How to start a startup in the first 90 days

Starting a startup is a sequence of tests, not a burst of paperwork or a three-month building sprint. In the first 90 days, your job is to find a painful problem, identify the people who feel it sharply, test the smallest possible solution, and watch whether anyone changes her behavior or pays. Incorporation, banking, and founder equity matter, but they should support real commitments rather than substitute for them.

The order matters. If you build before you understand the problem, your product teaches you little. If you incorporate before you know who owns what, you create cleanup work. If you fundraise before you can show why a customer cares, investor meetings become an expensive way to hear the questions you should have asked users.

Treat 90 days as a decision period, not a promise that every company should launch on day 90. A medical device, a financial product, and a neighborhood service cannot test risk in the same way. They can all move from assumption to evidence in the same order.

Days 1-10: Write down a problem you can observe

A startup begins with a specific person struggling with a specific task, not with a product category. "Software for small businesses" gives you nothing to test. "Independent physical therapists lose paid appointments when patients fail to complete intake forms before a first visit" gives you a person, an event, an existing cost, and a place to look.

Steve Blank makes a distinction founders routinely blur: an established company executes a known business model, while a startup searches for one. That changes your first task. You are not writing a smaller version of an operating plan. You are recording guesses so reality can disprove them cheaply.

Write a one-page problem brief with five lines:

• The customer is a narrowly described person with purchasing or adoption power.
• The difficult event happens in a named situation, not "often."
• Today she uses a workaround, competitor, employee, spreadsheet, or does nothing.
• The consequence costs time, money, risk, lost revenue, or status.
• You will know the problem is serious when she makes a concrete commitment.

That last line prevents wishful thinking. A commitment can be an introduction to a decision maker, access to a workflow, a scheduled second meeting, a deposit, a signed letter of intent with real terms, or payment. A compliment is not a commitment. An email address entered to be polite is weak evidence.

Choose a problem you can reach. A huge market populated by people who will not speak to you is worse in month one than a smaller entry point where you can conduct ten honest conversations. You are choosing where to begin learning, not declaring the permanent size of the company.

Do not spend these days naming the company, designing a logo, or calculating a top-down market from a research report. Check whether the problem could support a business, of course: who controls a budget, what they already spend, whether regulation blocks the obvious solution, and whether the pain recurs. But keep every claim labeled as an assumption until a customer supplies evidence.

Your day-10 output is not certainty. It is a falsifiable statement: "We believe [specific customer] will [specific commitment] to solve [specific recurring problem] because [observed consequence]." If no reachable customer recognizes that sentence, change the customer or problem before you build.

Days 11-20: Interview behavior, not opinions

Customer interviews should reconstruct what happened the last time the problem occurred. Future-tense questions invite fantasy. "Would you use an app that...?" tells the participant what answer you want and asks her to predict behavior in a situation with no cost.

Y Combinator's guidance on talking to users recommends questions about the hardest part of the task, the last encounter with the problem, why it was hard, what the person tried, and what she disliked about existing solutions. That advice is sound because every question asks for a memory or a fact. I would add two questions founders often avoid: who approved the current spending, and what happened when the customer did nothing?

Use this interview sequence and keep your pitch out of it:

1. "Tell me about the last time you had to [do the task]."
2. "Walk me through what happened, starting with the trigger."
3. "Where did time, money, or attention get lost?"
4. "What have you tried, and what did each option cost?"
5. "Who else had to approve or use the solution?"

Ask for artifacts when appropriate: the blank form, anonymized schedule, current invoice, spreadsheet columns, or screenshot with sensitive information removed. People summarize messy work into neat stories. An artifact exposes the extra handoff or workaround that memory skips.

Recruit interviewees from the narrow customer definition, not from a general founder audience and not only from close friends. Start with people one introduction away, professional groups, former colleagues, local businesses, and careful direct outreach. A short message works: name the workflow you are researching, state why you chose them, ask for 20 minutes, and make clear that you are not selling in the interview.

After each conversation, record exact facts under the same headings: trigger, current process, frequency, consequence, alternatives, buyer, and promised follow-up. Do not average stories too soon. You need to notice whether the same type of person reaches the same painful moment for the same reason.

