Domain Atlas / Child welfare & family services
Hackney / Xantura Early Help Profiling
Work with this case in the PAN Lab ↗
Hackney paid the analytics firm Xantura £361,400 over four years to run an Early Help Profiling System that flagged families for preventive intervention from council data, but scrapped the pilot in 2019 after finding that, despite flagging about 350 families, it surfaced only 7 children previously unknown to the council and the available data was too limited and variable to justify continuing.[2]
What happened
Between 2015 and 2019 Hackney paid the analytics firm Xantura £361,400 to run an Early Help Profiling System, procured through the London Councils and Ernst & Young "London Ventures" programme, that mined data across council services — social-care referrals, school attendance and exclusions, youth offending, domestic-abuse and antisocial-behaviour reports, truancy and benefits status — to flag families for preventive intervention. Each month it sent social workers a short list of the families it judged most at risk, written as narrative summaries rather than a numeric score, and the families themselves were not told: notice came only through a general online privacy notice, and the impact assessment recorded no opt-out. The council dropped the pilot in 2019, reporting that the available data was more limited and more variable than expected — despite flagging about 350 families, the tool surfaced only 7 children previously unknown to it, and it could not deliver enough genuinely new insight to justify further investment. It was a rare kind of ending: not a public scandal, but a cost-benefit judgment that the system did not work well enough to keep.
The sociotechnical reading
Hackney matters precisely because it is small. Most institutional AI is not a statewide system but a pilot bought from a vendor by a stretched team — and the governance surface of a pilot is thin: procurement is the main control event, monitoring is informal, and quiet failure is the likely end state. The vendor-gate and provenance patterns exist for this scale. A pilot that cannot say what data it runs on, and how families would ever learn of it, has already failed a governance test regardless of its accuracy. What is documented sharpens the point. The method was withheld as commercially sensitive and requests for the data-sharing agreements were refused, so no independent evaluation of the tool's accuracy or harm was ever published — the buyer could not fully say what it had bought. And the thing that finally stopped it was not the model at all but the records beneath it: variable, incomplete administrative data meant the profiles added little, which is why the useful levers here sit upstream of any model tuning — procurement with teeth, data minimization, and someone with the authority to ask who signed off that families could be profiled without being told.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.