How to start a business with AI: the whole first ninety days, in order
AI removes almost all of the typing and none of the risk. It will write your plan, build your site, price your work and draft every email — but it cannot tell you whether a stranger will hand you money, and that is the only question the first ninety days exist to answer. This is the order to do things in, week by week, with a gate at the end of each block that says whether you've earned the next one.
Ninety days is long enough to find out whether a business works and short enough that you can't hide in it. About 22% of new US businesses close inside their first year — that's roughly 600 a day, on the Bureau of Labor Statistics' own establishment data — and very few of them close because the founder couldn't write a plan. They close because nobody was ever going to pay that price for that thing, and it took eighteen months and a lot of building to find out.
AI changes the cost of the building, not the odds of the paying. Ten years ago the first ninety days were mostly production: writing copy, wrestling a website, formatting a plan, working out what a licence costs in your county. That work is now hours instead of weeks, which is genuinely enormous — but it means the risk has moved. When making things is cheap, the failure mode is making the wrong thing beautifully, fast. So the sequence below is built to spend the cheap thing (production) early and defend the expensive thing (your ninety days) with gates.
Read the gates first. Each block ends with one, and it's a pass/fail sentence, not a feeling. If a gate fails you don't move on — you go back and change the offer, the price or the customer. Founders who follow this and stop at day 14 because nobody would pay a deposit have had a very good fortnight. That's the point.
What AI does well, and what it can't do for you
Be precise about this and the whole ninety days gets easier. The split isn't "creative versus boring". It's whether the task has a verifiable right answer that exists somewhere in text, or whether it needs someone with skin in the game.
- AI is very good at · the blank page
- Turning a description into a first draft of anything — a landing page, five ad angles, a pricing table, a cold email, a job post. Not the final version. The version you can react to, which is the version that gets made.
- AI is very good at · boring completeness
- The forty things a new business must do that nobody remembers all of: the sales-tax permit, the city licence, the report due in ninety days. A checklist is a solved problem and a founder's memory isn't.
- AI is very good at · research with sources
- What competitors charge, what a market is worth, what a supplier's lead time is — gathered in an afternoon with links you can click and check. Check them. This is where confident nonsense hides.
- AI is very good at · arithmetic that has to be redone
- Runway, break-even, cost per customer, margin per job. The numbers that are only useful if they're current, and that a busy founder recalculates roughly once, in month one, and then never again.
- Only you can · pick the customer
- Every downstream decision — price, channel, copy, what you build — is determined by who you decided this is for. AI will happily optimise a business aimed at nobody in particular, forever.
- Only you can · ask a person for money
- The single highest-value act in the first fortnight, and the one nothing can do on your behalf. A card charge is the only research method that can't be polite to you.
- Only you can · deliver the first jobs
- Do the first ten yourself, badly if necessary. What you learn in someone's driveway — the objection, the upsell, the thing that took twice as long — is not in any dataset, and it's what you'll price on.
- Only you can · be accountable
- Your name is on the filing, the lease, the personal guarantee and the tax return. AI drafts; you sign. Treat anything it produces about law, tax or safety as a well-organised starting point to check, not advice.
The ninety days, block by block
Ticks are kept in this browser only — nothing is sent anywhere and there is no account. The steps marked only you are the ones no software can do on your behalf: a bank wants to meet you, a licence wants your signature, and five customers want a conversation.
Six blocks. Each one has a job, a division of labour, and a gate. The dates are a default, not a rule — a business with a licence to wait for will have a longer block four, and a consultancy with a first client already lined up can compress blocks one and two into a phone call. What shouldn't change is the order.
Days 1–7 · One sentence, five names
The job: stop having an idea and start having an offer. An idea is "mobile bike repair". An offer is "a 40-minute tune-up in your driveway, $60, anywhere inside the Denver city limits, Tuesday to Saturday".
Give AI: the market research — who else does this within ten miles, what they charge, what their reviews complain about, how many households are in the area. Ask for sources and open them. Also the one-page plan, which takes ten minutes and is worth doing once, mostly because writing down your assumed price and assumed volume makes them arguable.