Ten conversations are not a scientific sample, and twenty are not magic. Stop interviewing long enough to test a solution when you can predict important parts of the next conversation, several qualified people describe the same costly event without prompting, and at least a few agree to a concrete next action. Continue discovery alongside building. If every conversation reveals a different problem, your segment is still too broad.

Fear that someone will steal the idea makes many founders ask for an NDA before a basic interview. That usually blocks the conversation while protecting very little. Share the problem and the workflow; hold back genuine trade secrets, customer data, source code, and patent-sensitive technical details until counsel tells you how to handle them. If your advantage disappears when another person hears one sentence, you do not yet have much of an advantage.

Days 21-35: Build the smallest test that can fail

The smallest testable version is the least work that exposes your riskiest assumption to customer behavior. It may be software, but it may also be a manual service, clickable prototype, paid workshop, landing page with a real purchase path, or a report you assemble by hand. "MVP" has become an excuse for a small pile of features. A test needs a question and a result that could make you stop.

Name the assumption before choosing the artifact. If you doubt that clinic managers will share scheduling data, a polished dashboard tests the wrong thing; ask for a limited export under clear handling terms. If you doubt that a buyer will pay $200, a free beta cannot answer the pricing question. If the technical risk is whether a model can classify a difficult document reliably, a manual concierge service hides the risk and gives false comfort.

Set the decision rule in advance. For example: "We will offer five qualified customers a paid, manually delivered weekly report at $200 per month. If two pay and use the report in a decision within three weeks, we will automate the slowest production step. If none pays after we address the same two objections, we will revisit the customer and consequence."

The numbers in that example are not a universal benchmark. Their purpose is to force a decision before enthusiasm edits the result. Your threshold should fit the sales cycle, price, risk, and size of the reachable pool.

Keep the version honest. If a human performs a task behind the interface, tell the user enough to set accurate expectations about timing and data handling. Do not fake automation in a regulated or safety-sensitive decision. You can test demand without pretending that an unfinished system is finished.

Collect less data than your imagined full product needs. The Federal Trade Commission advises businesses to keep sensitive personal information only when they have a legitimate business need and to limit access. That is practical product advice, not paperwork for later. Every extra field creates security work, user hesitation, and a possible breach consequence before it has earned its place.

Your day-35 artifact should let a customer complete the core job once. Remove onboarding tours, settings for hypothetical segments, broad integrations, and scale work that no current usage requires. Keep authentication, consent, safety, accessibility, and payment controls that the test genuinely needs. "Small" never means careless with another person's money or information.

Days 36-50: Find first users by hand

Your first users should come through a channel you can work personally, because you need conversation more than reach. Paid acquisition can produce traffic, but it often hides why a stranger ignored you. Founder-led outreach is slow and revealing: you choose each prospect, hear the objection, observe onboarding, and see whether the promised result arrives.

Return first to interviewees who experienced the problem most sharply. Do not send a mass announcement. Write an individual note that connects their earlier account to the test: "You told me reconciliations fail when contractor receipts arrive after month end. I built a manual weekly check that flags missing receipts before close. I am taking on three teams this month at $150 each. Would you like to see the exact output?"

That message works only if each claim is true. It names the observed event, the result, the constrained test, and the price. It does not ask the customer to interpret a vague product vision.

Charge earlier than feels comfortable when payment is part of the business model. Free use proves that someone will accept a free thing. Payment tests priority, trust, procurement, and whether the outcome is worth more than the price. You can offer a narrow paid pilot or a cancellable first month without locking yourself into a discount that later customers will expect.

Some products cannot charge during the first test. A consumer network may need density, an enterprise deployment may require a long security review, and a clinical tool may need approval before use. Choose the strongest commitment legally and ethically available: repeated use, a signed design partnership with named obligations, integration work by the customer's team, or a scheduled purchasing review. Write down why it predicts eventual payment.

Track a short funnel with actual people: qualified contacts, conversations, offers, accepted tests, completed first use, repeated use, and payments. A large signup count can disguise a dead product if nobody reaches the promised outcome. At this stage, one customer's path from outreach to repeat use teaches more than a traffic chart detached from names and conversations.

Do not automate outreach until you have earned a repeatable sentence. If ten carefully chosen prospects misunderstand the offer, sending it to ten thousand people makes the confusion more expensive. Rewrite the customer, trigger, result, or offer and try again.

