Faster repairs, fair decisions: AI agents for a UK housing association under the new automated decision rules

The situation

A housing association in England managed tens of thousands of homes. Repair requests arrived by phone, web form, email and through housing officers, and were keyed into the housing management system by a contact centre. At busy times, requests waited days to be read. A message mentioning "black marks on the bedroom wall" or "the baby keeps coughing" could sit in the same queue as a dripping tap.

Why now. Two sets of rules changed the stakes. Awaab's Law requires social landlords to investigate and fix damp, mould and serious hazards within set timescales, with the first phase in force since October 2025⁵. And the Data (Use and Access) Act replaced the UK GDPR's rules on automated decisions with new Articles 22A to 22D from 5 February 2026, permitting more automated decision-making but with safeguards such as information, the right to contest and human intervention²³. The ICO has said it is working on updated guidance and a code of practice on AI and automated decisions¹, and in October 2026 it set out its expectations for AI agents and called for views⁴. The association needed speed and safeguards together.

What had to be true

• Every request read on arrival, whatever the channel.
• Hazards escalated by rules, with statutory clocks that start when the landlord becomes aware.
• People decide what significantly affects tenants; AI classifies, summarises and suggests.
• Every step traceable, so complaints and the Housing Ombudsman can see what happened and why.
• Tenants can challenge an outcome and reach a person.

Architecture

The data flow, end to end

The hard part: long-running cases that must not get lost

A repair case is not a single model call. It can run for weeks: classify, wait for a surveyor, wait for a tenant to confirm access, raise a job, wait for a contractor, check it is fixed. Agents built as simple request-and-response loops lose state when a service restarts, and nobody notices until a statutory deadline has passed.

The team ran each case as a durable workflow. The workflow engine records every step, so a case resumes exactly where it was after a failure, and it can wait days for a person or a contractor without holding resources. Tracing had to follow the whole case, not just individual calls; the OpenTelemetry and Dapr communities have worked on carrying trace context across the workflow engine's internal boundaries so steps form one coherent trace⁶.

The second hard part was deciding where automation stops. Classification and routing of ordinary repairs were automated end to end. Anything that could significantly affect a tenant, such as refusing a repair, deciding it was the tenant's responsibility or flagging possible vulnerability, went to a person with the agent's summary and evidence. That kept the association comfortably within the safeguards the new rules expect, and made challenges easy to handle³. See structured outputs for how the classifications were kept reliable.

How it ran

Key decisions and trade-offs

What I'd do again, and differently

Again: let housing staff define hazards and vulnerability in their own words, then turn those into rules and test cases. Differently: bring the complaints team in from the start. Their cases were the best test set we had. Related: evaluating AI agents, agents as users of enterprise systems, UK cyber resilience readiness and the UK page.

If you are bringing AI agents into public-facing services in the UK, see my AI architecture work, or analyse your challenge and bring the result to a call.