There is a conversation happening right now in housing that nobody is leading. It’s about whether the AI that was built for everyone is actually fit for the work they do – and the honest answer, the one that keeps getting dressed up in careful language, is no.
This is the opposite of an anti-AI rant; instead, it’s an argument for AI that takes housing seriously enough to be fit for purpose.
Generic AI was not designed for highly regulated compliance environments. There, I’ve said it.
The large language models that power most enterprise AI tools were optimised for scale. Useful to millions of people across thousands of tasks. That breadth is a genuine achievement, and it is precisely why these tools are a structural mismatch for regulated housing work.
Here is the thing: AI models produce more confident language when they are wrong than when they are right. Research published in January 2025 confirmed what housing professionals (who have used these tools) already suspect.
The failure mode is not the AI that refuses to answer or crashes. It is the AI that answers fluently, completely, and incorrectly, with nothing in its output to tell you which one it is doing!
Generic AI has no mechanism for knowing that the guidance in its training data has since been superseded. It will produce a polished, well-structured response based on regulation that may have changed eighteen months ago, and it will do so with the same confidence it brings to everything else. That is not a bug that can be patched.
For general enterprise use, that is a nuisance. In housing – where decisions must be traceable, explainable, and defensible to the Regulator of Social Housing and the Housing Ombudsman – it is a liability.
And the same dynamic is playing out across financial services, healthcare, and local government. Generic AI tools are built for scale, not sector.
The gap nobody wants to name
The National Housing Federation’s own research found that almost half of housing associations are already using AI in their operations; 87% report low levels of AI knowledge; 44% have no AI policy in place.
Did I mention that’s a liability?
When a housing officer uses a generic AI tool to assist with a policy document and the output reflects regulation that has since changed, the accountability for that error does not transfer to the AI provider. It stays with the organisation.
“The AI said so” does not satisfy a regulator. It does not satisfy an ombudsman. And it will not hold up when someone asks for the audit trail.
The amendment currently moving through Parliament – cross-party backing for a government kill switch over AI systems that pose a risk to essential services – is a political signal pointing at an operational question.
Can you explain what your AI is doing and why? For a significant number of housing organisations right now, the honest answer is no. Not because they have made bad decisions, but because the tools they are using were never designed to make that answer possible.
This is not a warning, it is an invitation
Generic AI was not built for housing organisations operating under live regulatory scrutiny. It was not built for environments where the evidence chain must be reproducible on demand.
It was not built to know the difference between current RSH standards and superseded ones, to cite its sources so the decision trail stays intact, or to flag what it does not know rather than fill the gap with fluent-sounding invention.
What replaces it is governed AI, a curated, validated knowledge base rather than open-internet retrieval. Vertical, sector-specific AI relies on architecture built from the ground up for a defined regulatory context.
In housing, that means outputs validated against current regulation, no confident assertions based on superseded guidance, and an audit trail that will satisfy the Housing Ombudsman. The same principle, applied to the FCA’s framework or CQC requirements, produces the same result – AI that can account for itself.
The housing professionals who are getting this right are asking themselves – Is the AI we are using built for the work we do?
If the answer is no (and for most generic tools, it is) that is the reason to demand something better.