Days 51-65: Let usage overrule your roadmap

Repeated behavior should decide what you change after launch. Users will request features, but a request can mean several things: the core job is blocked, the user belongs to another segment, the product lacks trust, or the requester is being helpful. Trace the request back to the moment where work stopped.

Maintain an evidence log with one row per observation. Record the date, customer segment, expected behavior, observed behavior, consequence, and decision. Separate facts such as "three invited teammates never opened the email" from interpretations such as "teams do not collaborate." The fact may point to the invitation copy, permissions, timing, or the wrong buyer.

Review the log twice a week. Fix failures that prevent the promised outcome for the target customer. Defer preferences that do not affect that outcome. When several users invent the same workaround, study it before replacing it; the workaround may reveal the workflow your design missed.

Measure retention according to the natural frequency of the job. Daily retention means little for quarterly tax preparation. For a weekly planning product, ask whether the user returns in the next relevant week and completes the core action. For a one-time formation service, retention may be the wrong measure, so track completion, referral, use of the resulting documents, and willingness to buy the next adjacent service.

Decide among three moves. Persist when the target customer reaches the promised result and returns or pays, even if onboarding is ugly. Revise the solution when the problem is strong but the test fails at a specific step. Change the customer or problem when interviews and behavior show weak consequences, no budget, or no urgency.

Do not call every edit a pivot. A new button label is an iteration. Selling the same capability to compliance teams instead of individual consumers changes the customer and often the business model. The distinction matters because a genuine change invalidates old assumptions about sales cycles, pricing, product requirements, and market access. Write a new problem brief rather than dragging old evidence into a new thesis.

By day 65, aim for a small number of users whose behavior you understand end to end. You should be able to explain how they found you, what caused them to try, where they hesitated, what result they received, and what they did next. If you cannot, more features will not repair the missing evidence.

Days 66-75: Make founder commitments explicit

Cofounder agreements should capture the working relationship before memory and incentives diverge. A long friendship, shared employer, or promising weekend project does not answer who owns prior work, who works full time, how decisions get made, or what happens when one founder leaves.

Run a direct founder conversation before issuing equity. Cover expected time commitment, cash contributions, compensation, roles, decision rights, fundraising appetite, personal constraints, and the circumstances that would cause either person to stop. Discuss the uncomfortable cases while both of you still expect the company to succeed.

Equity should reflect the future work and risk required to build the company, not only who mentioned the idea first. Investors and future hires will inspect the capitalization table, but that is not the reason to use vesting. Vesting prevents a founder who leaves early from retaining the same ownership as the person who spends years doing the work.

Put intellectual property assignments, confidentiality obligations, stock purchase terms, and vesting in signed documents prepared for your jurisdiction and situation. Confirm that a current or former employer, university, client, or contractor does not own work that the startup needs. Paying a contractor does not automatically answer every ownership question; the contract must address the relevant rights.

Do not split equity through an informal message and promise to document it after funding. Cleanup may require consents, amended filings, tax analysis, and a difficult renegotiation when the company has more value. A startup lawyer costs money, but targeted advice on formation and founder stock is usually cheaper than repairing an ambiguous cap table.

If you are an international founder in the United States, keep company formation and immigration authorization separate in your mind. Owning a US company does not by itself authorize you to work for it, and your visa facts can affect role, payroll, travel, and timing. Use qualified immigration counsel for your facts before treating incorporation as an employment answer.

Days 76-85: Form the company when commitments require it

Form a legal entity when you are taking money, signing meaningful contracts, issuing founder equity, hiring, accepting material liability, or preparing for institutional investment. You may need one earlier in a regulated field or to protect intellectual property. You may wait during low-risk interviews and prototypes, but do not take customer risk or mix substantial business money with personal accounts to preserve a tidy timeline.

In the United States, entity choice depends on liability, taxes, ownership, financing plans, and where you operate. The Small Business Administration warns that structure affects taxes, fundraising, paperwork, and personal liability. A venture-backed company commonly uses a Delaware C corporation because investors and their lawyers know that structure. A consulting business, local service, or self-funded company may prefer an LLC. Habit is not analysis, so ask a startup attorney and tax professional about your actual plan.