Keep for yourself: the sentence, and the list. Write the offer as one sentence with a price in it. Then name five real people — not personas, people with phone numbers — who fit it. If you can't name five, you've picked a customer you don't have access to, and that's a much bigger problem than the price.
Gate: you can say the sentence out loud without hedging, and you have five names and a way to reach each of them this week.
Days 8–14 · Money before anything gets built
The job: find out whether the offer survives contact with a stranger's wallet. Nothing else in the ninety days is as informative and nothing else is as often skipped.
Give AI: a one-page waitlist or pre-order page — headline, offer, price, one form — and drafts of the messages that go to the five names and to whatever list, group or forum your customers already sit in. Also the question list for the conversations, so you're not improvising.
Keep for yourself: the asking. Have the five conversations and, at the end of each, ask for money at the real price: a deposit, a pre-order, a first booking. Watch what happens between "that sounds great" and the card coming out. That gap is the whole business.
Gate: at least one person who isn't a friend or a relative has paid or firmly committed money at your real price. Enthusiasm is not a pass. If nobody pays, change the price, the offer or the customer — in that order, because price is cheapest to change — and run the fortnight again. The three tests and what failing each looks like →
Days 15–21 · A site that takes the money
The job: somewhere a stranger can land, understand you in ten seconds and pay without speaking to you. That's it. Not a brand. Not five pages.
Give AI: the whole build — copy, layout, the booking or order form, the logo if you want one. Generating a site from a description is one of the things AI is unambiguously good at, and a first version taking payments in an afternoon is now normal.
Keep for yourself: connecting the payment processor to an account in the business's name, setting the actual price, and testing it. Pay yourself. On a phone, on mobile data, from the link exactly as a customer would receive it, then refund it. About half of the "my site doesn't convert" problems are a checkout that's broken on mobile and has never been tried.
Gate: someone who has never heard of your business can find the page, tell you what you sell and what it costs, and complete a real payment on their phone.
Days 22–30 · The legal and money floor
The job: the unglamorous layer that only hurts if it's missing. Do it after the first payment, not before — a business nobody wants doesn't need an entity, and a filing fee paid in week one is the most common wasted money in a failed start.
Give AI: the checklist for your state, in order, with the fee and the official portal for each step, plus the deadlines that come later — the annual report, the sales-tax filing, the licence renewal. Also the first budget: what a month costs including your own pay.
Keep for yourself: the filing. You file with the state; you get the EIN free from irs.gov in ten minutes; you open the bank account. Nothing should ever file on your behalf, and nothing should charge you for the EIN, which is free. Every state's fee, portal and recurring deadline →
Gate: a business bank account is open, money from customers goes into it and nowhere else, and you can say what a month costs to the nearest hundred dollars.
Days 31–60 · One channel, ten customers
The job: ten paying customers from one channel, run properly for four weeks. One. The instinct at this point is to do a bit of everything — some posts, some ads, some emails, a stall at a market — and the result is four channels each run at a tenth of the effort needed to learn anything from them.
Give AI: the volume work. Finding and qualifying leads, writing the ad variants, drafting the outreach, making the posts, drafting replies to every enquiry so none sits for a day.
Keep for yourself: the conversations that convert, the judgement about which replies are real, and the delivery. Also the counting: leads found, messages sent, replies, quotes, paid. Four numbers, written down weekly.
Gate: ten paying customers, and you can say what one cost you to get in money and in hours. If the channel produced under three customers in four weeks of real effort, it's the wrong channel for what you sell — change it, don't add to it. Which channel by business type, and the numbers to expect →
Days 61–90 · Raise the price, kill the losers, write the numbers down
The job: turn ten customers into a thing that repeats. This block is mostly subtraction.
Give AI: the analysis — what each channel actually returned, measured from payments rather than from what the ad platform claims; which costs haven't earned their place; the runway number, recomputed; follow-ups to every past customer.
Keep for yourself: three decisions. Put the price up on new work — almost every first price is too low, and the first ten customers have told you why. Stop the thing that didn't pay, entirely, this week. And ask every one of the ten what nearly stopped them buying; fix the top answer.
Gate: you can state, without opening anything, your cost per customer, your average order value, your gross margin per job and your runway in months. And at least one customer has bought twice.
