The Window Is Open (and It Has a Sharp Edge)

The weekend that used to be a year

At 1.14 on a Sunday morning in March 2024, my friend Arjun sent me a screen recording with no message attached. I watched it on my phone in the dark, trying not to wake Meera, and it took me a full minute to understand what I was looking at.

It was a booking system for a chain of physiotherapy clinics in Pune. Patients picked a therapist, chose a slot, got a WhatsApp reminder the night before, and the front desk saw everything on one tidy screen. I knew this software. In 2019, Arjun had built it at the agency where he worked, with three other developers and a project manager who wore the same grey hoodie every single day. It took them seven months. The client paid ₹32 lakh, which was about $45,000 at the time, and there were two near-resignations and one actual one along the way.

Arjun had rebuilt the whole thing, from nothing, since Friday evening.

"How much of that did you write?" I typed back.

"Honestly? Maybe a fifth," he wrote. "I mostly argued with the AI until it got it right. Cost me the $20 subscription and about $11 in API credits. And a lot of chai."

An API, if you have not met the word, is simply a doorway one piece of software uses to ask another piece of software for something. In Arjun's case he was paying a few cents each time his code asked an AI model to write or fix more code. Eleven dollars. Against forty-five thousand.

I want to be careful here, because Arjun is a very good developer and that matters. He knew what the clinics actually needed, because he had sat through those meetings in 2019. He knew when the AI was confidently wrong, which it was, several times, including one memorable moment when it booked a patient into the same slot as herself. His weekend version was rougher than the original. It had no proper security review. Nobody had tested it on a ten-year-old Android phone in a clinic with patchy Wi-Fi.

But it worked. A receptionist could have used it on Monday.

Here is the part that kept me awake longer than the video did. The expensive, slow, specialist bit of making software (turning an idea into something that runs) has collapsed in price by something like a thousand times in about five years. I am rounding, and you could argue the number either way, but not by enough to change the point. That collapse does not only help people like Arjun. It reaches the dental receptionist who has an idea about insurance forms, the retired accountant in Leeds, the schoolteacher who knows exactly which spreadsheet ruins her Sunday evenings. People who would never have been allowed near a codebase can now describe what they want in plain English and watch it appear.

I lay there until nearly three, refreshing the video. In the corner of the frame, half hidden behind his code editor, I could see Arjun's old 2019 project timeline, still pinned to his wall, with "LAUNCH (hopefully)" circled in red.

Cheap code, expensive trust

When something gets a thousand times cheaper, the cost does not vanish. It moves. All that money and effort that used to go into building software is still out there. It has simply drained towards the parts that stayed hard.

Two of those parts matter more than the rest. The first is attention: getting one particular person, with one particular problem, to notice that your product exists at all. The second is trust: that same person deciding you are real, that you will still be around next June, and that typing their card number into your little form is not a mistake they will have to explain to their spouse. Neither of these got cheaper in 2024. They got dearer, if anything, because Arjun's weekend is now everybody's weekend, and every inbox on earth is filling up with people who also built something since Friday.

I learnt this before the AI boom made it obvious, and I learnt it slowly. My first product rewrote product descriptions for small online shops. It took me eleven days to build and four months to sell to a single human being. I will tell you the full, faintly humiliating story of those four months later in the book. For now, the ending is enough.

Her name was Linda. She ran a candle shop in Ohio, and she paid me $19 a month (about ₹1,600, which my cousin pointed out was less than he spent on samosas in a fortnight). I asked her, a few weeks in, why she had picked my tool. There were other description writers around by then. I am guessing here, but probably dozens, several of them better built than mine and backed by companies with actual offices.

She said it was the demo video. My neighbour's dog had barked all the way through the recording, and I had been too tired to redo it. To Linda, the barking meant a person was behind the screen. A person she could email. A person who might answer.

Think about what she was not paying for. She was not paying for my prompts, which anybody could copy in an afternoon. She was not paying for my code, which was, to be generous, held together with hope. She was paying $19 for the feeling that if something broke on a Saturday before Mother's Day, somebody in particular would care.

That feeling is the scarce thing now. You cannot generate it in a weekend, and you cannot prompt your way to it. You earn it one Linda at a time, usually in ways that look embarrassingly small from the outside: a reply sent within the hour, a bug fixed while she was still on the phone, a note remembering that her shop does a winter sale.

The first email Linda ever sent me, after her card went through, had nothing to do with software.

"Is the dog okay?" she wrote. "He sounded upset."

