Your employer bought the AI. You carry the risk.

And, yes, there is a risk. A big one, in fact.

The software writes it, but the name on the record is still yours — and nobody has yet settled who answers when it gets something wrong.

AI has become part of ordinary working life in health and social care without much of an announcement. Copilot drafting an email. A chatbot tidying up a report or a referral. Software turning a conversation into a set of notes. For a lot of people it turned up before any clear rules about how to use it did.

This is our own take on where that leaves the person doing the work. It is not a summary of anybody else's reporting, and it is not advice — just the picture as we read it from what has actually been published.

Not everyone, and not everywhere

Plenty of people reading this will not be using these tools at all. Social Work England's research described the rollout across the profession as "uneven", and a 2025 report for the Department for Education found adoption in English children's social care still "limited and experimental". Some services have deliberately kept AI out of certain work.

It is spreading, though. A Community Care poll of 713 social workers in October 2024 found 21% had already used AI tools for daily tasks. Social Work England's later survey found 40% had used AI at their employer's direction, and a further 24% had used it without any.

By February 2026 a further Community Care poll of 400 practitioners found most reporting a positive experience: nearly two-thirds said using AI for admin had gone very or moderately well, only 9% called it negative, and 26% had still not used it at all. That is worth holding on to while reading the rest, because for most people, most of the time, these tools are helping.

So this is not a claim about everybody's experience. It is about what the evidence shows, and what seems worth knowing before it becomes your experience.

What it is actually being used for

Almost all of it falls into one category: getting through the paperwork and spotting patterns. Very little of it involves the software deciding anything.

Across the NHS and health services, the King's Fund's 2025 review of emerging use found AI being used to:

• Read scans and images — flagging possible abnormalities in radiology, pathology and dermatology, so the most urgent cases move up the list.
• Write up consultations — so-called ambient scribes that listen and produce a structured summary.
• Draft routine paperwork — fit notes, benefits applications, housing letters, responses to information requests.
• Predict who will miss an appointment, so they can be contacted and offered a different option rather than simply recorded as a no-show.
• Read thousands of safety incident reports to find recurring themes, at a scale no team could manage by hand.

Across social care and social work, Social Work England's research found the four most common tools to be virtual assistants, transcription software, case recording support and chatbots. In practice that means:

• Turning visits and meetings into case notes — the single most common use, and the one with the most evidence behind it.
• Searching case files — pulling a timeline or a specific document out of years of records in seconds rather than minutes.
• Rewriting dense material in plain language — turning a medical or legal explanation into something a family can actually use.
• Scheduling and rostering — matching home visits to staff and building sensible routes.
• Monitoring and pattern-spotting — analysing large volumes of care notes to identify things like fall risk.

Social Work England's survey found practitioners broadly positive: 83% thought AI could reduce administrative burden. It also found something quieter and worth noting — that these tools can help bridge language and learning gaps, and can act as assistive technology for neurodivergent staff.

What it is mostly not doing — and one place it clearly failed

Notice what is missing from those lists. Almost none of it is the software making a judgement. Practitioners themselves are far less convinced on that front: in the same survey, only 48% thought AI could help with decision-making, and 46% with identifying risk and need.

That caution is well placed, because there is one area where prediction was tried seriously and the results were poor. Machine learning models built to identify children at risk were trialled over 18 months across four local authority areas. A published review of the results found that, on average:

• the models missed around four out of every five children who were actually at risk;
• and where they did flag a child, they were wrong roughly six times out of ten.

Department for Education guidance issued in 2024 now says these tools are "higher risk, especially where predictions relate to individual children and families", and advises that a service should be able to show it handles simpler analysis confidently before attempting prediction at all. Eileen Munro, who led the national review of child protection, has put it more bluntly still, arguing that the underlying data is "imbued with prejudice, distortion and inaccuracy".

There is a second catch, and it is easy to miss. Saving time in one place does not create capacity in another. One social care supplier told the King's Fund its system had identified 650 people at risk of falls — and that services had the capacity to respond to about 40% of them. A tool that finds more need than a service can meet has not solved a problem. It has relocated it.

