The Government Has Published AI Guidance for Care Providers. Here’s What It Asks of You.

Earlier this year a provider showed me a care plan. It was tidy, well structured, used all the right words, and had been written in about ninety seconds by a free online AI tool. The manager was proud of it. Then I asked who had typed the person’s name, date of birth, diagnosis and medication list into that free tool, and where that information had gone.

Silence.

That moment is why the new government guidance matters. On 29 September 2026 the Department of Health and Social Care (DHSC) published “Using AI in adult social care” on GOV.UK, written for providers. It has already been updated once, on 3 November. It is not statutory guidance and it does not create new legal duties. What it does is tell you, in the government’s own voice, what good looks like. I have read it in full. Here is what it asks of you, in plain English.

First, what it is and what it is not

The guidance describes itself as an introduction to AI in adult social care and best practice for providers. It is aimed at England, as DHSC guidance is, though the sense of it travels well across the UK. It says openly that many providers are still in the early stages and have not adopted AI at all, and it says, just as openly, that AI may not always be the right answer. I like that honesty. It is not a sales document.

It does not mention the Care Quality Commission. If you want the regulator’s view, I wrote about CQC’s May 2026 AI statement separately. The two documents fit together: CQC tells you how AI will be judged, DHSC tells you how to use it well.

Where AI is already being used in care

The guidance lists the kinds of AI providers are using now. Some will be familiar; some surprised me.

• Sensors and acoustic monitoring. Movement and sound sensors in care homes that switch a light on when someone gets up at night, detect falls, and reduce disruptive night checks. The guidance cites one care home that reported an 84% reduction in falls after introducing a night-time smart lamp.
• Chatbots. A first point of contact on helplines, available 24 hours a day.
• Facial recognition for pain assessment. Software that picks up facial movements linked to pain and guides staff through a structured pain score. Useful for people who cannot tell you they hurt.
• Data analytics. One home analysed its incident reports and found falls clustering at the same spot at the same time of day. The cause was strong sunlight. They fitted blinds. Falls and hospital admissions dropped. No robot required, just the data they already had, looked at properly.
• Generative AI for care plans and assessments. Drafting plans, helping with audits, logging daily data.
• Admin. Meeting notes, action lists, draft business plans, recruitment and onboarding administration, multi-site audit summaries.

Notice how much of that list is about back-office time, not care decisions. That matches what I see. The wins are in the paperwork first.

The sentence every manager should read twice

On generative AI and care planning, the guidance says the evidence that it can produce truly personalised care plans is currently limited, and that a member of care staff must review any AI-assisted plan for accuracy.

That is the government saying what experienced managers already know. A care plan is not a document. It is a promise about how we will treat a particular person on a particular Tuesday when she does not want her shower. A tool that has never met Mary cannot make that promise. It can help you write it down faster. Someone who knows Mary has to make it true.

The ten best-practice points, grouped

DHSC sets out ten headings of best practice. Here they are, grouped the way I would explain them to a team.

Get the inputs right. “Input high quality data.” AI is only as reliable as what you feed it. Incomplete, badly structured records produce confident, wrong summaries.

Protect the data. “Consider data protection laws.” Comply with UK GDPR. Know where your data is processed and whether it leaves the EEA. Tell people how their data is used. And the line I would put on a poster: do not enter personally identifiable information into free or online AI tools, because there is no contract protecting it. That is the ninety-second care plan problem, named by the government.

Be fair and ethical. “Maintaining ethics” and “Be aware of and mitigate bias.” The guidance points to the Oxford Project principles: truth, transparency, equity, trust, accessibility, humanity and responsiveness. Bias can come from the data, the design or the people curating it, and it can quietly distort decisions.

Keep a human in charge. “Always have a human review.” Review every output for accuracy and fairness, set review guidelines, define the human’s role in writing.

Train your people. “Upskill your workforce.” Build skills and confidence, and consider appointing digital champions.

Buy carefully and count the cost. “Outline the process for choosing AI tools” and “Evaluate the financial cost.” Decide what you want the tool to achieve before you look at any product. Write a specification. Involve staff and the people you support. Compare benefits and risks, and watch for hidden costs in hardware, training and changed processes.