The Wrapper Test

In the spring of 2023 a friend of a friend, a quiet engineer in Pune I will call Nikhil, built a tool that let you upload a PDF and ask it questions. Drop in a 90-page rental agreement, type "can my landlord keep the deposit if the geyser breaks?", and get an answer with the page number. It was lovely. It took him nine days. Within four months he had about 1,100 people paying $12 a month, which is roughly $13,000 a month, and he had started looking at flats with a balcony.

Then the company whose model he was renting held a product announcement. Somewhere around minute twenty, a presenter dragged a PDF into the ordinary chat window and asked it a question. No extra charge. Nikhil watched it live. By the end of the following month he had lost more than half his customers, and the rest left slowly, politely, the way people leave a party once the host starts stacking chairs.

"I didn't have a company," he said to me later, over a very long phone call. "I had a feature they hadn't got round to yet."

That sentence is the whole test. If the company that makes your AI model, or some teenager with a free browser extension, could ship your core feature within a month, you have a demo and no business. The model is the engine. Everybody can buy the same engine. You need something bolted around it that the giants will not bother to build.

There are four things that tend to work, and Nikhil could have reached for any of them.

He could have owned a workflow: instead of "chat with any PDF", the full Thursday routine of a small law office, where the contract arrives by email, gets checked against the firm's own clause list, and lands in the right client folder with a summary attached. He could have picked a niche so narrow that a big company would find it beneath notice: shipping documents for spice exporters in Kochi, say. He could have built up data nobody else had, such as thousands of annotated tenancy agreements showing which clauses actually caused disputes. Or he could have held a relationship: the kind of customer who emails you about the dog.

Napkin: the Wrapper Test Write your product idea in one line. Then answer honestly:

• Could a big AI provider add this as a button next Tuesday? Yes / No
• Could a free browser extension do 80% of it? Yes / No
• Which of the four layers do you own today? Workflow / Niche / Data / Relationship / None

Two "yes" answers and a "none" means you have a weekend project. That is fine, as long as you know it.

Nikhil is building again now. Freight invoices for small importers, a deeply boring problem. He sent me a photo last month of his new desk. Taped to the monitor was a printout of that announcement's date, with nothing else written on it.

The Small Machine

A developer friend of mine in Bengaluru, Karan, raised $2.4 million in 2021 for an app that helped restaurants plan their staff rotas. He had a proper office with a beanbag nobody sat on, eleven employees and a slide deck with a line on it that went up and to the right. Three years later the company was sold for $1.9 million. The investors had a clause, standard in these deals, that let them take their money back first. They took all of it. Karan walked away with nothing except a very good chair he bought from the liquidation.

"I worked harder than I've ever worked," he said, "to make other people slightly less poor."

I do not tell that story to sneer at him. He is cleverer than me, and the rota app was better built than anything I have shipped. I tell it because his ending is normal. A Harvard researcher, Shikhar Ghosh, looked at venture-backed companies and estimated that about three in four never return their investors' money. I am guessing here, but if the investors usually get paid before the founders do, the share of founders who come out with a life-changing sum must be smaller still. Perhaps one in twenty. Perhaps fewer. Nobody publishes those numbers with much enthusiasm.

The funded route is a rocket ship. It is built to reach orbit or explode, and the people paying for the fuel are quite relaxed about the explosions, because they own a hundred rockets and need only one to land on the moon. You own one rocket. You are sitting in it.

So here is my strong opinion, and you are free to disagree with it over chai: for almost everyone reading this, raising money is the wrong goal, and small is a strategy rather than a consolation prize.

What I want you to build instead I call the Small Machine. It is a lean software business run by one to three people. It solves one dull problem for a group of customers who know your name. It earns more each month than it costs, and it keeps working when you step away for a fortnight to attend a cousin's wedding in Udaipur. Nobody will write about it in a magazine. You will not ring a bell anywhere.

The arithmetic is gentler than you might expect. Three hundred customers paying $49 a month comes to $14,700 (about ₹12.5 lakh) before costs. You need no investors, no beanbag and no board meeting. Most people who build a Small Machine never needed a moon. They needed their rent covered, their evenings back and the dignity of answering to customers rather than a boss.

My own machine is not glamorous. It pays our bills, and some months only just. When I explained the plan to my wife, Meera, years ago, she listened carefully and then asked the only question that mattered.

"Fine," she said. "But what does it actually do, and who has that problem on a Friday afternoon?"

I did not have an answer. Finding one took me a long time, and it began with a dentist's receptionist in Leeds who retyped the same invoice every single week.