NORTH OS · The North Report
Seven paths. Eleven cases. Every claim carries its source.
Why enterprise-AI programmes fail — assessed the way this platform assesses everything: eleven documented failures across six sectors, each anchored on a royal commission, a court ruling, a regulator's order, a peer-reviewed study, or a company's own signed disclosure. Journalism only corroborates. A separate red team then attacked every load-bearing claim: of 20, twelve survived as written, eight were cut to precision, and none was quietly dropped.
“Time saved is capacity, not value” — and the value conversation should begin with the business constraint, the intended use of released capacity, the accountable owner, and the evidence that will show whether the expected outcome actually arrived. The market is converging on this question. This report is our answer to it, codified.
1. The vendor knew — The failure evidence lived inside the vendor's organisation before or while the buyer deployed — the buyer's room carried risk it never got to price.
Pre-mortem check: Which risks about this system already live in the vendor's room — and what surfaces them to us in writing, before signature and again before launch?
Confidence: 0.85
2. Unlawful from day one — The core mechanism was known, in writing and before launch, to be unlawful or non-compliant; the programme launched anyway and the legal position eventually killed it.
Pre-mortem check: Does the core mechanism have a written, current legal and compliance read — and has it reached the people funding and building it?
Confidence: 0.90
3. Governance theatre — The responsible-AI checklist was fully executed and treated as evidence to proceed, while the sharpest objection the process produced was logged as input rather than acted on.
Pre-mortem check: Find the sharpest objection your own process has produced. Who could act on it — and did anything actually stop or change?
Confidence: 0.80
4. Scaling past the red line — A known threshold — accuracy, cost, safety — was defined and repeatedly missed, and the programme scaled anyway for years; termination came framed as success.
Pre-mortem check: What threshold did we set before launch, who owns stopping the rollout when it is missed — and how long has it already been missed?
Confidence: 0.85
5. Said but never heard — The concern was in the room — said aloud, held silently, or said and discarded — and the group's decision process never pooled it into the decision.
Pre-mortem check: Who holds a concern about this plan and hasn't said it — what makes it unsayable here — and when concerns were said, did they reach the decision?
Confidence: 0.80
6. Validation never caught up — Deployment ran ahead of independent validation, and the exposure compounded silently until the model's errors surfaced in the world or on the balance sheet.
Pre-mortem check: What independent validation exists before this model touches customers at scale — and how much exposure accumulates per week deployment runs ahead of validation?
Confidence: 0.85
7. The wrong yardstick — The programme was judged on a metric that could not survive contact with the baseline it replaced, and the metric was corrected only after the damage.
Pre-mortem check: What number decides whether this programme is working — and would the human baseline, the customer's quality bar and the P&L all agree with that number?
Confidence: 0.75
The asymmetry the cases expose is itself a finding: the vendor's room often knew first while the buyer carried the risk — and states write their knowledge down and get contradicted by it, while companies delete, deny, or call the knower a troll.
Verified cases (headline)
Robodebt (advised unlawful 18 months pre-launch; ~430,000 debts / more than $720M settlement) · England's A-level algorithm (experts NDA'd; 40% downgraded; scrapped within days) · SyRI, Netherlands (a decade of operation, then ruled incompatible with ECHR Article 8) · IBM Watson for Oncology (internal decks “shared widely with management” while marketed worldwide) · MD Anderson × IBM (a $2.4M contract extended 12 times to $39.2M) · Epic Sepsis Model (hundreds of hospitals, zero independent validation at adoption) · Amazon recruiting (knew by 2015, scrapped quietly) · McDonald's × IBM (three years past a stated threshold) · Zillow Offers (the write-down in their own release) · Walmart × Bossa Nova (humans matched the robots) · Babylon (regulator questioned safety in 2017; administration 2023).
What we could not verify ships beside the findings — Moffatt v. Air Canada primary, the RSS April 2020 letter, the gov.uk visa-streaming primary, the full Mitchell 1989 paper, Amsterdam's council record, and the CQC 2017 report are all logged as open, with the attributed relays we used instead. The full ledger and the per-claim provenance live in the repository run record alongside rules/paths.v1.yaml and rules/checks.v1.yaml.
Definitions. Failure-path coverage: the share of the seven named enterprise-AI failure paths — vendor knowledge, day-one unlawfulness, governance theatre, threshold overrun, unpooled dissent, validation lag, yardstick error — that a programme has surfaced, evidenced and assigned a named owner. Reported with a date and a ruleset version. Pre-mortem (operational definition): an independent, before-decision assessment in which participants name failure paths under consequence, producing evidence-graded findings rather than agreement. It measures confidence calibration and risk identification; it does not claim prevention.
The check for your programme is the same seven questions, asked independently, before the decision. That instrument is what this company builds. Start free, individually, no account.
Every claim on this page is checkable: /proof · the desk: /verify · the kernel: /os/kernel. This surface is deliberately still — one loop per viewport, proof viewports never flicker; motion arrives only with a provable STILL MODE.
UseNorth is a product of Lodecrest (Pty) Ltd (registration number 2026/557367/07, South Africa).