Write it down and keep improving. “Introduce an AI policy” and “Continuously improve.” Have a policy even if your policy is that staff may not use AI. Cover how each role interacts with AI and include a review checklist. Then gather feedback and adjust.

What this looks like in home care and in care homes

In a care home, the guidance’s examples lean towards sensors, night monitoring and incident analytics. The practical question is governance: who owns the sensor data, who acts on the alert at 3 a.m., and how does that get back into the care record and the risk assessment? A smart lamp that nobody reviews is a gadget, not care.

In home care, the risk is more personal and more mundane. A coordinator under pressure, a free chatbot open in another tab, a client’s details pasted in to “just tidy up this assessment.” No malice, no awareness, and a data breach that would take months to unpick. The policy point and the free-tools point are the two that matter most for domiciliary providers, and they cost nothing to implement.

Four things I would do before Christmas

1. Write the AI policy. One page is enough to start. Who may use what, for what, with what data, and who reviews it. I will share a template for this in my next post.
2. Audit the free tools. Ask every office-based member of staff what AI tools they have used at work in the last month. You may be surprised. Then give them a safe alternative.
3. Name the reviewer. For any AI output that touches a care record, write down who checks it and how. Make it part of the job description, not a hope.
4. Look at your own data first. Before buying anything new, run the sunlight test. Where do your incidents cluster? Your reporting tools may already hold the answer.

How this shaped what we built

I will be open about where Careberry stands. We were early adopters, in two senses. We use AI every day in how we build and run Careberry, from our development team to how I write and plan, and I have written before about how I use it. And we were early to put AI inside a care management platform, carefully, with a human approving everything. Those two experiences taught us the same lesson the guidance now sets out: AI is at its best when it removes paperwork and at its worst when it is trusted blindly.

Careberry AI is not a separate app. It switches on inside the platform teams already use, with the same login, the same role permissions and the same audit trail. Three modules are live today.

• AI Assist reads the daily notes for the people you choose, over the period you choose, and highlights risks, actions and errors so nothing important is missed. It reviews care plans for gaps and inconsistencies, and helps staff write clearer notes and family updates. A qualified person approves every suggestion before it touches a record.
• Knowledge AI lets you upload your own policies, procedures and manuals once. Staff then ask a plain-English question in the carer app and get an answer drawn only from your documents, citing the exact section. No per-user charges, so the night carer has the same access as the manager.
• Capture AI is the one I am proudest of. An assessment, supervision or one-to-one is run as a natural conversation. Consent is asked for and logged before any recording starts, and if the person declines, the meeting carries on with written notes. The conversation is transcribed and the details land in the right care plan fields, ready for a coordinator to edit and approve. Eye contact, not keyboards. We estimate it saves three to four hours per care plan.

Read the government’s best-practice list again and you will see why it was built this way. Human review: a person approves everything, and every AI action is logged with who requested it, what was suggested and who approved it. Data protection: UK GDPR compliant, servers and database hosted in the UK, and client, staff and policy data never used to train AI models. Transparency and choice: consent before recording, every time, and the whole thing is optional. AI in Careberry is off until you switch a module on, and you can switch it off again at any point without losing anything else. The guidance says AI may not always be the right answer, and we agree, so we built it so you can say no. Choosing tools and counting cost: one fixed monthly fee, no tokens, no credits, no surprise bills.

The analytics example in the guidance, the falls that clustered by time and place, is exactly the kind of question Report Hub was built to answer with the data providers already record every day. The most useful “AI” is often just your own information, organised well enough to see the pattern.

Careberry is on the NHS Assured Solutions list, meets the Data Security and Protection Toolkit, and is certified to ISO/IEC 27001:2022. When a provider asks where their data is processed, we can answer in one sentence. The guidance says to ask your supplier. Please do.

There is more coming. Our rule for every new AI feature is the same: it must save a carer or a manager real time, and a human must stay in charge of the record.

People first. Process next. Technology always.

A closing thought

The manager with the ninety-second care plan was not careless. She was busy, and the tool was there, and nobody had told her where the line was. The government has now drawn that line, gently and sensibly. Read the guidance. Write the policy. Keep a human in charge. And remember that the person the plan is about has never read a line of it, and does not care how fast it was written. They care whether we got them right.

If you would like to see AI Assist, Knowledge AI and Capture AI working inside an assured care record, with a human approving every step, book a viewing of Careberry.