Before we blame the chatbot, look at the queue

I can understand why someone with a tenancy problem opens a chatbot. It is available when they are. It will read the long explanation. It does not appear impatient when they ask the same question again.

Those qualities matter whether the person is worried about the home they rent, a property they own or a difficult situation they are responsible for managing.

They also help explain why warnings about AI, however necessary, will never be the whole response. People are looking for help they can reach. If we want them to use better help, we need to make it available at the moment they need it.

“What should I say back?” is a question any of them might ask. A tenant may want someone to check a claim that their landlord is breaking the law. A landlord may want help understanding a disputed deduction. A property manager may need to respond to conflicting accounts without making the situation worse. We should meet each need with care.

The risks are real. A confident answer can contain the wrong law. A convincing citation can lead nowhere. A person who began with a repair complaint can end up with pages of allegations they cannot substantiate. The other party must then respond, and the Tribunal may have to untangle material that should never have been included.

The Principal Tenancy Adjudicator’s guidance recognises that AI can help people identify issues and organise documents. It also warns about invented cases, false rights, altered material and inaccurate timelines. People must check the material themselves; asking an AI tool to verify it is not enough. The party remains responsible for what they submit. Errors can lead to corrections, reduced weight or delay; serious cases may be struck out. Costs can follow where the material makes proceedings frivolous or vexatious. The guidance also prohibits using AI to record, transcribe, summarise or analyse Tribunal hearings or mediation discussions.

What interests me is the design of the help itself.

Suppose a tenant writes: “They are keeping my bond and I think this is unfair.” The landlord may see an unpaid bill or damage they believe needs explaining. The property manager may hold the inspection records. A useful service would establish what amount is disputed, what explanation has been given and what records exist, giving each account room to be checked. It might discover that one deduction is contested and the rest is agreed. It might discover a more serious problem. Either way, the next question should follow the facts.

We should expect technology to help a person recognise uncertainty. There should be room for “I do not know”, “that needs checking” and “you need someone qualified to look at this”. Those answers can be enormously valuable when the alternative is acting on misplaced confidence.

Our check-before-send demonstration uses fictional material to explore where checking is needed. It illustrates a workflow; it does not provide the personal verification the Tribunal expects before someone relies on a document.

I worry about measuring progress by how much a system can produce. The person dealing with the problem has to read it. So does the person receiving it. A long document can transfer work from the machine to everybody else while giving its author the impression that something substantial has happened.

A better measure is whether the person can now explain the issue, identify the relevant evidence and understand their options. Have we reduced the work needed to make a sound decision? Have we preserved the point that matters?

Access to justice includes language, confidence, time and the ability to participate. A person should not have to write like a lawyer to be taken seriously. They should also be able to trust that a tool helping them has not invented a legal argument or altered their account.

That trust requires real discipline behind the conversation: reliable sources, careful handling of private material, clear limits and human responsibility. Technology should help people keep control of their decisions, including the decision to seek advice or stop negotiating.

I am optimistic about what we can build. The same advances that make it easy to generate another ten pages could help someone ask one useful question, find the missing record or understand an answer they have already received.

When the public debate moves on from the next extraordinary AI story, the person with the tenancy problem will still need help. That is the test I want us to keep in view.

Read next: Pets, permission and the cost of getting stuck →