Insight Bristol's CSE/CCE Models — a Decommissioned Predictive Risk Score for Child Exploitation

Insight Bristol was a joint data-analytics unit set up in 2015 by Bristol City Council and Avon and Somerset Police, combining police, council, school, housing and charity records -- including data from Barnardo's on roughly 1,000 children with a known history of abuse -- into a shared 'Think Family' database.

From around 2016 the unit built two predictive risk-scoring models -- one for child sexual exploitation (CSE) and one for child criminal exploitation (CCE) -- to help identify children at risk, using variables such as housing support, rent arrears, free-school-meal eligibility, school absences, missing-person reports and mental-health concerns; gender and ethnicity were excluded from the scores.

The CSE and CCE models were used operationally from around 2016. In 2023 the council concluded they were 'not fit for operational use' and quietly decommissioned them.

A subsequent review by Social Finance, reported by the Bristol Cable in 2026, found the models were the 'weakest element' of the programme: known victims of sexual exploitation sometimes scored lower than people involved in burglaries, some vulnerable children did not appear in the results at all, and frontline staff reported low confidence in the scores. Reviewers could not fully audit the models because the source code and list of variables could not be located, and neither the council nor the police retained records explaining why the models were withdrawn.

England's Children's Commissioner, Dame Rachel de Souza, called the findings 'deeply concerning' and said predictive tools used on children 'must be transparent [and] well understood by professionals.' The UK's Independent Inquiry into Child Sexual Abuse (IICSA) has since cited predictive-analytics cases like this one as evidence of the risks of unaudited algorithmic profiling in child protection, while the OECD's Observatory of Public Sector Innovation had earlier profiled Insight Bristol's broader data-sharing work without flagging these specific model failures -- a reminder of how hard such flaws can be to detect from outside.