Nobody kills one of these on a bad feeling. They kill it on a number they wrote down eighteen months earlier, or they do not kill it at all.
Every threshold below arrives attached to a reason it does not count: the concentration is temporary because the second contract is nearly signed; the margin is low because this quarter had unusual rework; utilisation is bad because the buyer paused a project. Sometimes that is true. Set the thresholds now, while it is cheap to be honest, and require a written case to override one.
Buyer concentration
Stop or restructure when a single customer exceeds 25% of revenue for two consecutive quarters with no second contract in signed diligence. Treat anything above 15% as a binary event, not a growth line.
The base rate is Appen: a US$4.3B peak with 80% of revenue in five clients, Google terminated in January 2024, shares down 40–41% in a day and 97% from peak. Its top five are still 74.3% — concentration does not necessarily fall as you shrink. TaskUs discloses Meta at 26% and trades near 0.61x price/sales. One layer down, Sama issued redundancy notices to 1,108 Nairobi employees in April 2026 after Meta terminated one contract.
Two amplifiers that should move the threshold down rather than up.
If your buyers compete, neutrality is the product. Meta's $14.3B for 49% of Scale AI was simultaneously the largest financing and the largest customer-loss event in the sector's history: within weeks Google cut ties, OpenAI departed, Microsoft pulled back, and Scale cut 200 FTEs and ~500 contractor relationships. Equity, a board seat or exclusivity from one buyer in a competitive set is a concentration event whatever the percentage says.
The in-housing clock is legible. A buyer in-houses when your function becomes (a) large enough to matter in their P&L, (b) strategic enough to differentiate, and (c) legible enough to hire for. Track all three per customer; when all three go true you have roughly one renewal left. OpenAI's Project Mercury hired 100+ ex-bankers directly at $150/hour; xAI cut 500 generalist annotators while building a specialist team in-house.
Gross margin
Stop when gross margin measured after rework, payout fees and QC overhead falls below 25% with no written path to 40% inside four quarters. Below ~40% the revenue multiple is capped near 1.6x, permanently — Accenture, the best-run services firm on earth, trades at 1.56x on a 32% gross margin.
Three common ways this gets faked. Measuring before rework, which is where first contracts actually lose money. Measuring before payout costs — FX spread times payout frequency is cost of revenue. And measuring before the gold-standard overhead of 5–15% of task volume, which is not optional if you sell verified human work.
Two calibration points. Mercor's leaked gross margin was 27% in 2025 rising to 33% in Q2 2026 — the best-funded company in the sector, with 30,000+ vetted contributors and years of tooling. And high gross margin alone does not save you: Fiverr runs 82.0% and trades at 0.09x EV/revenue, because volume is shrinking. The software multiple is a conjunction — margin and recurring revenue and NRR above 100% and fragmented buyers and volume growth. See What the public market pays for labour.
Leakage
Nobody can measure this. Upwork, twenty years in with full payment and messaging telemetry, states in its 10-K that circumvention losses are "difficult or impossible to measure." So use proxies, and set the thresholds on the proxies.
Stop when all three of these are true at once: fewer than half of buyers who transact once transact again within one purchase cycle; the value of a single relationship exceeds roughly a month of your revenue; and the only anti-leakage mechanisms you actually hold are ratings, escrow and a non-circumvention clause.
That is the formula stated operationally: leakage ≈ value per relationship ÷ (frequency × switching friction). One-shot, high-ticket, low-friction is the maximum-leakage corner and no drafting fixes it. Only two mechanisms survive a sophisticated buyer — liability transfer and workflow embedding. Concentrix's largest-client relationships average 16 years because the client's process lives inside Concentrix; Robert Half's filing calls its staffing contracts "generally terminable on short notice and without penalty."
A faster proxy: raise the fee five points on a subset and watch one quarter. If volume moves measurably, your hold is priced and you now know the price. Note Upwork's asymmetry — circumvention rises in a downturn, exactly when your rate is under strain. See Getting cut out.
Supply utilisation
Stop when the median active contributor cannot fill 20 hours a week for eight consecutive weeks at a rate they will accept, and 8-week retention of a new cohort is below 40%. Either alone is a warning; together they are the failure loop already running.
This kills more of these companies than demand does. Rideshare worked because a driver could fill 20–50 hours a week; the "Uber for X" cohort could not, paid the same $300+ per supplier acquisition cost, and never dug out. The end state is Mechanical Turk: 500,000 registered workers by 2011, about 2,000 active simultaneously by 2018, median $2/hr.
Watch the leading indicator. Attrition moves last; what moves first is inter-annotator agreement — the contributors with alternatives leave first, so median quality falls before headcount does. The wage record is the context: Mercor moved Meta project workers to $16/hr from $21; Appen disclosed crowd NPS falling from 33 to 22. See Building the supply side.
Regulatory events
Most regulatory news is not a kill signal. Federal deregulation in particular is not: Field Assistance Bulletin 2025-1 stopped DOL enforcement of the 2024 independent-contractor rule, but the same bulletin preserves it "for purposes of private litigation," and a friendly federal rule retires zero state ABC tests.
These five are different.
- An ABC prong-B finding against you or a close comparator in California, Massachusetts or New Jersey. Prong B — work outside the usual course of the hiring entity's business — is unsurvivable for a company whose product is the crowd's output. There is no version of your business that passes it.
- 2 December 2026 transposition in a Member State where you have EU-resident contributors — especially more than 250 in one Member State, which triggers Art. 13. The presumption flips the burden onto you, Art. 3 makes intermediaries jointly liable, and Art. 7 makes the anti-fraud stack that works unlawful. If your supply is EU-resident, that is a stop-and-restructure date, not a compliance project. See The law is about to arrive and Operating from Europe, selling to the US.
- A joint-employer rule finalised such that your buyer demands an uncapped, uninsurable worker-classification indemnity. That converts a legal risk into a commercial one you cannot price.
- Your own non-payment or deactivation docket. A second legal surface, separate from misclassification, opening the moment payment timing becomes a cash-management lever — see the Mercor filings of April–May 2026 and the Handshake withheld-pay record. It destroys the recruiting funnel and the legal position at once.
- A breach involving contributor identity data. Mercor lost ~4TB in a March/April 2026 breach including passports, SSNs and biometric face and voice data, drawing six class actions and causing Meta to pause work. The MSA carve-outs for confidentiality and data breach are uncapped.
The table
| Signal | Watch | Stop |
|---|---|---|
| Largest customer, % of revenue | >15% | >25% for two quarters, no second contract in diligence |
| Buyer set competes and one takes equity | Any | Immediately restructure or exit that buyer |
| Gross margin after rework, payouts and QC | <40% | <25% with no written path to 40% in four quarters |
| Repeat rate within one purchase cycle | <60% | <50% with high ticket and only speed-bump defences |
| Median contributor weekly hours | <20 | <20 for eight weeks with <40% cohort retention |
| Days sales outstanding vs contracted | +15 days | +30 days, or an unresolved acceptance dispute |
| Regulatory | Any of the five above | Prong-B finding, or EU transposition with EU-resident supply |
What should not kill it
A bad quarter. One lost logo below 15%. A competitor raising more — Surge AI passed Scale AI while bootstrapped against $1.6B of venture funding, and Toptal outlived every funded competitor in its cohort on $1.4M raised in 2012. Growth slower than the deck, if margin and retention hold.
The failure mode this page exists to prevent is not stopping too late for lack of information. It is stopping too late because each number arrived separately, with a plausible explanation attached, and nobody had written down in advance which combination meant the experiment was over.