The gates on one page
Print this bit. A gate is failed until it's passed, and the correct response to a failed gate is always to go backwards, never sideways into a new task that feels like progress.
| Days | Pass looks like | On a fail, go back to |
|---|---|---|
| 1–7 | One sentence with a price in it, and five named people you can reach this week. | The customer. You've chosen someone you have no route to. |
| 8–14 | A stranger has paid or committed money at the real price. | Price first, then offer, then customer. Re-run the fortnight. |
| 15–21 | A stranger can pay you on their phone in under two minutes. | The checkout. Test it yourself before blaming the traffic. |
| 22–30 | Business bank account open, all customer money landing in it, monthly cost known. | The budget. If you can't name the monthly number, you can't price. |
| 31–60 | Ten paying customers from one channel, cost per customer known. | The channel. Change it; don't add a second one. |
| 61–90 | Cost per customer, average order, gross margin and runway known from memory. One repeat customer. | Delivery. A business with no repeat customers has a product problem, not a marketing one. |
The two gates people argue with are day 14 and day 60, and they're the two that matter. Day 14 is uncomfortable because asking for money is uncomfortable. Day 60 is uncomfortable because "the channel didn't work" sounds like "the business didn't work" — it isn't. A florist failing at cold email has learned that florists don't buy flowers by cold email, which took four weeks and cost almost nothing.
Ninety days, for one real-shaped business
Northline Bike Repair is the example business that runs through these pages: one founder, a van, mobile bike repair in Denver. The figures below are this example's own — not benchmarks — and the shape is what matters. Notice that revenue in month one is $380 and that nothing has gone wrong.
| Days | What happened | The number that mattered |
|---|---|---|
| 1–7 | Six competitors priced (four shops, two mobile). Shops charge $75–$95 for a tune-up and take four days. The offer became: 40 minutes, in your driveway, $60. | 4-day wait — the thing to beat |
| 8–14 | Twelve conversations: eight cyclists from a Saturday group ride, four from a neighbourhood group. Nine said it sounded useful. Three paid a $20 deposit on the spot for a first Saturday. | 3 deposits · $60 |
| 15–21 | A one-page site with the service area, three prices and a booking form that takes a deposit. Tested from a phone in a car park, then refunded. | 1 test payment, refunded |
| 22–30 | LLC filed with Colorado, EIN from the IRS, business account opened, van insurance switched to commercial. First three deposits moved into the business account. | $1,410/month to run |
| 31–58 | One channel: local Facebook and Instagram ads inside a five-mile radius, four angles, $300 over three weeks. 214 visits, 4 paid bookings. | $75 per customer |
| 31–58 | Two of the first four ad customers referred a neighbour; one deposit-holder booked again. Customer ten arrived on day 47. | 10 customers · day 47 |
| 59–75 | Every customer asked what nearly stopped them. Six of ten said the same thing: they didn't know whether he'd come to their block. A service-area map went on the site. | 6 of 10 — same objection |
| 76–90 | Tune-up went from $60 to $66 on new bookings. Nobody objected. Two of the four ad angles were switched off. Month three closed at $2,180. | $2,180 · 22 customers to date |
Three things in that table are the actual lesson. The first is that the deposit on day 12 — sixty dollars, from three people — was worth more than any amount of research, because it converted an opinion into a fact. The second is that the objection that cost the most sales ("does he come to my block?") was invisible until ten people had been asked directly, and cost nothing to fix. The third is that at day 90 this business is turning over about $2,180 a month and is nowhere near paying a salary. That's normal, it's on schedule, and a founder who expected month three to look like a job is the one who quits in month five.
Where ninety days usually go
Building for six weeks because building is calm
Selling is exposing; building is soothing. AI makes building faster, which means it makes the avoidance more productive-looking. The tell is a founder in week seven with a beautiful site, a logo in three formats, a fourteen-day content calendar and zero conversations. If you've spent more days making things than talking to people, you are behind regardless of what exists.
Forming the company on day two
An entity is a container for a business's risk. Before there's a business it's a filing fee, an annual report and a deadline you'll miss. There are real exceptions — a lease to sign, a client who won't contract with an individual, a trade that legally requires it — and if one applies to you, form it now. Otherwise it belongs in block four, after somebody has paid.
