AI and self-directed support: where it fits

Self-directed support rests on a simple idea: the person decides. They shape their own support, and where they hold a direct payment, they arrange it.

AI can support that idea or undermine it, and the two can look similar from the outside. This page is about telling them apart.

The test

For any proposed use of AI in this area, one question does most of the work:

Does this give the person more control over their support, or less?

A tool that helps someone understand their options and say what they want in their own words widens control. A tool that produces the plan and then asks the person to agree with it narrows control, however collaborative the language around it.

Where it genuinely helps

Making information understandable. Care and support information is written in the language of the system. AI is good at rewriting dense text into plainer English or towards an easy-read structure. See making letters and leaflets easier to read.

Helping someone say what they mean. Some people find it hard to put what they want into the form the system expects. Someone talking about their week and having that shaped into the headings an assessment uses keeps the content theirs and changes only the packaging.

Reducing the paperwork between a worker and a person. If drafting takes less time, there is more time in the conversation. That is the strongest honest argument for AI in this area, and it depends entirely on whether the saved time actually goes back to the person.

Explaining what a direct payment can be used for, in general terms — with the caution in the next section.

Where it goes wrong

Treating it as a source of truth on entitlements. These tools predict likely text. Asked what someone is entitled to, or what a direct payment can cover, a model will produce a confident, plausible answer that may be wrong, out of date, or from another country’s system. Entitlement questions go to the local authority, not to a chatbot. This is the single most likely way AI causes harm in self-directed support — not through a dramatic failure, but by someone acting on a confidently wrong answer about their own money.

Pre-filling the plan. A support plan generated from previous records and presented for approval is not co-production. Most people, asked to approve a document that already looks finished, will approve it. The form of consultation is preserved and the substance is gone.

If AI drafts anything in a support plan, the person’s own words should be the input, and the draft should be visibly a draft.

Assuming digital access. Not everyone has a smartphone, reliable internet, or the confidence to use them. CQC’s principles require “transparency and choice”, including that “non-digital routes to care are offered if needed, also considering digital skills and connectivity”.

A digital-only route is not neutral. It moves effort onto the people least able to carry it, and the people it excludes are disproportionately those who need support most.

Flattening the unusual request. Self-directed support exists partly so people can arrange things the system would not have thought of. AI works by producing likely continuations, so an unusual but entirely reasonable arrangement is exactly what it will tend to smooth away. If a draft plan looks like every other draft plan, something has been lost.

Telling people

If AI is involved in producing someone’s support plan or the information they receive, tell them. That is a legal duty as much as a courtesy — see telling the people you support.

For people arranging their own support, it also matters the other way round: some will use AI themselves, to understand their options or draft their plan. That is their choice and it is not for a provider to police. What you can usefully do is be honest that these tools get entitlements wrong, and make sure a real person is available to check anything that matters.

A reasonable position

AI is a drafting and explaining tool. In self-directed support that makes it useful at the edges — turning system language into human language, and giving a worker more time in the room.

It is not a route to deciding what someone needs, and it should never be the thing that makes a plan look finished before the person has had their say.

If you are considering it, start with ways to use AI and the readiness rating on each page.