Better care notes in less time, often by voice, and early warning of patterns such as falls, weight loss, missed medicines and safeguarding signs, put in front of a manager who decides what to do.
At a glance
An illustrative scenario. The numbers are typical examples, not a client case study. The cloud services and UK rules were checked on 9 November 2026.
Care notes are the record of someone's life in care: what they ate and drank, how they seemed, what help they had, whether they took their medicines, and anything that worried the carer. The law expects them to be accurate, complete and contemporaneous. In practice, a home care worker writes them on a phone at the end of a 30-minute visit before driving to the next, and a care home carer writes them at the end of a long shift. Notes become short and repetitive, such as 'all fine, ate well', and the detail that matters is the first thing lost.
Most providers now use a digital social care record. Government figures show adoption rose from 41% of registered providers in 2021 to 80% by July 2025. That means the information is there, but it is spread across thousands of notes, incident forms, medicines records and visit logs. A slow pattern, such as a resident eating a little less each week, a third fall in a month, missed medicines on weekend visits, or a change in someone's mood after a particular visitor, is easy to miss when each note is read alone.
The stakes are high. Providers must notify CQC of serious injuries and allegations of abuse without delay, raise safeguarding concerns with the local authority, and show inspectors evidence of safe, well-led care. Managers spend days pulling that evidence together before an assessment. CQC is also changing how it assesses social care, with new sector-specific frameworks expected from the end of 2026, so the evidence a provider keeps needs to be easy to reorganise.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Carer time spent writing notes after each visit or shift, measured in the pilot services.
ExampleDown by a third
Share of notes that cover what the care plan asks for, checked by a monthly audit sample read by senior carers.
ExampleAbove 90%
How long from the first sign (a second fall, a weight drop) to a manager reviewing it.
ExampleFrom weeks to within 2 days
Share of flags the manager found useful. Very low means noise; near 100% can mean the threshold is too high and things are missed.
ExampleStable, and reviewed monthly
CQC notifications and safeguarding referrals made without delay, from the date the concern was first recorded.
ExampleEvery case in the audit
Manager time spent gathering evidence before a CQC assessment.
ExampleDown by half
Cross-sector rules are on the UK regulation map.
Regulation 17 requires an accurate, complete and contemporaneous record for each person. A draft written by the AI only becomes the record when the carer who gave the care checks and confirms it, at the time. Flags support safe care (regulation 12) and safeguarding (regulation 13), but responsibility stays with the registered manager.
Deaths, serious injuries and abuse or allegations of abuse must be notified without delay. The AI can spot a possible notifiable event and remind the manager, but it never decides whether something is notifiable and never sends a notification.
The single assessment framework still applies. CQC consulted in late 2025 on sector-specific frameworks, published drafts in March 2026 that remove scoring and bring back rating characteristics and key lines of enquiry, and planned to start using them from the end of 2026. The five key questions stay, so evidence is tagged to those and can be re-mapped when the final framework lands.
Local authorities lead safeguarding enquiries for adults at risk. The provider's job is to recognise and report concerns promptly. The AI raises possible signs with the manager and safeguarding lead; a person decides whether to make a referral, following the local multi-agency procedures.
Care notes are special category health data, often about people who may lack capacity. That means a DPIA before the pilot, processing in UK regions, a clear retention period, and a contract that stops providers training their models on the data. People and their families are told how notes are used. Since 5 February 2026 the strict limit on solely automated significant decisions applies mainly to special category data, so every decision here is made by a person.
