The Margin Math

The customer who cost me money

There is a particular stillness that comes over you when a bill is wrong in the wrong direction. Your hands stop on the keyboard. You stop breathing through your nose. Somewhere in the house a pressure cooker whistles and you do not hear it.

Meera's question sent me to the API dashboard that night. (API stands for application programming interface; in plain terms it is the doorway my software knocks on every time it asks the AI model to do something, and the model company charges me for each knock.) I had glanced at that dashboard before, the way you glance at a dentist's waiting room. I had never actually gone in and sat down.

That February's bill was $212. Twenty-three agencies paying me, and a model bill of $212 seemed fine, even modest, until I sorted the usage by account and one name sat at the top like a boulder.

The account belonged to a man called Darren in Salford. He was on the $15 Solo plan, a cheaper tier I had added in January after a few independent landlords in the Facebook group asked for something smaller than the agency price. Darren was not an independent landlord. Darren, it turned out, ran a small business that tidied up listings for other people, and he had written a little script that fed my tool every property description he could find, one after another, all day and most of the night.

Forty thousand one hundred and twelve requests in a month.

A normal agency used somewhere around 600. Each request cost me roughly a third of a cent (I worked it out later at about $0.0035, though the price moved with the length of the listing), which sounds like nothing until you multiply. Darren alone had cost me about $140 in model fees, roughly ₹11,700, which was more than my whole electricity bill for the winter. He had paid me fifteen dollars. After the card fee, a bit less.

So my most enthusiastic user was also, by a distance, my most expensive. He loved the product. He used it more than anyone alive. And every morning he logged in, I was quietly paying him for the privilege.

I sat very still for what Meera later claimed was four full minutes. Then I scrolled down to check whether anyone else looked like Darren. Two more Solo accounts were climbing the same way, smaller but steeper.

"Can you switch him off?" Meera said.

"He hasn't done anything wrong. The plan says unlimited."

"Who wrote 'unlimited'?"

I had. In January, at about midnight, because it looked generous on the pricing page and the agency plan said it too and I never imagined anyone would take the word literally. I opened the page again. There it was in a friendly green tick, right next to the price.

"So he's your best customer," she said, "or your worst."

The pressure cooker had boiled dry in the kitchen. Neither of us had noticed.

One page, five numbers

The next morning I took the whiteboard in my office, rubbed out a half-finished sketch of a logo I would never use, and wrote five words down the left-hand side. Price. Model. Hosting. Fees. Support. That was the whole system. It still is, give or take the smudges.

The idea is simple enough to embarrass me now. For one ordinary customer, you write down what they pay you in a month, then subtract every cost they cause. Whatever is left is your margin per customer, the money that actually belongs to you. If that number is negative, every new sign-up makes you poorer, and growth only digs the hole faster.

Here is the box I wish I had drawn in January. Do it with your own numbers. Five minutes, a pencil and the back of anything.

Napkin: Margin Math for one customer

Take one typical customer on your main plan and fill in a month:

• Price: what they pay. My agency plan: $49 (about ₹4,100)
• Model cost: their requests × cost per request. 600 × $0.0035 = $2.10 (₹176)
• Hosting: your server and database bill ÷ number of customers. $40 ÷ 23 = $1.74 (₹146)
• Payment fees: roughly 2.9% plus 30 cents per card charge. $1.72 (₹144)
• Support: hours spent helping all customers × what an hour of you is worth, ÷ customers. 6 hours × $15 ÷ 23 = $3.91 (₹328)

Margin = $49 − $9.47 = $39.53 a month (about ₹3,300)

Now do the same sum for your heaviest user.

The support line is the one people skip, and I skipped it too for a long time, because my hours felt free. They are not. I picked $15 an hour because that is roughly what I would pay a freelancer in Jaipur to answer emails competently, and I am guessing a bit, but leaving that line blank means pretending you work for nothing. You will not keep doing that cheerfully for very long.

Then I did the last line of the box, the one that matters.

Darren paid $15. His model cost was $140. Hosting, the same $1.74 as everyone. Card fee, about 74 cents. Support, almost nothing, because Darren never once wrote to me (he was far too busy). Total cost: roughly $142.50.

Margin: minus $127.50 a month. About ₹10,700, gone, every month, to a man in Salford I had never spoken to.

What unsettled me most was that the agency plan looked healthy. My average across all twenty-three accounts was still positive, and if I had only ever calculated the average I would have felt fine for another six months. Averages are generous like that. They let one Darren hide behind twenty agencies in Leeds and Harrogate who write a few listings on a Tuesday and then go home.

