पाठशाला Pathshala · उत्पाद Utpād, The product · Lesson 19 · Build
Your first ten paying customers and what they get wrong
The first ten customers are the best signal a company will ever get and the worst product managers it will ever hire. Use what they do; be careful with what they ask for.
Pathshala, The Founder Library · 11 October 2026 · 7 min read
The first ten customers who pay are the most valuable people the company will meet, and the most dangerous. They prove that someone will part with money. They also arrive with a list of requests, and a founder who builds the list ends up with a product shaped for ten companies that are not like the next thousand.
This lesson is about the gap between those two facts. It sets out why the first ten are not a sample of the market, the four ways they mislead, how to separate the signal in what they do from the instruction in what they ask, how to measure the risk when one of them pays most of the bills, a worked example, and a monthly review that keeps the product pointed at the next hundred.
Why the first ten are not your market
Paul Graham’s Do Things That Don’t Scale is the standard instruction for this stage: recruit users by hand, and make the first ones so happy that signing up feels like one of the best choices they ever made. He also says something founders read past: your initial model of users is always inaccurate, even if you are one of them. The first ten are where that inaccurate model gets corrected. They are not where it gets confirmed.
The reason is selection. A customer who pays for a product with half its features, no references and a founder answering support on WhatsApp is unusual by definition. They feel the problem more sharply than most, they tolerate more friction, and they often bought because they know the founder, a former colleague, a batchmate, a client from a previous job. Each of those traits makes them a good first customer. Each also makes them a poor guide to what the mainstream buyer in the same segment will need.
The four ways they mislead
They tolerate what others will not. An early customer will import data by hand, put up with a missing report and forgive a bad week. Their retention says the core value is real. It does not say the onboarding works, because they never needed it to. They are too close to you. Customers from your network renew partly out of loyalty and soften their complaints. Their churn is lower and their feedback is kinder than the market’s will be.
They are not the same size. In business software the first ten usually include one or two accounts several times larger than the rest, and the founder treats their requests as orders. They describe solutions. Each request arrives as a feature, phrased in the customer’s own workflow: add a column for our branch code, export in the format our auditor uses, let us approve in two steps because our MD wants to. Built one at a time, these turn a product into ten consulting projects that share a login screen.
Graham describes the deliberate version of that trap as a technique, not a mistake: pick a single user and act as consultants building something just for them. It works at the start because it teaches the founder the job in detail. It fails when it continues past the point where the founder has learned what the job is, and the first ten paying customers are usually that point.
Signal in what they do, instruction in what they ask
Treat the first ten as a panel to observe rather than a committee to satisfy. The strongest evidence is behaviour: which features each customer uses weekly, which they opened once, where they stall, what they do outside the product to finish the job. Pull it from the [events you instrumented](/library/instrumenting-product-twelve-events) and from watching them work. The second strongest is what they would lose. The weakest is the feature they ask for, which is a hypothesis about a problem, written by someone who is not paid to solve it.
Superhuman’s method is the cleanest way to sort the panel. Rahul Vohra describes asking users how they would feel if they could no longer use the product, with Sean Ellis’s benchmark of 40 per cent very disappointed as the bar. Superhuman scored 22 per cent at first. The team then looked at who the very disappointed users were and narrowed the target to those personas, which lifted the score to 33 per cent before a line of code changed. Vohra’s rule for the rest is blunt: politely disregard those who would not be disappointed. Among the somewhat disappointed, listen only to those who named the same main benefit as the lovers, because something small is holding them back. Spend half the time deepening what the lovers love and half removing what holds that second group back. Within three quarters the score reached 58 per cent.
With ten customers the percentages are noise, as the [lesson on the Sean Ellis test](/library/sean-ellis-test-and-where-it-misleads) explains, but the sort still works. Ask the question in person. Put each customer in one of three groups. Read their requests only through the lens of the group they sit in.