Once you form, complete the operational chain rather than stopping at a certificate:

• Approve formation and founder actions through the required board and stockholder documents.
• Issue founder stock correctly, record purchases, and maintain a capitalization table.
• Get an EIN directly from the IRS, which issues it free, then open a company bank account.
• Start bookkeeping, save receipts, and separate company income and expenses immediately.
• Calendar state filings, taxes, licenses, registered-agent duties, payroll, and insurance reviews.

Restricted founder stock creates a deadline that founders miss because it arrives when everything else feels urgent. IRS Form 15620 says a Section 83(b) election must be filed no later than 30 days after the property transfer. Ask your tax adviser whether the election applies, file it correctly and on time if advised, give required copies, and retain proof. Do not wait for the first tax return or funding round.

Check current requirements rather than copying a formation checklist from an old blog post. FinCEN changed its Corporate Transparency Act rule in March 2025 so entities created in the United States and their beneficial owners are exempt from federal beneficial ownership reporting; some foreign entities registered in the United States still must report. State ownership reports, bank verification, tax filings, and future federal changes are separate issues.

Delaware also does not forget a corporation that has not launched. Its Division of Corporations says active domestic corporations must file the prior year's annual report and franchise tax by March 1. Calendar the obligation when you form, not when revenue arrives.

Before choosing a name, search more than domains and state entity records. The US Patent and Trademark Office tells applicants to search for conflicting marks in connection with the relevant goods and services. A state may let you register a company name that still creates trademark risk. Delay an expensive identity system, but investigate a name before printing it everywhere.

Days 86-90: Build a cash and decision cadence

A 13-week cash view and a weekly decision meeting are enough financial machinery for most new startups. Build the forecast from the bank balance, committed receipts, realistic collection dates, payroll, contractors, software, taxes, insurance, and one-time obligations. Keep hoped-for investment in a separate scenario until signed documents and transferred funds make it real.

Use cash dates, not accounting optimism. An invoice sent this month may arrive next month or later. An annual software contract may leave the bank today even though accounting recognizes the cost over time. Your near-term survival depends on when money moves.

Each week, review four things: cash and runway under the base case; customer evidence added; the largest blocked test; and one decision that changes the next week's work. Assign an owner and date to that decision. A meeting that merely reports activity will not protect the company from spending six weeks on an assumption nobody tested.

Set a monthly close even when transactions are few. Reconcile the bank and card accounts, classify expenses, review outstanding invoices, update the cap table after any equity action, and save formation and tax correspondence in one controlled location. Clean books make tax work, diligence, and founder decisions less painful. They also show whether your "cheap" experiment is quietly consuming cash.

Fundraising belongs in the first 90 days only when capital is necessary to run the next decisive test or when the company already has evidence that makes the process rational. Investor interest is not customer validation. A warm meeting can help you learn how the financing market sees the business, but a long raise can consume the founder time needed to create the evidence investors request.

Sisters gives women building companies a place to ask experienced peers about formation, customer development, fundraising, and the daily decisions that rarely fit a generic checklist. Use peer advice to expose questions, then use a lawyer, accountant, or immigration professional where an answer depends on your facts or creates a filing obligation.

Defer work that assumes success

Defer anything whose value depends on scale you have not earned, unless law, safety, or a signed customer commitment requires it. Premature polish feels productive because nobody can reject it. It also consumes the weeks when rejection would teach you the most.

Wait on a broad brand campaign, a large launch event, an office, executive titles, a complicated analytics stack, generalized automation, and hires for work the founders have never done. Delay a long investor deck if customer evidence is still vague. Do not build permissions for five future roles when one founder and three test users can operate safely with a simpler design.

Deferral does not mean neglect. Keep a parking-lot document with the request, who raised it, what evidence would activate it, and the likely cost. "Build an enterprise audit export after two qualified buyers make it a purchase condition" is a decision. "Enterprise later" is a hiding place.

Some work never belongs on the defer list. Address legal restrictions, security appropriate to the data, accessibility for the core experience, truthful claims, customer consent, tax deadlines, and safety before exposure grows. The cheapest test still has to respect the person taking part in it.

At day 90, you may have a small product, a manual service, a waiting list with deposits, or evidence that the original idea is wrong. The last outcome can save a year. Judge the period by the quality of decisions now available: a named customer, an observed problem, a tested offer, behavior you can explain, clean founder commitments, and enough financial control to run the next test. If those pieces are missing, another quarter of building will only make the guess more expensive.