I wrote "−127.50" in red marker under Darren's name, circled it twice, and left it on the whiteboard. It stayed there until the summer, slowly fading under my thumb every time I reached past it for the duster.

Nine dollars and a nervous stomach

The Solo plan nearly cost $9. I had typed that number into the pricing page first, late one January night, because $9 felt like a price nobody could be angry about. Arjun saw it on a video call the next day and laughed so hard his camera froze.

"Who is it for?" he said.

"Small landlords. People who can't afford forty-nine."

"Did any of them tell you that, or did you decide it for them?"

I had decided it for them. That is the whole problem with $9 in one sentence. It is a price set by the founder's nerves, not by the buyer's wallet, and I hold this opinion more firmly than almost anything else in this book: for a small AI product, $9 a month is usually a mistake. Run the napkin on it. The card fee takes about 56 cents. Hosting takes $1.74. A customer who uses the tool properly burns $2 of model costs, and if they email you twice you have spent half an hour you will never bill for. You are left with perhaps $3 (₹250), and one keen user wipes out the profit from twenty quiet ones. Then there is the other cost, which is harder to put on a whiteboard. Cheap plans attract people who were never sure they needed you, and they leave first.

The agents taught me how to think about price instead. A single let in Headingley earns Claire's agency a fee of roughly $1,000. A listing that goes live a day sooner, or that doesn't get flagged for a missing EPC rating, is worth far more than $49 to her. She was never comparing my price with Netflix. She was comparing it with an hour of her own Friday.

So I rebuilt the page in an afternoon:

• Solo, $19: 150 listings a month, then a polite message.
• Agency, $49: 1,000 listings, which is almost twice what a normal agency uses.
• Volume, $249: 50,000 listings, for people like Darren.
• Annual on any plan: pay for ten months, get twelve.

The word "unlimited" went in the bin. The limits are generous enough that ordinary customers never see them, and that is the point. A limit is there for the one account in fifty that would otherwise drown you, not to nag everyone else.

The annual plans surprised me most. Four agencies switched in the first week, which put $1,960 in my account at once (about ₹1.6 lakh), and an agency that has paid for a year does not quietly cancel in March because the owner is tidying up his card statement.

Linda, when I told her I was raising prices for new customers, wrote back: "About time. I was starting to worry about you."

Then I emailed Darren. I kept it short and did not apologise. He replied eleven minutes later.

"Fair enough, mate," he wrote. "Didn't think it'd last."

He upgraded that afternoon. I rubbed nothing off the whiteboard. I just wrote "+$106" beside the red number, in blue, and stood back to look at the pair of them.

The quiet leak

Churn is the share of your paying customers who leave each month. Mine was about 6%, and for a long time I thought that sounded fine. Ninety-four out of a hundred people stay. Lovely.

Then one evening in May, with the fan rattling and the chai gone cold, I did the sum properly. Start with 100 customers and lose 6% every month. After one year you have 48 left. After two years you have 23. More than three-quarters of the people who once trusted you enough to type in a card number have walked out, and most of them never said why.

It gets worse when you add new customers. Say you win ten a month, which for me was a good month. At 6% churn, your business stops growing at around 167 customers, because by then you are losing ten a month as fast as you win them. You can work harder and harder and the number simply will not climb. I had been feeling vaguely stuck since February. This was the reason, written in my own handwriting.

So I went looking for the leaks, and they were small and embarrassing.

• Failed cards. Nearly a third of my cancellations were not cancellations at all. Cards expired, and the payment failed quietly. I set up three polite reminder emails before an account closed.
• The silent first week. Accounts that wrote no listings in their first seven days almost always left. I started sending a short personal email on day three, from me, asking what they were working on.
• One question on the way out. "What made you cancel?" with a free-text box. Nobody had to answer. About half did.
• Annual plans, which I have already confessed to loving.

The exit question was the humbling one. A woman called Bev, from an agency in Knaresborough, wrote: "Lovely tool. Forgot I was paying for it until my accountant asked." She had not left because the product was bad. She had left because I had never once reminded her it was doing something useful. I added a monthly email telling each agency how many listings they had written and roughly how many hours that saved. Bev came back in September.

By the end of summer, churn was a little under 3%. Same product, same prompts, same price. At 3%, ten new customers a month stops at around 333 instead of 167, and after two years you keep 48 of every hundred rather than 23. Nothing I have ever built in a weekend did as much as those four emails.

I showed Meera the new numbers on the whiteboard, the blue beside the red, the churn curve bending upwards at last.

She looked at it for a while. "So how many is enough?" she said.

I opened my mouth to answer and found I had nothing to say. I had a ceiling, a margin and a churn rate. I did not have a number for our actual life.