost about $5.83 a ton. That April, Malcom McLean sailed the Ideal-X from Newark to Houston with 58 containers on deck.
What did McLean's experts figure it cost per ton?
About 15.8 cents a ton, roughly 37x cheaper. That figure was the company's own estimate.
Cut the cost of a step that everything has to pass through, and you get uses nobody planned for. Global supply chains came out of a metal box.
Source Richmond Fed
The Southwest effect
When Southwest Airlines entered a route, fares on it fell by about half.
What happened to the number of passengers on that route?
Traffic more than tripled, as Southwest summarises a 1993 US Department of Transportation study. Herb Kelleher said Southwest was competing with ground transportation.
A 2x price cut can triple use when the new customers come from outside the market.
Source Southwest history
Roads and induced demand
A US city adds 10% more interstate lanes to fix its traffic.
What happens to the amount of driving?
Duranton and Turner found driving rises in proportion to interstate lane kilometres. Faster roads make trips worth taking that people had been skipping, and the traffic comes back.
New capacity fills up. More use and a solved problem are two different things.
Chik shampoo sachets
In 1983 a bottle of shampoo was a big one-time spend for a rural Indian family. Chik sold shampoo in single-use sachets for a rupee or less.
What was Chik's share of southern rural markets by 1990?
Over 50%, overtaking the market leader. The price per millilitre was not the barrier. The size of one purchase was.
Make the smallest purchase small enough and people who could never buy the big pack become customers.
Source Business Standard
ATMs and bank tellers
ATMs took over the bank teller's core job of handing out cash.
What happened to US teller jobs over the next 15 years?
Tellers per urban branch fell from about 20 to 13, but banks opened about 43% more urban branches because each one got cheaper to run. Then mobile banking arrived. Teller jobs are down nearly 30% since 2010, and the BLS projects another 13% fall by 2035. Spreadsheets ran a similar course. About 400,000 bookkeeping clerk jobs went and about 600,000 accounting jobs came.
A cheaper service can grow jobs for a while. A second technology can reverse it.
Iridium satellite phones
Iridium spent over $5 billion on phones that worked anywhere on Earth. Handsets cost about $3,000. Calls ran up to $5 to $9 a minute, depending on the source.
How many subscribers did it have five months after launch?
About 10,300. Iridium filed for bankruptcy in August 1999. Cheap cellphones got to most customers first.
If the cost doesn't fall far enough, or a cheaper substitute gets there first, the demand stays hidden.
Source TIME Via Satellite
Webvan
Webvan wanted to deliver groceries to your door. Before it had many customers per neighbourhood, it placed a $1 billion order with Bechtel to build warehouses.
What happened next?
Webvan went bankrupt in 2001 and laid off 2,000 people. Grocery delivery later worked for others.
The demand was real. The cost base and timing were wrong. A misfire is often a right guess at the wrong price.
Source Wikipedia
Segway
Dean Kamen expected to sell 10,000 Segways a week by the end of 2002. The consumer price was $4,950.
How many sold over the next six years?
About 30,000. Six years of sales added up to three weeks at the pace Kamen had predicted.
To test that forecast, find a commuter who would spend $4,950 to replace their usual ride.
Source Wired IEEE Spectrum
Cipla's generic HIV drugs
In 2001, triple-therapy HIV drugs cost about $10,400 a patient a year at US prices. Cipla, an Indian generic maker, offered them for $350.
About 400,000 people in developing countries were on treatment in 2003. How many people were on HIV treatment in 2024?
Air Deccan and cheap flying
In 2003-04 India had about 32 million domestic air passengers. Air Deccan started selling a few seats per flight for one rupee.
How many domestic passengers did India have in 2024-25?
About 339 million. Air Deccan itself took two years to fly its first million passengers, made losses, and was sold to Kingfisher in 2008.
The first flyers had never been airline customers. And the pioneer is often not the one who profits.
Rainfall insurance
Indian farm households named rainfall as their biggest risk. Insurers offered cheap policies that pay out when the monsoon fails.
What share of households bought one?
Only 5 to 10%. Price cuts moved take-up less than you would expect.
Cash in hand at planting time and trust in the insurer were the barriers that mattered.
Source Cole et al., NY Fed
M-Pesa
Most Kenyans had no bank account. M-Pesa let people store and send money by text message, cashing in and out at corner shops.
How many households did it lift out of extreme poverty, by one estimate?
Suri and Jack estimated 194,000 households, about 2% of all households in Kenya. The barrier was access and trust. Price had little to do with it.
Some latent demand waits for a door to open, not a price to fall.
Source MIT News, Science 2016
Airbnb
In 2008 sleeping in a stranger's spare room sounded unsafe to most people.
