पाठशाला Pathshala · उत्पाद Utpād, The product · Lesson 02 · Start

Concierge and Wizard-of-Oz MVPs: selling before you build

Deliver the service by hand to ten paying customers, write down every step you take, and the software specification writes itself. Then watch the founder hours, because the hand-run version has a ceiling and the ceiling is a date.

Pathshala, The Founder Library · 11 October 2026 · 8 min read

The cheapest version of a software product is a person pretending to be one. Before Zappos had a warehouse it had a founder walking into a shoe shop in Sunnyvale with a camera. Before Dunzo had an app it had a man on a motorbike reading a WhatsApp group. Neither had built anything, and both knew more about their customers after a month than most teams know after a release.

This lesson covers the two forms of hand-run first version, how to set one up with ten paying customers, what to write down while running it, the ceiling at which it breaks and how to turn the log into a specification. It ends with the thing a concierge test cannot tell you, which is whether the business works once the hands are replaced.

Two ways to do the job by hand

A concierge MVP delivers the outcome the product promises, personally, and the customer knows it. Manuel Rosso’s meal-planning company Food on the Table began that way: his account of users two to twenty describes a first meeting face to face, then a phone call, then recipes and grocery lists sent by email, with Google Apps standing in for dynamic web pages. The team’s rule was to focus on the experience rather than the code and to write code only when it would make learning cheaper. The sentence that justifies the method is his: if you cannot get people to adopt the idea with high-touch personal service they are not going to buy into a cold web page.

A Wizard-of-Oz MVP delivers the same outcome by hand behind an interface that looks automated. Nick Swinmurn started Zappos in 1999 by offering a shoe shop a deal: he would photograph its stock and put it online, and if anyone bought a pair he would buy it from the shop at full price. The website looked like a retailer. Behind it was a man driving to a shop. The assumption under test was whether people would buy shoes without trying them on, and no amount of software would have answered it faster.

The difference is what the customer believes. Concierge is honest about the human and therefore learns more, because customers talk to a person in ways they do not talk to a screen. Wizard of Oz tests whether the automated experience is wanted at the price, and is the right form when the customer’s willingness to use a self-serve product is itself the risky assumption. Choose by asking which belief you need to test: that the outcome is wanted, or that it is wanted from a machine.

Why the hand-run version teaches more than the built one

Software encodes rules. A team writing software before it has done the job has to guess the rules, and the guesses become the spec. A team that does the job by hand for ten customers discovers the rules by being forced to make them: what to do when the retailer has two names for the same product, when the diabetic customer’s doctor changes the diet mid-month, when the pickup address is a landmark rather than a street. Every improvisation is a requirement nobody would have written down in advance, and it arrives with its frequency attached. By week four the founders know which exceptions happen daily and which happened once.

Paul Graham’s Do Things That Don’t Scale describes the same effect from the other side. The Stripe founders, when someone agreed to try the product, asked for the laptop and set it up on the spot, which YC still calls a Collison installation. Airbnb’s founders went door to door in New York helping hosts improve their listings. In each case the unscalable act was also the research: the founder watching the customer’s face at the moment of use.

An Indian case, with its ending

Dunzo began in Bengaluru in 2014 as a WhatsApp group in which friends asked Kabeer Biswas to run errands, and he ran them, on a bike: collecting laundry, delivering cold drinks, getting a clock repaired. Strangers started posting in the group. The company was incorporated in January 2015, co-founders joined that September and the app replaced WhatsApp in February 2016; by April, Biswas told Scroll, traffic was doubling roughly every three weeks. That is a concierge MVP run for over a year before a line of consumer-facing code, and it answered the question it was built to answer: people in a congested city will pay a stranger to do their errands.

The ending belongs in the lesson too. Dunzo raised more than $450 million over the following years and went offline in January 2025 after a long decline. A hand-run test proves that demand exists. It says nothing about whether the demand can be served at a margin once the founder on the bike is replaced by four thousand riders and the customer’s willingness to pay meets the rider’s cost per task. The concierge phase answers the first question. The second needs the [arithmetic of unit economics](/library/cac-ltv-and-payback-the-three-numbers), and it needs it before the money arrives rather than after.

