Ninety days is enough to answer three questions and not one more: is the urgency real, can you assemble the supply, and does the spread survive contact with delivery. Everything else is next quarter's problem.
Days 0–10: pick the thimble and name the line item
Every canonical success launched inside a constraint tight enough to make liquidity achievable at trivial absolute scale — Etsy's handmade rule, Poshmark's six months on fashion, Uber black cars in one city. The constraint is not modesty; it is what keeps the few suppliers you have busy.
Write down, before anything else: three named buyers, the specific line item the money comes out of, and the person whose budget it is. If the money comes out of a headcount line the buyer controls, you are priced against a salary they can look up and you should say so out loud now rather than discover it in month four.
Be honest that the record does not tell you who signs. Epoch's reporting shows lab deals originating with researchers — one neolab researcher cited seeing $300K–$500K deals — which implies research-led budget authority rather than a central procurement function. No public source names a title. That is a gap you close with three conversations, and it is the highest-value use of week one.
Days 10–25: the cheapest test of buyer urgency
Urgency does not show up as enthusiasm. It shows up as a date and a purchase order.
The test: ask for a paid pilot, at a price that hurts slightly, with a delivery date the buyer chooses. Not a letter of intent. Not a free design-partner arrangement. Not a discount for being first. A buyer with genuine urgency will pay to move a date forward; a buyer without it will offer you a logo and a testimonial.
Two supporting reads on the same call. What breaks if this is not done by the date? If nothing breaks, the budget is discretionary and will be the first thing cut. And what did you try before calling us? If the answer is "we posted the role three months ago," the capability gap is real. If it is "we already have a team, we just want capacity," you are bidding against a marginal hire.
Do not build supply until you have this. Version One's instruction to seed supply first is right for a marketplace with an existing buyer side; it is wrong as a reason to recruit a thousand people against a hypothesis. Recruit the number the pilot consumes, plus a small buffer for the rework you have not yet learned to price.
Days 25–45: what a first contract should look like
The first contract sets the shape of every subsequent one, and it is the only moment you have leverage on terms rather than price. Eight things to get right:
- Fixed scope, milestone-billed. Projects with budgets that end are worse than functions with budgets that renew, but a first contract should be a project so both sides can stop cleanly.
- A deposit or an advance. The cheapest capital there is, and nearly impossible to retrofit later.
- Net-30, not net-60. At a 25% gross margin that difference is about 6 points of annual revenue in trapped cash. Terms matter more than margin — see You pay weekly, they pay in sixty days.
- Named acceptance criteria and an acceptance clock. The clock starts on delivery, and silence past N days is acceptance. Without this, "net-60 on acceptance" is net-90.
- A defined rework allowance. Say the number — a percentage of delivered volume, re-delivered within a stated window. Uncapped rework is an uncapped margin leak, and it is where first contracts actually lose money.
- A liability cap at fees for v1. You will lose the carve-outs for confidentiality, IP and data breach eventually. Do not concede them on the first deal in exchange for nothing.
- No exclusivity given away. Exclusivity is worth 4–5x when sold deliberately (Epoch AI); giving it away as a goodwill gesture is the single most expensive thing in this list.
- IP drafted as assignment with a fallback licence. A single US-form clause silently converts into a five-year, India-only licence in India and an exploitation right in Germany. See Where the supply can legally live.
And one thing to decide, not drift into: whether you are the principal. If you set the spec, direct the work and carry rework risk, you book gross, the crowd payments are your cost of revenue, and you own the funding gap. That is usually right, and it must be consistent across the audit file, the MSA and the worker agreement. See Pricing and the contract and The law is about to arrive.
Days 45–90: deliver, and instrument the delivery
Three operational things worth doing from the first batch rather than the tenth.
Inject gold-standard tasks from day one, at 5–15% of volume. Retrofitting a quality baseline onto a running programme is nearly impossible, and the contamination base rate is high: 33–46% of Mechanical Turk workers used LLMs on a summarisation task in 2023 (Veselovsky, Ribeiro & West). Scale's Google programme ran with spam problems for eleven months before they surfaced (Inc).
Measure gross margin after rework and payout costs, not before. The number in your model is not the number in your bank account. Payout rails, FX spread and re-delivered work are all cost of revenue.
Track the cohort, not the headcount. The metric that moves first when a supply pool starts to fail is inter-annotator agreement, not attrition.
Day 90: the seven numbers
| Metric | Continue | Worry | Stop |
|---|---|---|---|
| Second order from the first buyer, no discount | Signed | Verbal | Nothing |
| Median weekly hours a contributor can fill | 20+ | 10–20 | <10 |
| 8-week retention of the first supply cohort | >60% | 40–60% | <40% |
| Gross margin after rework and payout costs | >40% | 25–40% | <25% with no written path to 40% |
| Largest buyer as a share of booked revenue | <25% and falling | 25–50% | >50% with no second contract in diligence |
| Actual days sales outstanding vs contracted | At terms | +15 days | +30 days or an unresolved acceptance dispute |
| Rework / rejected-work rate on gold tasks | <5% | 5–15% | >15% |
The one that decides everything is the first row. A second order at full price from a buyer who has now seen your delivery is the only evidence in the list that cannot be manufactured. Everything above it can be bought with a subsidy, and subsidies are a happiness crutch — a competitor copies a subsidy in a week.
The second row is the one operators skip and Andrew Chen's whole post-mortem is about: if a contributor cannot fill a week, your supply cost is a recurring re-acquisition cost. See Which side you build first and Building the supply side.
Two failure patterns that look like success at day 90. A large first contract from a single buyer is a concentration problem being mistaken for product-market fit — above 25% of revenue you are a division of that customer (One customer is a binary event). And high gross margin before rework is the standard first-contract illusion; the sector's best-funded company runs 27–33% after everything, and it has 30,000 contributors and years of tooling.
What ninety days cannot tell you
It cannot tell you whether the position holds. Leakage is invisible early — the first buyer and the first suppliers have not yet had the second transaction, and nobody defects on deal one. It cannot tell you whether the budget survives a downturn; circumvention and price pressure both rise exactly when the buyer gets cost-sensitive. And it cannot tell you whether models eat the work, because the tier you are serving on day 90 is not necessarily the tier that exists in a year — xAI cut 500 of ~1,500 generalist annotators while pledging a tenfold larger specialist team.
Those are 12- and 24-month questions. When to stop gives the numbers at which to answer them by stopping.