Turning spoken notes into visit records

Care workers write up each visit, often at the end of a long shift, in the car before the next call, or hours later at home when the detail has faded. It is the most common complaint about the job’s paperwork, and the most asked-about use of AI in domiciliary care.

How it works

The care worker opens an app after the visit and talks for a minute or two about what happened. The tool converts speech to text, then arranges it under the headings your service already uses — personal care, nutrition, mood, anything to report.

The care worker reads the draft, corrects it, and saves it. Nothing goes on the record until they do.

That last sentence is the whole design. If your process lets a draft save itself, you have built something different and considerably worse.

Where it helps

Speed. Talking is faster than typing on a phone for almost everyone.

Timeliness. A note spoken in the car after a visit is more accurate than one typed at ten that night, simply because less has been forgotten. This may matter more than the time saved.

Consistency. Records come out in the same structure, which helps whoever reads them next — including a social worker, a nurse or an inspector.

Accessibility. Staff who find written English difficult, including many excellent care workers for whom English is a second language, can produce clear records. This is the benefit we hear most from managers and see discussed least.

Where it fails

Medication names. Transcription does not fail by producing gibberish. It fails by producing a different real word. Many drug names sound alike, and all of them are plausible to the tool. Check every one against the MAR chart.

Names and places. Local street names, unusual surnames and place names are the next most common failure.

Noise. A television on, a busy kitchen, traffic through an open car window — accuracy drops fast, and the tool does not tell you it is struggling.

The detail that did not fit. A care worker mentions six routine things and one odd one: she seemed confused about what day it was. The summary comes back organised under neat headings and the confusion is gone, because it did not belong under any of them. Nothing was invented and the record is now worse than useless, because it reads complete.

Smoothing. “He shouted and threw the cup” becomes “he became agitated”. Each rewrite is defensible; the cumulative effect can hide an escalating pattern from whoever reads the records later. See safeguarding and AI.

Choosing the tool matters more here than anywhere else

This is the use where the difference between tools is starkest, because you are handling audio of an identified person discussing their health.

When we researched our directory in September 2026, we found tools at both ends:

• One is built for social care, and the councils using it state that all processing happens in the UK, with recordings and transcripts deleted after 12 weeks.
• Another states in its own privacy policy that it uses content “for training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)”, draws no distinction between its free and paid tiers on this, describes no opt-out, and hosts on US infrastructure.

Both appear in our directory, because the directory is neutral and hiding a tool helps nobody. But they are not interchangeable for this task. Read the entry, not the name.

How to trial it

1. Test on an invented person first. Never a real one. See testing prompts with a made-up person.
2. Do the DPIA. Before any real visit. See do you need a DPIA.
3. Tell the people you support, before the trial rather than after. The ICO requires it, and it is also simply the decent order to do things in.
4. Measure the baseline first. Time the task as it is now, for two weeks. See measuring a baseline.
5. Start with two or three willing staff, not the whole team.
6. Sample and compare. Pick a set of records during the trial and check them against what actually happened. Count the corrections, not just the minutes.
7. Measure again once it has settled, a few weeks in — not in week one, when everyone is still learning.

What it costs

Licence costs for transcription tools in our directory start around £15 per user per month on business plans. The real cost includes the DPIA, the training, and the checking time that never goes away.

If a care worker saves six minutes and spends four checking, the saving is two minutes, not six. That is still worth having across a team. It is not the number in the sales deck. See paying for AI.

When to stop

If more than about one draft in five needs a substantive correction, the tool is making work for that task. If staff have started saving drafts without reading them, stop immediately — the record now carries an implied check that did not happen, and that is worse than no tool at all.