HAILIE AI Development Sprint

Executive Summary

The Housing Ombudsman Service (HOS) publishes some of the most objective evidence there is on how social landlords perform. Each of its determinations is an independent ruling on how a landlord treated a resident, and it has issued more than 16,000 over the past five years.

Complaint outcomes are only one lens on performance, alongside measures such as tenant satisfaction, but they are among the hardest for a landlord to present in a flattering light. The difficulty is not the data itself, it is access to it. The evidence sits across three disconnected forms — summary statistics, an in-year report for each landlord, and more than 16,000 individual determination PDFs — and none of them, on its own, allows a fair, like-for-like comparison or shows how a landlord is changing over time.

What was the goal?

The primary objective of the sprint was to build a system that automatically gathers, cleans and categorises the Ombudsman's data into a single database that can be enriched and analysed. There were two phases:

The Data: Three views of the sector

The sprint drew on the three forms in which the Ombudsman publishes its complaints data. Each is valuable, and each is incomplete on its own. Bringing them together is what makes fair, longitudinal comparison possible, and the sprint captured all three.

The first two sources give an authoritative but static picture of the present. The third, once extracted and structured, is what unlocks the trends and the like-for-like comparisons that follow.

What did we do?

The sprint ran as two parallel workstreams, governed by a shared data ontology.

1. Data Engineering

• Built a custom data ontology: a set of guidelines that let an AI model categorise determinations without relying on rigid text formatting.
• Taught language models to read the semantic intent behind the Ombudsman's wording, so the system could categorise determinations and their severity accurately.
• Created an ethical extraction pipeline to move data from the HOS website into our own database, then structured and cleansed the PDF text.
• Captured, for every determination: landlord name, landlord type, case ID, decision date, determinations and findings.
• Captured individual landlord reports: adding official current-year figures (maladministration rate, stock size, compensation and category detail) to validate and complement the extracted determinations.

2. Analytics and Quality Assurance

• Validation: benchmarked the extracted data against the HOS 2024-25 Annual Report. Where summary statistics could not be reproduced, we investigated the Annual Report's own parameters and exclusions.
• Enrichment: appended external landlord metadata, adding organisational size (stock volume), regional data and Tenant Satisfaction Measures (TSMs).
• Normalisation: to compare providers of very different scales fairly, we expressed volumes as adverse findings per 1,000 homes. Adverse findings are defined as determinations of service failure, maladministration or severe maladministration.
• Analytics: formulated hypotheses and designed the visualisations that follow, which will feed into the dashboard.

What were the challenges?

The sprint met several challenges. Each became a learning opportunity, and getting the database right matters because everything the dashboard will do depends on it.

Pre-Extraction Challenges

Post-Extraction Challenges

What were the findings?

The consolidated database does two things the published record cannot. It lets us compare landlords of very different sizes on a fair basis, and it lets us see how performance has changed across five years. We begin with the current picture, then turn to the trends the data unlocks.

Two views of the same year: rate versus intensity

The Ombudsman's headline measure is the maladministration rate, the proportion of findings upheld. On that measure the regions of England look much alike, ranging only from 65% to 73%. But normalise adverse findings to the size of a landlord's stock and a sharp divide appears: London landlords record 5.9 adverse findings per 1,000 homes, close to five times the North's 1.2. The rate conceals what the intensity reveals. This distinction, between how often a landlord is found at fault and how much maladministration is reported and escalated to HOS, frames everything that follows.

The trend behind the headline: a steep rise, now easing

Across the five years the sector's maladministration rate climbed from the mid-40s to around 70%, as complaint volumes grew and residents became more aware of their right to redress, before easing to about 62% as the statutory Complaint Handling Code embedded through 2024 and 2025. This recent drop in the maladministration rate may reflect a mixture of shifting regulatory definitions such as internal thresholds or legal language adjustments, alongside actual operational improvements by landlords.

Up until this point, individual landlords took very different routes: Peabody's rate more than doubled, London & Quadrant sat above the sector line throughout, Notting Hill Genesis began below the average before crossing above it and Clarion sat in the middle of the pack before significantly lowering their rates. A single year's figure cannot tell a board whether its landlord is improving or deteriorating. The trend can, and this is the central capability the sprint unlocked.

The London gap is widening

The regional divide is not a fixed feature of geography; it is growing. In 2020-21 every region sat clustered near the bottom of the scale. Since then London has pulled steadily away, reaching about 4.2 adverse findings per 1,000 homes while the North has stayed near 0.8, so the gap has widened from roughly three-fold to five-fold. Two checks give us confidence: the finding aligns with the Ombudsman's own published analysis, which records London as having the highest volume of complaints per property of any region, and it holds even after the three largest London landlords are removed, so it reflects a general regional pressure rather than a few providers.

The changing face of failure: damp and mould

What drives maladministration is shifting. Damp and mould's share of all adverse findings roughly tripled over the period, from around 5% to 14%, taking ‘property condition’ as a whole past a third of all adverse findings. The rise tracks the heightened attention that followed the Awaab Ishak inquest and anticipates the phased introduction of Awaab's Law. Complaint handling itself remains the single largest category throughout, a reminder that how a landlord responds to a complaint can matter as much as the fault that prompted it.

The Scale Myth: bigger is not worse

There is no linear relationship between a landlord's size and its rate of adverse findings. The largest providers are not systematically worse per home than smaller ones; scale, in itself, neither protects nor condemns. The highest-intensity outliers are small and mid-sized landlords, but almost all of them are in London (shown highlighted), so what first looks like a size effect is largely the same geography effect seen above. The practical lesson is that landlords should be benchmarked against genuine peers of similar size and region, not against the sector as a whole.

Perception and reality align: satisfaction predicts adverse findings

We had expected that self-reported satisfaction and objective Ombudsman findings might diverge. The data shows the opposite. Landlords whose tenants report lower overall satisfaction (the TP01 measure) record markedly higher adverse-findings intensity, a strong negative correlation of r = -0.62. Tenant Satisfaction Measures, far from being disconnected from regulatory outcomes, are among their better predictors. This strengthens the case for reading satisfaction data and complaints data together, exactly the kind of cross-referencing the enriched database is built to support.

What were the “AI wins”?

What are the next steps?

Sources and Licence

• Housing Ombudsman Service determinations and annual reviews (official data sources).
• Regulator of Social Housing TSM dataset 2024-25.

The code and methodology developed in this sprint are open artefacts. Content is available under CC BY-SA 4.0.