Domain Atlas / Housing & homelessness services
Xantura OneView (predictive homelessness flagging)
Work with this case in the PAN Lab ↗
Xantura's OneView integrates more than 15 multi-agency data feeds into a single household view and flags residents as likely to become homeless months ahead. In Maidstone's pilot year it produced 650-plus alerts that a single financial-inclusion officer could contact only about 260 of. Its headline effectiveness figures - a reported 40 percent fall in homelessness, savings and an ROI over 600 percent, and the widely quoted contrast between contacted and uncontacted households - are vendor- and council-reported pre/post numbers from one COVID-affected pilot year; the contact-versus-no-contact contrast reflects capacity-driven selection rather than a randomised comparison, and the independent randomised controlled trial commissioned to test the causal claim was still in progress into 2026.[5]
What happened
Xantura Ltd built OneView, a data-integration and predictive-analytics platform that combines 15 or more internal and external council data feeds into a single household view and applies risk models to flag households at risk of homelessness before crisis. Barking & Dagenham began an engagement with Xantura and EY in 2018 and procured OneView in 2019 to develop analytics across children's social care, homelessness and adult services; the platform later generated a Covid-19 "OneView" risk model in 2020. Maidstone Borough Council partnered with Xantura and integrated 15 services and external agencies — housing register, council tax, housing benefit, tenancy and rent-arrears debt from a housing association, and domestic-abuse and "troubled families" data from the county council — with agreed thresholds triggering alerts to the housing team. In Maidstone's pilot year, the tool generated more than 650 alerts for households 3 to 6 months from crisis; a single financial-inclusion officer had capacity to contact only about 260 of them, and a second officer was later recruited. The council and vendor reported that 0.4% of the highest-risk contacted households later presented as homeless, versus about 40% of the alerted households the officer could not reach — a contrast widely quoted, but one produced by capacity-driven selection rather than a randomised comparison. Vendor and council materials also reported a 40% fall in homelessness, roughly 100 households prevented, savings of about £225k (up to ~£578k at full capacity), a vendor-stated ROI over 600% and around £2.5m in societal savings; these are single-pilot-year, pre and post, council- and vendor-reported figures from the COVID period, with no published methodology, base rate or confidence interval. In Barking & Dagenham, the OneView "Single View of Vulnerability" is governed by a statutory Digital Economy Act 2017 data-sharing agreement (public register #376, running 1 June 2023 to 31 May 2026) listing DWP, the Ministry of Justice, the council and the Metropolitan Police as controllers and Xantura, INBest and CareTech among processors, covering identifiers, health, social-care, education, offending, benefits and debt data. An independent Ada Lovelace Institute ethnography of that early deployment (fieldwork 2020, published July 2024; its study areas were children's social care and the COVID-19 response) found social workers and managers lacked a transparent explanation of which factors drove the case summaries and predictive alerts, and that some frontline staff were unconvinced the analytics were as objective, neutral or accurate as described. Big Brother Watch's 2021 "Poverty Panopticon" investigation criticised the vendor's Covid OneView, used by two London councils, as building on thousands of data points including sensitive items unrelated to housing, operating without residents' knowledge. To test the causal claim the pilot figures could not establish, the Ministry of Housing, Communities and Local Government commissioned a "Using Data to Prevent Homelessness" Test and Learn randomised controlled trial — managed by the Centre for Homelessness Impact, with Xantura supplying the pseudonymised at-risk list and Verian Group and Simetrica-Jacobs as independent evaluators, across Barking & Dagenham, Newham, Stockport, Test Valley and (per the MHCLG privacy notice) Kensington & Chelsea. In the trial, households are randomly assigned to proactive support calls or control; outbound calls began around April 2025, and, with full causal results due in 2026, the trial was still in progress into that year.
The sociotechnical reading
Almost every other predictive tool in this Atlas is judged on how accurate it is. OneView asks a prior question: even a correct flag does nothing until someone acts on it — and here far more households are flagged than one officer can reach. That makes reach a governed resource. The model's real coverage was set by staffing, not by the classifier, and the harm did not land on the households the tool got wrong; it landed on the ones it may have got right and no one ever contacted. That is the distinctive lesson of this case, and it is why the marquee numbers mislead. The widely quoted contrast between contacted and uncontacted households looks like an effectiveness result, but contact was rationed by capacity, not assigned at random, so the gap measures who got selected, not what the intervention did. A pre and post number from a single, COVID-affected, capacity-rationed pilot year is not evidence of effect; it is the reason to run a controlled trial — which is precisely what the government then commissioned, and which was still running as this was written. Two further controls are load-bearing and both sit off the accuracy axis. The first is transparency as a working condition: the independent ethnography found frontline staff could not see which factors drove an alert, so their discretion could not be the calibrated check the design assumes — an opaque flag is either over-trusted or quietly dismissed. The second is the legitimacy of the integration itself: a single view assembled from sensitive non-housing records — offending, health, benefits, debt — on residents not told they are scored is governed here by a statutory data-sharing agreement and a data-protection function, but external scrutiny still framed it as automated suspicion. In map terms, the levers that matter are not another model upgrade but the ones that govern reach, evidence and data: concentrate scarce review where stakes are highest, hold the tool to an independent evidence bar before crediting its numbers, make the model inspectable through procurement, and minimize and authorize the feeds that build the flag. The honest reading also keeps the benefit in view: unlike the punitive scoring cases, OneView's intended output is proactive help, and the double-edged question is whether prediction widens that help or simply widens the reach of surveillance faster than the capacity to act on it.
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.