What problem does it solve?
Statutory social work runs on paperwork. A Care Act assessment, a social circumstance report, a visit record or a child protection case note each has a required structure, and most are required by the same statutory duty that requires the visit itself. Writing it up well takes real time: a full assessment can take hours after the conversation itself, on top of the visit, and it competes with the next family or person waiting for a visit.
Dictation software that turns speech into text has long helped with typing the words, but it still leaves the social worker to structure and write the note from a blank page. The newer generation goes further: a captured or described account of the conversation feeds a language model that drafts the note directly into the format the service uses, so the social worker's job becomes checking and correcting rather than writing from nothing. That is also where the risk sits. A record that a court, a panel or another professional relies on later must say what actually happened, not what a model guessed was likely to have been said, and a rushed sign off turns a helpful draft into an unreliable record.
This is distinct from summarizing a meeting into general notes and actions: a statutory case note or assessment has to follow the case record's own required fields and format, not a generic summary structure, and it becomes part of a legal record rather than an internal recap.
How does it work?
- Get consent. The social worker tells the person that the conversation will be recorded and used to help write up the visit, and records that consent; the person can decline.
- Record and transcribe, or prompt. A phone, tablet or meeting app captures the conversation and turns it into text, ideally with speakers separated; some deployments instead have the social worker enter a prompt describing what they heard and saw.
- Draft into the required format. A language model turns the transcript or prompt into a first draft of the specific document the service needs (a Care Act assessment, a social circumstance report, a visit note), using the council's own template and headings.
- Check, correct and own it. The social worker reads the draft against what they heard and saw, cross references anything from other professionals or relatives, corrects anything wrong or missing, and is the person who signs and submits it.
- File and audit. The finished note goes into the case management system with a record of who wrote and approved it; the audio and raw transcript follow the council's retention policy.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Handling time reduction | Too few to pool | 63% | 1 | 1 organization |
Value drivers: Employee productivity, Lower cost to serve, Customer experience.
Indicative value
A council with 1,000 social workers across children's and adult social care
USD 500,000 to USD 6.4 million
Social worker time cost avoided on case note and assessment drafting per year
How this is calculated
Formula: socialWorkers * documentsPerWorkerPerYear * hoursSavedPerDocument * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Social workers socialWorkers, social workers | 1,000 | 1,000 | The reference council. |
| Case notes and assessments per social worker per year documentsPerWorkerPerYear, documents per social worker per year | 40 | 80 | Editorial assumption, replace with your own caseload and documentation data. |
| Hours saved per document hoursSavedPerDocument, hours saved per document | 0.5 | 2 | Swindon Borough Council's is the only precise per document saving that can be computed from a source on this page: its reported write up time for a Care Act assessment fell from four hours to one hour thirty minutes, a saving of two and a half hours (Swindon Borough Council, "Council completes AI trial to enhance Adult Social Care services"). Lancashire County Council also reported a write up time falling, for a social circumstance report, from up to two days to around three to four hours, but that range is too wide to turn into a single figure. The high end of this range is set at 2 hours, below Swindon's 2.5 hour saving, and the low end at 0.5 hours, so the range also covers shorter, routine case notes and visit records, which take less time to write up than a full assessment and should save less. |
| Cost of a social worker hour, fully loaded costPerHour, USD per hour | 25 | 40 | Editorial assumption for a fully loaded social worker cost. Replace with your own. |
What it leaves out: Gross avoided admin time only. It leaves out the tool's licence cost, the time spent reviewing and correcting drafts (which is real and necessary, not zero), training time, and any change in the accuracy or consistency of the records produced.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Lancashire County Council
United Kingdom · Government and public sector · 2026
Lancashire County Council trained more than 1,400 social workers, educational psychologists and support officers in Adult Services and Education and Children's Services to use responsible AI. Staff were also supported to use Microsoft 365 Copilot prompts, a set of simple instructions that draft a social circumstance report (a complex assessment describing a person's living situation and support system) in the council's format; the source does not say the responsible AI training itself covered these prompts. The social worker checks and amends the draft, and still meets the person and speaks to relevant professionals and relatives for the report; a social care lead and a councillor both state the social worker remains responsible for what is submitted and that the tool does not replace professional judgment. The source does not say whether the visit itself is recorded or dictated, only that a Copilot prompt generates the draft.