Asking AI what your business should be
You'll get an answer, it'll be plausible, and it'll be the average of everything written on the internet about businesses like the one you described. Useful for a list of options to react to. Useless as a decision, because it has no information about your town, your skills, your tolerance for phone calls or the fact that your brother-in-law owns a warehouse.
Running four channels at a quarter strength
Covered above, and it's the most common reason day 60 arrives with six customers and no idea which thing produced them. Four weeks on one channel gives you a number. Four weeks on four channels gives you four anecdotes.
Not counting
Leads, replies, quotes, paid. Four numbers a week, in a notes app if that's all you'll keep up. Without them the day-90 gate is unpassable, every decision afterwards is a guess, and the single most useful sentence in a small business — "customers from this cost me $75 and are worth $180" — can never be said.
Day 91
If the gates passed, you have something rare: a small business with paying customers, a known cost to acquire one, a known margin and a known number of months of cash. That's not a big business, and the ninety days weren't supposed to produce one. What they produced is the right to make decisions with evidence instead of hope — whether to raise the price again, whether to add the second channel, whether to hire, whether to keep going at all.
If a gate didn't pass, the honest answer is that you now know something in ninety days that most people take two years and their savings to learn. Change the customer, or the offer, or the price, and run it again. The site, the checklist and the entity are all still there; almost none of the ninety days is wasted on a second attempt, which is exactly why the cheap-production era is a good time to try.
Common questions
Can AI actually start a business for me?
It can do the production — research, copy, a live website, a legal checklist, pricing, ads, drafted replies, the nightly arithmetic — and that's most of the typing. It cannot decide who your customer is, ask a person for money, deliver the first jobs or sign anything. Those four are the ones that determine whether it works, so the honest description is that AI compresses the ninety days rather than replaces them.
How much money do I need to start?
For a service business run from home, the floor is the state filing fee if you form an entity ($35–$520 depending on the state), a domain, and enough test budget for one channel — a few hundred dollars is a real test. Stock, equipment, a vehicle or a lease change the answer completely. The number that matters more is what a month costs to run once you've started, because that's what runway is measured against.
Should I form an LLC before I start selling?
Usually not. Take the first payments as a sole proprietor while you're finding out whether anyone wants this, and form the entity in block four once someone has paid. Form it earlier if you're signing a lease, hiring, taking on real physical risk, or dealing with a client who won't contract with an individual.
What if nobody pays in the first two weeks?
That's a result, not a failure, and it's the cheapest one you'll ever get. Change the price first, because it's the easiest thing to change and the most common thing that's wrong. Then the offer — same customer, different thing. Then the customer, which is the biggest change and needs a fresh set of five names. Give each version a fortnight.
How long until it pays me a salary?
Longer than ninety days for almost every business that isn't consulting. In the worked example above the business turns over about $2,180 in month three against $1,410 of costs, which is a real business and not yet a job. Plan your own living costs on the assumption that the business contributes little for the first six months, and put your intended pay into the cost side so the runway number tells you the truth.
Do I need a business plan?
A one-page one, once, in the first week — mostly because writing down your assumed price, assumed volume and assumed cost makes them arguable and lets you check them against reality in week two. A thirty-page plan is for a bank or an investor who asked for one, and writing it before you have customers is a way of spending a fortnight without leaving the house.
Which parts of AI output do I have to check?
Anything with a number, a deadline or a legal consequence. Fees, filing dates, licence requirements and tax rules change and models are confidently out of date; open the state's own portal and the IRS page before you act. Research claims should come with links you actually click. Copy and drafts are lower risk — the worst case is that they're bland.
Is ninety days realistic for a business that needs a licence or premises?
The sequence holds; block four stretches. A food business waiting on a health inspection, a trade waiting on a state licence, or anyone signing a lease should start that clock in week one in parallel — it's the one thing worth doing early, because it's the one thing you can't compress. Everything else stays in order behind the day-14 gate.
The ninety days, with the production already done.
Describe your business once. The website goes live on its own address ready to take card payments, the legal checklist for your state fills in, and the overnight team starts on the leads, the ads and the numbers. The gates are still yours. Free to start, no card.
Start free →