What it gets wrong

The errors are not usually gibberish, which would be easy enough to spot. The dangerous ones read perfectly well.

Practitioners interviewed for published research described an AI tool recording suicidal ideation that had never been discussed. One said: "If I hadn't checked that and it had gone into the case note… that could have had implications further on down the road." Others found the wrong names attached to the wrong people.

There is a related problem that no amount of accuracy fixes. A transcript captures words. It does not capture the pause before an answer, the body language, or the fact that someone said "I'm fine" while looking at the floor. Practice guidance is consistent on this: keep taking your own notes on what you observed, because the tool is only listening.

Why that lands on you

In February 2026 the Ada Lovelace Institute published research based on interviews with practitioners using these tools across a range of local authorities. Its central finding is not about the technology. It is about where the risk sits.

Practitioners, the researchers found, are "responsible for the additional work of ensuring the accuracy and quality of outputs" — carrying full professional accountability for tools chosen and introduced above them, while training and guidance are still catching up. Social Work England's research found the same ambiguity from the other direction: where AI is in use, it is often unclear whether responsibility for its output sits with the practitioner, the line manager, or the company that built it.

In practice, it resolves in one direction. The name on the record is yours.

The preparation is not there yet either. Among social workers who qualified in the last five years, 86% had no preparation for AI during their training.

The rules are still being written

On 17 September 2026 the Professional Standards Authority, the ten health and social care regulators it oversees — Social Work England among them — and a body representing 28 further professional registers issued a joint statement of intent. That phrase matters: a statement of intent is not a rule, it is a promise to write one. They committed to producing broad principles on how practitioners' use of AI will be regulated.

Sarah Blackmore of Social Work England: "Social workers need clarity about what is expected of them when using AI, and the people they support need confidence that their care and wellbeing remain firmly in human hands."

That statement did not come out of nowhere, and the picture has become clearer since we first published this.

A week earlier, on 10 September 2026, an independent commission set up by the medicines and devices regulator reported on how AI in healthcare should be regulated. It made 44 recommendations. The regulators' joint statement a week later was a direct response to it.

Then on 6 October 2026 the government accepted all 44. Draft guidance follows in December 2026, a consultation on classifying AI tools during 2027, and a full roadmap in spring 2027.

Elsewhere: the NMC is consulting on a new Code covering the safe use of technology and AI, open from September to December 2026 — so nurses and midwives can shape it now. The Care Quality Commission has published its expectations, built on the principle of "AI to support, not to replace". And in Ireland, HIQA published national guidance on AI in health and social care in July 2026 — published guidance, not a promise of some.

The pressure it is landing on

None of this is happening in a calm sector. Skills for Care reported in June 2026 that the adult social care vacancy rate in England had fallen to its lowest in a decade — genuinely good news — while still sitting at roughly three times the rate in the wider UK economy.

Set that beside a finding that recurs throughout the research: checking the output takes time, and that checking eats into the time the tool saved. These tools do not so much remove work as change what the work is, from writing things down to checking what has been written.

That is a design problem, not a character problem. It is not about diligence, or grit, or caring enough. A tool that produces confident, plausible text and leaves the checking to whoever is still at their desk at the end of a long day has put the burden in the wrong place, and it will do that to anyone anywhere.

It is not all bad practice

It would be easy to read the above as employers versus practitioners. That is not the picture — much of what is known about the weaknesses of these tools is known because organisations went looking and then published what they found. Recurring patterns in the better published examples:

• Keeping it in a closed system rather than pointing staff at a free public chatbot — a contained version, a written contract covering how the information can be used, and a formal check of the privacy risks before anything starts.
• Having someone else mark the homework — commissioning an outside evaluation instead of accepting the numbers from the company selling the software.
• Drawing a clear line. Some organisations have kept AI out of particular kinds of work entirely; some prohibit it during face-to-face contact. Deciding where the line sits and writing it down is worth far more than leaving it to whoever is on shift.
• Publishing the bad news too — saying openly where a tool performed poorly, which no supplier would volunteer.

