A jobcentre sits at 30 Drysdale Street, metres from a tech gold rush. The state's AI there is real, funded and pointed at fraud. The AI pointed at getting you a job — that one doesn't exist yet. This is the case for building it, offline, for the price of a phone.
Employment expansion in Shoreditch & Hoxton wards — fastest in Hackney. The boom is literally across the street.
Hackney is the third most income-deprived borough in London. Boom and poverty share one postcode.
of working-age residents hold no formal qualifications. 15,000+ more are trapped in insecure work.
Digital-skills levels in the borough — the interface of every government service is the exact skill in shortest supply.
THE JOBS ARE ACROSS THE STREET. THE TALENT IS HERE. WHAT'S MISSING IS THE MACHINE IN THE MIDDLE.
A Work Coach sees you for 10–15 minutes, eighty to a hundred caseload deep. The friction between a claimant and a live vacancy isn't ambition — it's paperwork, language, digital access and time. Friction is exactly what machines are for.
Sources: Hackney Economic Development Plan 2025–2030 (draft); DWP labour-market statistics. Office-level visitor breakdowns are not published — cohort mix here is modelled from regime data, not observed.
30 // Hoxton Jobcenter (N · 5F) · 26 // Unbound (N · 2F + terrace) · 7–117 // The Timber Yard (S · 3–5F courtyard) — the state node sits one gate away from the tech gold rush. 51.52815°N · 0.07982°W.
DWP's fraud-detection ML is officially 2.5× more effective than the old system. Real. Funded. Works.
DWP's own fairness monitoring: non-UK nationals are 2.27× more likely to be flagged for review. The machine has a tilt.
of conditionality claimants sanctioned. And 89.9% of adverse decisions are for missed appointments — not fraud. Procedural failure, punished like crime.
licences reached frontline Work Coaches. Central staff got AI helpers. The people who help you, didn't.
Welfare-fraud risk scoring struck down in The Hague — Article 8, ECHR. Opaque, aimed at poor neighbourhoods, unchallengeable. The blueprint for what to never build.
Profiling model scored "female" as a −0.14 penalty and applied childcare only to women. Independent audit: >30% misclassified. Deployed anyway.
Sources: DWP algorithmic-transparency records 2024/25; UC sanctions statistics to May 2026; Rechtspraak SyRI ruling; AlgorithmWatch/LMU AMAS evaluation.
employment likelihood; 30% faster to a job. Applications went down as fit went up. Fewer, better shots.
hires, +8.4% wages, biggest for the weakest initial writers. A writing machine for people the CV screen discards.
job-finding for women; +4pp stable contracts. Tells you which firms are about to hire — signals, not advice.
employment in six months; +5% match per pair. Gains largest for the least-educated — the Hoxton demographic.
caseworker documentation time at ~98% accuracy, with 100% of advisers in one pilot wanting to keep it. The human gets their hour back.
Every win shares one shape: concrete, wired into real vacancies, and human-adjacent. Kill a named friction between a person and a real job — not "engagement", not logins, not chat.
A beautifully engineered personal coaching chatbot, randomised across 212,277 jobseekers: a precisely-estimated zero on every employment outcome. Advice at scale is a rounding error.
Cross-agency risk scoring of benefit claimants — Article 8 breach. The court killed it. Opaque state profiling is a legal dead end, not a shortcut.
the model's literal coefficient for being a woman. Care duties applied to women only. >30% misclassified. Automated triage of humans reproduces the bias you feed it.
Every failure shares one shape too: it either performs helpfulness instead of doing work, or it profiles people for the state without consent, transparency or appeal. Both fail. One fails quietly, one in court.
Full source trail in the companion research report. Effect sizes as published; pp = percentage points. Bob Emploi rules out any effect >0.4pp.
A KIOSK THAT CAN'T SNITCH. OFFLINE BY ARCHITECTURE, NOT BY POLICY PAPER.
Aligned with current UK policy direction: sovereign-AI MOU, UK data-residency option (MoJ first adopter, Oct '25), National Data Strategy "data infrastructure sovereignty". Legal posture: UK GDPR satisfied by design, not paperwork.
Hackney Central Library. Where the jobcentre-avoidant already go.
Hoxton Trust / food-bank floor. Where crisis cohorts actually are.
2–3 kiosks. Booth, mic, speaker, the box. Under £5k all-in.
Navigator 2 days/week + referral line to Hackney Citizens Advice. The machine opens doors; people walk through.
DO THE MATH YOURSELF. DRAG THE FADERS.
Time-to-first-interview, not "sessions logged".
Benefit-uptake checks passed — money people were owed but never claimed.
User trust score (NVLQ). If it reads as the state's ear, usage collapses to zero — covenant is the product.
Caseworker hours returned to human conversation (Beam baseline: 63%).