How many hosts does Airbnb have now?
Over 5.5 million hosts and over 2.5 billion guest arrivals. Reviews and held payments fixed trust on both sides.
Latent demand can sit on the supply side too. Millions of people wanted to earn from a spare room.
Source Airbnb
Tata Nano
The Nano cost about one lakh rupees, the cheapest car in the world. Tata planned for 250,000 a year.
How many sold in its best year?
74,527, in 2011-12. In June 2018 Tata made one. Ratan Tata later said the car "was termed the cheapest car by the public and, I am sorry to say, by ourselves".
Price fell and stigma stayed. Nobody wanted to be seen in the cheapest car.
Source Wikipedia Business Standard
Uber and the taxi-market mistake
In 2014, NYU finance professor Aswath Damodaran valued Uber at about $5.9 billion. He assumed it would take 10% of a $100 billion global taxi market.
How much did customers book on Uber in 2025, counting rides, deliveries and freight?
About $193 billion in gross bookings. That’s nearly twice the entire global taxi market Damodaran estimated in 2014. Uber’s total includes deliveries and freight as well as rides. Bill Gurley had already spotted a trip a taxi survey would miss. On visits to Los Angeles, he’d started using Uber instead of renting a car.
Who else wants a ride but would never show up in a survey of taxi customers?
Reliance Jio
In 2014, mobile data in India cost about ₹269 per GB. In September 2016 Jio launched with free data.
How long did Jio take to reach 100 million subscribers?
170 days. By 2018 data cost about ₹11.78 per GB, and use per person went from about 8 GB a month in 2018 to 31.1 GB by December 2025. Free data ended in mid-2017, and people kept paying.
The interest in the internet was always there. The price was hiding it.
Source RIL TRAI via TOI Nokia MBiT 2026
UPI
Before UPI, paying ₹10 for chai digitally was more hassle than handing over cash.
How many UPI transactions happened in October 2026?
25.47 billion in one month. Merchant fees have been zero by law since 2020, and the government pays banks an incentive instead.
Tiny payments nobody bothered to make digitally became most of the volume. And the state can be the payer that keeps it going.
Source Entrackr, NPCI data PMO India
MoviePass
MoviePass cut its price from $50 to $9.95 a month for a film a day. It passed 3 million members within a year, paying cinemas full price for every ticket.
What did three million sign-ups actually prove?
MoviePass was losing at least $40 million a month by July 2018 and shut in September 2019. But AMC's own A-List plan, at a higher price, reached 800,000 members and was reported profitable. The demand was real. The $9.95 price was the mistake.
A sign-up measures interest at the price on offer. MoviePass also had to cover the cost of each visit.
Source Business Insider NPR Engadget
Prices as quoted at the time, not adjusted for inflation, except the light series.
Who pays when everybody shows up?
MoviePass found plenty of people who wanted to go to the cinema more often. Its problem was that a heavy user could spend the monthly fee in a single evening. Compare that with a shampoo sachet. A family can afford a small packet long before it can afford a bottle. Both offers get someone through the door. Only one requires the seller to buy another cinema ticket every time the customer goes back. Try the sachet case.
A flat-rate AI plan has its own version of that MoviePass ticket bill. Each extra conversation uses computing power the provider pays for, while the customer’s monthly fee stays the same.
To see who’s paying for AI at work, look at the bills going through Ramp. In January 2023, 7.5% of businesses on its platform made a qualifying payment to an AI vendor. By October 2026, 56.8% did. The sample skews toward tech. That makes it a useful view of early buyers, and a poor estimate for every business in America. Ramp’s series and methodology.
Anthropic’s quarterly revenue reached $17 billion in Q3 2026, about 12 times what it was a year earlier. During that quarter it also cut some prices.
My bet is that paid AI work will stick. A lawyer can bill for checking a draft. A business can keep paying for a tool that saves it enough time.
Free consumer chat is a harder call. In February 2026, OpenAI reported more than 50 million paying consumer subscribers and more than 900 million weekly ChatGPT users. In India, people got a free year of ChatGPT Go. How many renew when the first bill lands?
Who gave up before they became a customer?
Boris Cherny, who created Claude Code, calls latent demand “the single most important principle in product.” His example is Facebook Marketplace. He says 40% of posts in Facebook Groups were already people buying and selling. Cherny tells the story at 5:40.
My terminal setup went from five commands to one. It saved me time on something I was already doing. The work I’d look for next is the kind someone wanted to do and abandoned.
Before your next product meeting, find someone who tried to do the job your product helps with and gave up halfway. Ask to see the unfinished bit. Help them past the step where they stopped, then see whether they finish.