High Street to Housing: Hard-Won Lessons in AI Transformation

The lessons below are presented for housing providers embarking on, or accelerating, their own AI journeys. The operational parallels are substantial: customer-facing services, large and often siloed datasets, complex organisational cultures, constrained budgets, and a public-interest mission that demands both ambition and care.

1. Start with problems, not with AI

One of the most common mistakes organisations make is beginning with the question, 'How do we use AI?' The more productive question is: 'What are our most important customer and operational problems, and where do we already have data that could help?'

This reframing led to a focus on digital touchpoints where interaction data was richest: website behaviour, app usage, email engagement, loyalty data. Crucially, these were areas where the impact of any change could be measured, which made investment decisions defensible.

2. Build your data foundation before your major investments in AI capability

Early investment in stitching together previously siloed customer data proved to be a force multiplier. By connecting data sources into a single, real-time view, it became possible to make decisions in the moment rather than acting on yesterday's information.

The practical implication for housing is not that a specific platform is required. It is that data integration is a prerequisite for large scale AI, not an afterthought. If you cannot see a tenant's full journey, your AI will be working with fragments.

3. Make experimentation work is a cultural change, not just a technical one

Experimentation, running small, controlled tests to validate ideas before scaling, was not just a technical practice. It became the primary mechanism for shifting organisational culture. A counterintuitive principle had to be established at leadership level: a high failure rate in experiments is a sign of health, not dysfunction.

Early on, roughly 60-70% of experiments succeeded, but only a dozen were run each quarter. Over time, the success rate fell to around 50%, but experiment volume grew to approximately 10x per quarter. The organisation was learning faster precisely because it was failing more.

Two practical points:

• Set experiment volume as a KPI alongside success rate. Reward teams for running tests, not just for winning them.
• Capture and share what you learn from failures. Today's failure is often tomorrow's success as tenant expectations evolve.

4. The highest-value use cases are often the simplest

The use cases that drove the greatest value were not the most technically complex. They were the ones where the right data already existed and where the impact could be cleanly measured. Three are particularly instructive for housing:

Personalisation

Showing each customer relevant content rather than a one-size-fits-all experience drove significant commercial impact. For housing, this might mean surfacing the most relevant self-service options, repair tracking, or support information based on a tenant's prior interactions.

Content sequencing and suppression

Not every message is relevant to every customer. Showing irrelevant content erodes trust. Housing equivalents: suppressing rent arrears messages to tenants who are current, prioritising damp and mould communications to those in affected property types, tailoring content to tenants' tenancy stage.

Language optimisation

By testing >10 variants of email subject lines, it emerged that customers fell into six distinct language profiles, from formal to informal, emoji to no emoji. Tailoring subject lines to these profiles improved open rates by nearly 20%. Housing providers communicating with diverse tenant populations have significant potential here, particularly for reaching tenants who currently disengage from written communications.

5. Avoid the MVP trap

Moving quickly to get AI capabilities live has a hidden cost. Building many minimum viable solutions creates technical debt. Over time, engineering resource gets consumed maintaining legacy MVPs rather than building new capability.

The specific failure mode was allowing AI models to remain the responsibility of the teams that built them, rather than establishing operational functions to monitor, maintain, and update them. As the number of deployed AI capabilities grew, this became unsustainable.

The advice: start with MVPs, but plan deliberately for the point when you will need dedicated AI operations capability. That means clear ownership, monitoring for model drift, and governance processes. The mistake is waiting too long to make that shift.

6. The real challenge is cultural, not technical

Every AI use case encountered organisational resistance. Analysts worried about being replaced. Content specialists felt their expertise was being automated. Creative teams feared loss of brand control. This friction was consistent and predictable.

What worked was not pushing through resistance, but pulling resistors in. Making domain experts part of the design process, using their knowledge to define what good looks like, and demonstrating that AI amplified rather than replaced their contribution, turned critics into advocates.

The evidence over time was that AI created more roles than it destroyed. It augmented existing jobs rather than eliminating them. This is the reality that tends to get lost in broader public debate about AI and employment.

A good example was store operations where there are complex problems to solve which historically sit with the store manager. Rather than taking decision rights away from the manager, a more effective solution was found to be providing them with the tools to make better decisions.

7. Know when not to use AI

AI is not cheap to run. The important distinction is between problems that need a probabilistic answer (where AI earns its cost) and problems where the answer is always the same (where a simple rule works better and costs far less). Here, AI can be used to build a sophisticated rule set but is not necessarily required on an ongoing basis.

A clear example: when a customer calls to ask about the status of their order, automated lookup and playback of the relevant information serves the customer perfectly well. No AI inference required.

8. Check what you already have

Many organisations assume they need to build AI capability from scratch. In practice, AI is now embedded in a large number of existing vendor platforms, and most organisations are significantly underutilising what they have already paid for.

Early on the journey, building internal models outperformed vendor tools. But as AI has matured, embedded capabilities in third-party systems have become genuinely competitive, and the case for custom build is narrowing.

9. Prepare for AI-enabled tenants, not just an AI-enabled organisation

Perhaps the most forward-looking insight concerns what happens when your customers become AI-enabled. Over half of UK adults now regularly use AI assistants such as ChatGPT and Copilot. Increasingly, customers will not start their service journeys on your website or app. They will ask an AI assistant their question and receive an answer drawn from whatever information about your organisation is publicly available.

If a tenant asks an AI assistant 'my boiler broke three days ago and my housing association has not fixed it, what are my rights?', the assistant will answer. The question for housing providers is whether the information those AI systems draw on is accurate, comprehensive, and reflects well on your organisation.

Summary: Ten lessons for the housing sector