When one customer is a quarter of your revenue
Concentration is the measurable form of the problem. If one customer pays 35 per cent of revenue and two more pay another 25 per cent, the company does not have ten customers in any sense that matters for risk or for the roadmap. Losing the biggest would take out a third of revenue in one email, and every request from that customer carries the weight of three or four smaller ones in the founder’s mind, whether or not it should.
The figure turns the share into a single number. The effective number of customers is the inverse of the sum of squared revenue shares, the Herfindahl–Hirschman index competition regulators use for market concentration. Ten customers paying the same count as ten. Ten where one pays 35 per cent and the next two 25 per cent count as about four.
Set the biggest customer at 10 per cent and the next two at 20 per cent: close to ten effective customers and no single exit that ends the company. Push the biggest to 50 per cent and the effective number falls below four, whatever the logo slide says. Two decisions follow. Until the effective number is above half the headcount, treat any request from the biggest customer as a risk to price, not a priority to schedule. And make the next five sales deliberately smaller and more typical, because they repair the panel faster than any feature can.
Build for what the first ten do and what they would miss. Read what they ask for as evidence about a problem, never as a specification.
A worked example: ten jewellery workshops in Jaipur
A Jaipur company sells stock and job-work software to jewellery manufacturers. After seven months it has ten paying customers and ₹2.1 lakh of monthly revenue. One export house pays ₹70,000 a month; two mid-sized workshops pay ₹28,000 each; seven small workshops pay between ₹10,000 and ₹15,000. Four of the seven were introduced by the founder’s uncle, who trades in the same market.
The request list has 31 items. The export house wants multi-currency invoices, a second approval step and a report in its auditor’s format. Two workshops want karigar-wise piece-rate tracking. The founder’s plan was to build the export house’s three items first, because it pays a third of revenue.
The review changes the plan. Usage shows that all ten use the issue-and-receive register for gold sent to karigars every day, and that five of the seven small workshops also keep a paper register beside it, because the software cannot record weight loss in the process. Asked how they would feel without the product, six say very disappointed, all of them small or mid-sized workshops whose main benefit is knowing where the gold is. The export house says somewhat disappointed; its main benefit is invoicing, which three other products do. Effective customers by revenue: about six.
The quarter’s work becomes process-loss recording and karigar-wise tracking, which serve eight of the ten and the market of small workshops behind them. Multi-currency invoicing goes on the list of things the company will not build this year, and the export house is offered an export file its own accountant can use. The sales target is fifteen more small workshops, none introduced by family, which will bring the effective number above ten and test whether the uncle’s introductions were the reason the first four stayed.
The monthly first-ten review
Once a month, for an hour, until the company has fifty paying customers. Compute the effective number of customers by revenue and write it beside the headcount. For each customer, record weekly usage of the three core features and one sentence on what they do outside the product to finish the job. Ask each, in person or on a call, how they would feel without the product and what its main benefit is, and keep the answers in one sheet.
Then read the request list against that sheet. A request moves forward only if it comes from a very disappointed customer, or a somewhat disappointed one who names the same main benefit, and if at least three customers with no personal tie to the founder face the same problem. Everything else gets a written answer, as the [lesson on saying no](/library/feature-requests-discipline-of-saying-no) describes. Finally, check the next five deals in the pipeline against the [ideal customer profile](/library/ideal-customer-profile-on-one-page): the fastest way to stop the first ten designing the product is to make sure they are soon a minority.
The Jaipur company is illustrative. The method is Graham’s and Vohra’s; the concentration measure is the Herfindahl–Hirschman index.
Sources
- Paul Graham, Do Things That Don’t Scale, July 2013
- Rahul Vohra, How Superhuman Built an Engine to Find Product/Market Fit, First Round Review, November 2018 — 22 per cent very disappointed at first, 33 per cent after narrowing the segment, 58 per cent within three quarters.
- US Department of Justice, Herfindahl–Hirschman Index — The index squares each share and sums the results.