So, no: this is not a reckless sector. If anything, the published material gives the impression that the tools arrived first, the evidence is still arriving, and the rules are arriving last.

But that is only an impression, and it comes from what happens to have been written about. An organisation that paused, trained its staff first and moved slowly has nothing dramatic to announce and may never have published a word. Careful work is exactly the kind that does not become a case study. What can be said with more confidence is narrower, and it is the reason for this article: where the rules are unsettled, a share of the risk tends to end up with the individual rather than with the system around them.

If English is not your first language

The Ada Lovelace research notes that these tools "perform differently for people with characteristics that are less represented in the data used to train" them — including speech patterns tied to regional dialect or ethnicity. A 2025 study of clinical speech transcription found a measurably higher error rate for non-native English speakers than for native speakers.

The same concern runs wider than accents. Practice guidance warns that these tools can reproduce racist and sexist assumptions present in the data they were trained on — so an output about a family whose background is poorly represented in that data deserves a more sceptical read, not a quicker one.

What to do about it

• Never put identifiable details into a free public chatbot. BASW's guidance is explicit that general tools are not a lawful route for personal information. A tool your employer has contracted is a different thing from the same brand used from your own account.
• Read every line before your name goes on it — not a skim for tone, a read for facts that were never said.
• Check what carries the most weight: names, dates, medication, numbers, risk, and anything that reads like a disclosure.
• Keep your own notes on what you observed, not just what was said. The tool cannot see the room.
• Edit rather than accept. Rewriting the output in your own voice catches more than approving it as written.
• Think about consent before the conversation, not after — and treat it as genuinely harder where someone may lack capacity.

Questions worth asking where you work

None of these are confrontational. Anywhere that has thought this through will have answers ready.

• Where does the information go, who processes it, and has a privacy risk assessment been done?
• Are there conversations where we have decided this should not be used?
• Who is accountable if an AI-generated error reaches a formal record — and is that written down?
• How much time is allowed for checking, and is that time real or assumed?

How we use AI here

We would be poor company if we wrote all of that and said nothing about ourselves. We use AI heavily. The short, checkable version:

• Your CV is not typed into a public chatbot. It is processed by contracted providers at defined points, under agreements that bar them from using it for their own purposes.
• An assessment we give you decides nothing about you. It is not shared with employers, it gates nothing, and it does not stop you applying for any role.
• You can tell us we are wrong and ask a person to look again. We would rather you did that than assume the machine was right.
• If how we use AI changes materially, we will say so here rather than quietly amend a policy page.

The detail — every lawful basis, every provider category, every retention period — is in our privacy notice, written to be read rather than survived.

Sources and last check. Last checked 11 October 2026. This article is our own analysis. The published material behind it: the Professional Standards Authority and health and social care regulators' joint statement of intent (17 September 2026); Social Work England's commissioned research on the emerging use of AI in social work, by The Open University and Research in Practice; the King's Fund, More than just hype: how emerging AI use is assisting health and social care (May 2025); Ada Lovelace Institute, Scribe and prejudice? (February 2026); the British Association of Social Workers' practice guidance and statement on generative AI; Care Quality Commission guidance on artificial intelligence; the Nursing and Midwifery Council's Code consultation; HIQA national guidance (July 2026); Research in Practice's report on AI in case recording for the Department for Education; published reviews of machine learning trials in children's social care and the Department for Education's 2024 guidance on data analytic tools; Community Care's October 2024 poll of social workers; Skills for Care workforce data (June 2026); the government's response to the National Commission into the Regulation of AI in Healthcare (6 October 2026); Social Work England on the cross-sector AI principles (17 September 2026); and Community Care's February 2026 practitioner poll. Where a figure comes from a trial rather than settled practice, we have said so. If you spot something out of date, please tell us — we would rather fix it than leave it.