No outcome disclosed.
Swindon Borough Council
United Kingdom · Government and public sector · 2024
Swindon Borough Council trialled Magic Notes, a tool built by the social enterprise Beam, across 184 frontline meetings with 19 social workers and one leadership support officer in its Adult Social Care department between April and July 2024. The tool records conversations between social workers and their clients during Care Act assessments, mental capacity assessments and other supporting conversations, and automatically generates detailed, high quality assessments. Following the trial the council signed a new six month contract with Beam for the tool.
- Handling time reduction: 63%, April to July 2024 trial
"The trial has been carried out in the Council's Adult Social Care department and has led to a 63 per cent reduction in the time social workers spend compiling their assessments and logging their case notes."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A template for each statutory document type with its required structure and headings
- A written consent process and script for telling the person their conversation will be used
- A clear rule on which conversations should never be recorded this way, such as one where an immediate safeguarding disclosure needs an instant response rather than a drafted note
Systems to integrate
- Case management or client record system, to file the finished, signed document
- The council's existing productivity or meeting platform, if the tool is built on top of one
- Secure storage for audio and transcripts that meets the council's data residency policy
Complexity: Low
The speech to text and drafting problem itself is well understood. The real work is getting the output into the exact statutory format each document type requires, and building the habit, not just the rule, that a social worker always checks a draft before it is signed.
- 1
Start with one document type and one team
Pick a single high volume, lower risk document (for example a routine visit note or a Care Act assessment write up) in one team, rather than every statutory form across the service.
- 2
Encode the exact template
Build the prompt or template around the headings and structure the document already has to have, so the draft needs editing, not restructuring.
- 3
Build in consent and an off switch
Give staff a simple, consistent way to ask for consent, to pause or stop recording, and to not record at all when that is the right call.
- 4
Keep the social worker as the author
The draft populates an ordinary editable document; the social worker reviews it line by line, adds their own analysis and professional judgment, and is accountable for what is filed. Never auto file a draft.
- 5
Measure quality alongside speed
Sample completed documents against the original recording or notes for anything invented, wrong or missing, not only for how much time was saved.
- 6
Widen by team and document type
Once time and quality both hold up, add more teams and document types, and keep the same review discipline as volume grows.
Guardrails
- No document is filed until the social worker has reviewed, corrected and signed it
- Explicit, recorded consent before any conversation is captured, with the option to decline
- An escalation path for urgent safeguarding concerns that does not wait for a drafted note
- Audio and transcripts stored securely, kept only as long as policy requires, and access logged
KPIs to instrument
- Time to complete each document type, before and after, by team
- Sample audit rate and findings on drafts checked against the original recording
- Social worker adoption, usage and satisfaction
- Time spent in direct contact with people, versus time spent on admin
Human in the loop
The social worker stays the author and the accountable professional. They check the draft against what they actually heard and saw, add their own analysis, and are the person who signs and submits the final record; the tool's job is to produce a first draft in the right format, not a final document.
Common failure modes
- A fluent draft states something that was not said
- A model can produce a confident sentence that is not supported by the recording, and a rushed reviewer can miss it. Prevent with mandatory review before filing and periodic audits of drafts against the source recording.
- Consent becomes a formality
- Under time pressure, staff record without properly explaining what is happening. Build consent into the workflow as a required step, not an assumption, and check for it in audits.
- A safeguarding disclosure waits for a drafted note
- A visit reveals an immediate risk, and staff treat the AI note as the next step instead of an immediate safeguarding referral. Train and test that the drafting tool is never the safeguarding process.
- Time saved becomes headcount cut, not more contact time
- If released time is only used to reduce posts, the case for the tool weakens with the workforce that has to adopt it. Track and report time released against direct contact time and caseload, not only against cost.
What are the risks and rules?