How to run one: ten customers, real money, a log

Ten customers, recruited by hand. Not a waitlist and not friends. Ten people in the segment you intend to serve, found the way the customer interview lesson finds them, each of whom has the problem this month. Rosso’s early users came through relationships built during discovery and through a few mavens who recommended others; yours will too.

A price, paid in advance. Charge what the software will eventually charge, or more. A customer paying ₹4,000 a month on UPI behaves differently from one on a free pilot: they call when the service fails, they tell you what they would stop paying for, and their continued payment is the measurement. Free pilots produce politeness.

A log of every step. A shared sheet with one row per task: what the customer asked, what you did, how long it took, what you had to decide, what went wrong. This is the most important artefact of the whole phase and the one most founders skip because they are busy delivering. The log is the specification. Without it you will build what you remember, and you will remember the dramatic cases rather than the frequent ones.

A rule book. Each time you improvise, write the rule you followed in a second sheet. When you notice the same rule applied five times it is a feature. When a rule has one entry after a month it is not.

A date to stop. Hand delivery has a ceiling and the ceiling is arithmetic. The figure below computes it.

A worked example: books for D2C sellers

Two founders in Mumbai believe small D2C brands need bookkeeping software that understands marketplace settlements. Instead of building it they offer to do the books by hand for ten brands at ₹6,000 a month. Each brand costs about three hours a week: pulling settlement reports from the payment gateway and the marketplaces, matching them to orders, entering returns, preparing the GST summary. Thirty founder hours a week for ₹60,000 a month. At ₹500 an hour of founder time the service costs more than it earns, which is fine; the margin is not the point yet.

After six weeks the log shows where the hours went: 70 per cent of the time was matching settlement lines to orders, because every marketplace reports in a different shape and deducts fees in a different place. 20 per cent was returns. 10 per cent was everything else, including the GST summary the founders assumed was the product. The specification for version one is now obvious and it is one workflow: ingest the settlement files, match them to orders, flag the unmatched. The rule book has forty rules for matching and eleven for returns. The ceiling, at thirty hours a week and three hours a customer, is ten brands, and the founders are at it, so the code starts now, and it starts with matching.

Do the job by hand until it hurts. Where it hurts is the specification.

The three signals that it is time to write code

First, the ceiling: the founder hours needed exceed the founder hours available and customers are waiting. Second, repetition: the log shows the same step performed the same way more than twenty times with no new rules added in a fortnight, which means the rules have stabilised enough to encode. Third, error: a step done by hand has started producing mistakes customers notice, because tired people doing repetitive work make errors software does not. Any one signal is reason to automate that step. All three is reason to stop taking new customers until it is automated.

Automate in the order the log dictates, by hours consumed, not by what is most interesting to build. The step that ate 70 per cent of the time is the product. The GST summary that ate 10 per cent is a feature for month four, and it may turn out that a founder doing it by hand for fifty customers is cheaper than building it at all.

One line should not be crossed. Wizard of Oz has a limit. Do not pretend to automation where the pretence has consequences: a lending decision presented as an algorithm when a founder is eyeballing bank statements, a health recommendation presented as clinical software when it is a founder with a spreadsheet, a delivery promise the human behind the screen cannot keep. If a customer asks whether a person is involved, say yes. Personal data handled by hand is still personal data under India’s data protection law, and a sheet of customers’ bank statements on a founder’s laptop is a liability the software version would not have.

A weekly ritual for the concierge phase

Every Friday afternoon for the length of the phase: read the week’s rows in the log and total the hours by step. Update the rule book and mark any rule that has crossed five uses. Recompute the ceiling with this week’s minutes per customer; it falls as you get faster and rises as customers ask for more. Ask each paying customer one question, which is what they would stop paying for first. Then decide one thing: take two more customers, automate one step, or raise the price. When the ceiling is within four weeks of the current count, the Friday meeting becomes a planning meeting for version one, and the log is the agenda.

The examples are from the public record and the sources are below. The worked figures are illustrative, and the ₹500 an hour shadow wage for founder time is a convention of the figure, not a benchmark.