EU AI Act
Depends on design
The tool drafts documentation for a social worker to check and sign rather than deciding on services or eligibility itself, but a Care Act assessment or similar eligibility document is what an adult social care eligibility decision rests on, which sits close to Annex III 5(a): AI systems used by a public authority to evaluate eligibility for essential public assistance benefits and services. A deployment designed as a preparatory drafting step, feeding into but not replacing that eligibility assessment, can rely on the Article 6(3)(d) derogation for AI performing a preparatory task to an assessment relevant for the purpose of the Annex III use case, rather than the assessment itself. Relying on that derogation carries its own duties, and they fall on the provider: a council that builds its own tool this way is the provider under the Regulation, and it must document why the system is judged non high risk (Article 6(4)) and register itself and the system in the EU database (Article 49(2)). That is different from Article 49(3), which requires a public authority to register its use of a system that actually is high risk, and does not apply once the derogation holds. Separately, the provider of the generative model may owe the Article 50(2) duty to mark the drafted text as AI generated. If a council reused the same transcripts or drafts to feed a scoring or triage model, that downstream use would need its own risk assessment (see the referral risk triage use case).
Controls to put in place
- Human review and sign off required before any AI drafted document is filed
- The person being visited is told a recording will be used to help write up the visit, and can decline
- Access controls and an audit trail on who viewed, edited or approved each AI assisted record
- Regular sampling of AI drafted documents against the original recording for accuracy
Frequently asked questions
- Does the AI decide what goes in the case note?
- No. It produces a first draft in the required format. Lancashire County Council's own account says Copilot "puts the information into the right format and is a great place to start," and that "the social worker is still responsible for the report that's submitted": they cross reference it and add details about the person and any professionals they have spoken to before it goes in. Swindon Borough Council's own account of its trial does not say whether that review step is mandatory before a document is submitted.
- How much time can this realistically save?
- It depends heavily on the document type. Swindon Borough Council's own reported trial result was a 63% reduction in the time social workers spent compiling assessments and case notes between April and July 2024, with the write up time for a Care Act assessment falling from four hours to one hour thirty minutes. Lancashire County Council's social circumstance reports fell from up to two days to around three to four hours. Lancashire also gave an early estimate of more than 200,000 staff hours a year freed up from targeted use of AI on routine tasks, not case note drafting specifically and not a measured result, so plan conservatively for a first team against the like for like write up figures above rather than that estimate.
- What happens if a visit reveals something urgent, like a safeguarding risk?
- That has to go through the normal safeguarding process immediately, not through the case note drafting tool. The drafting workflow should never be the route by which an urgent concern gets raised or actioned.
- Is this tool itself high risk under the EU AI Act?
- Treat it as context dependent, not automatically minimal. Drafting a Care Act or other eligibility assessment sits close to Annex III 5(a), which covers AI used by a public authority to evaluate eligibility for essential public assistance benefits and services, so a deployment designed as a preparatory drafting step can rely on the Article 6(3) derogation for AI that performs a preparatory task to that assessment. That derogation brings its own documentation and registration duties, and they fall on the provider, which for a council that builds its own tool is the council itself. The provider may also owe the Article 50(2) duty to mark the generated text as AI produced. Reusing the same transcripts or drafts to feed a scoring or triage model is a separate, additional risk that needs its own assessment.
How to cite this page
Blits.ai AI Use Case Library, "AI drafting of social work case notes and assessments", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/social-worker-case-note-drafting. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 30 September 2026: Published after review by an automated review workflow (independent skeptic review).
- 30 September 2026: Editorial pass on the skeptic review blockers: corrected the EU AI Act basis and matching FAQ answer (Article 6(3)(d) covers a preparatory task, not a narrow one; the Article 6(4) documentation duty and Article 49(2) registration duty fall on the provider, which for a council that builds its own tool is the council, not "the provider or deployer"; Article 49(3) deployer registration only applies to an actual high risk Annex III system, not one covered by the derogation; dropped the unsourced "most deployments" claim for conditional wording), removed the mobile-app channel (neither evidence record shows an app of the organization's own), aligned "social circumstances report" to "social circumstance report" in howItWorks, and reworded the indicativeValue note and the FAQ time saving answer to stay closer to source wording. Set the Lancashire evidence record's outcomeDisclosed to true (the source describes a report time falling from up to two days to three to four hours; kept without a metric because the range is too wide for one figure), removed the unsourced "statutory" claim from its summary, split its training statement from its Copilot prompt statement (the source does not say the training covered the prompts), and renamed the evidence file from `lancashire-county-council-copilot-case-notes` to `lancashire-county-council-copilot-reports` (the source never mentions case notes).
- 30 September 2026: Unpublished by an automated review workflow (independent skeptic review).
- 29 September 2026: First published