DHSC's guidance asks providers to have a human review every AI output, keep personal data out of free online tools, write an AI policy that sets out who is accountable, and tell people how their data is used. It points to the Oxford Project's principles on responsible generative AI. The new system is covered in the provider's Data Security and Protection Toolkit submission.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Channels
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Voice capture and transcriptionThe carer dictates after the visit, not during it, so the person's conversation isn't recorded. Test on regional accents, carers whose first language isn't English, and noisy homes. | Amazon Transcribe | Azure Speech in Foundry Tools | Open-weight speech models (e.g. Whisper) on your own GPUs |
| Drafting and summarisingBoth clouds offer a choice of models, including Claude. Drafts only use what the carer said and the care plan; nothing is added. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Pattern detectionSimple, clear rules (three falls in 30 days, a 5% weight drop) do much of the work. Language models read free-text notes for signs that rules can't see. | Amazon SageMaker AI, plus rules in the workflow | Azure Machine Learning, plus rules in the workflow | MLflow and Kubeflow on Kubernetes, plus rules in the workflow |
| Care plans and policies (retrieval)Each person's current care plan, plus the provider's own policies, so prompts and evidence come from approved sources. | Bedrock Knowledge Bases | Foundry IQ, built on Azure AI Search | OpenSearch, or PostgreSQL with pgvector |
| Agent runtimeThe original Bedrock Agents closed to new customers on 30 July 2026; new builds use AgentCore. | Amazon Bedrock AgentCore | Foundry Agent Service | An open-source agent framework (e.g. LangGraph) on Kubernetes |
| WorkflowRuns the nightly checks and routes each flag to the right manager. | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Middleware and events | Amazon EventBridge, SQS | Event Grid, Service Bus | Apache Kafka or RabbitMQ |
| Link to the care record and eMARThis decides the design. Agree in discovery whether drafts can be written into the record for the carer to confirm, or shown alongside it. | Through your care record supplier's published API | Through your care record supplier's published API | Through your care record supplier's published API |
| Human reviewCarers confirm notes where they already write them. Managers see flags where they already work. | In the care record and the manager's existing dashboard | In the care record and the manager's existing dashboard | In the care record and the manager's existing dashboard |
| GuardrailsBlocks drafts that add facts the carer didn't say, or that read like a clinical judgement. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionCarers only see the people they support. Managers only see their own services. | IAM Identity Center, KMS, Macie | Entra ID, Key Vault, Purview | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditLog every draft, every edit, every confirmation and every flag, with what the manager decided. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. Commissioners' contracts may require UK hosting. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK or EU provider |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.
Stays on-prem
The digital social care record, the eMAR and the incident system stay with their suppliers, and remain the record. Rostering and visit monitoring stay where they are. The AI never becomes the record.
Uses cloud AI
Transcription, drafting, nightly pattern checks and evidence packs run in a UK cloud region, or on the provider's own GPUs for the sovereign option. Only the notes needed are sent, and drafts and audio are deleted once the note is confirmed, within an agreed retention period.
How they connect
Through the care record supplier's API and an API gateway. Head office links over a private connection such as AWS Direct Connect or Azure ExpressRoute; carers' phones and care home tablets use the care record app as they do now.
If the link drops
Carers write notes in the care record app as they do today, including offline in homes with no signal. Voice notes wait on the device and are drafted when it reconnects. The nightly checks run late, and managers are told if a run was missed.
What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.
Note assistant
Where a person signs offThe carer who gave the care checks and confirms every note. A draft is clearly marked as a draft until they do.
Pattern review
Where a person signs offThe registered manager reviews every flag and records what they decided. Safeguarding referrals and CQC notifications are made by people.
Evidence pack assistant
Where a person signs offThe quality lead and registered manager check every statement before it is used. The agents' permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
The digital social care record stays the record. The AI reads and drafts through the supplier's published API, and nothing it produces becomes part of the record until a carer or manager confirms it. If the supplier only offers read access, drafts are shown alongside the record for the carer to copy, which is slower but still safe.
Medicines stay in the eMAR. Missed or refused doses are read from it, never inferred from notes, so the pattern checks don't double-count or contradict the medicines record.
Events keep the AI decoupled. A confirmed note, a new incident or a weight entry becomes a message on a queue. If the AI service is slow or down, carers keep working in the care record as they do today, and the checks catch up later. Every message carries an ID, so nothing is flagged twice.
Map how notes, incidents and notifications work today, audit a sample of notes, agree what the care record supplier allows, and choose the first patterns with managers and the safeguarding lead.
Gate · must be true to passBaseline note quality and time measured, supplier route confirmed, DPIA started, AI policy drafted.
Voice notes and pattern review in four to six services, in shadow mode first: flags go to a review group, not to services, and are compared with what managers already knew.
Gate · must be true to passNote quality up, carers want to keep it, flags judged useful, no missed safeguarding signs in the audit.
Live in the pilot services, with the evidence pack for the next assessment, monitoring, support and a rollback plan.
Gate · must be true to passMonitoring, support model, rollback plan and carer training in place.
Roll out by region, add patterns such as hydration and pressure care, and re-map the evidence pack to CQC's final framework.
Gate · must be true to passEach new service and pattern passes the same pilot gate.
| Activity | Director of operations | Product owner | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Quality and compliance | Safeguarding lead | DPO |
|---|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Informed | Informed |
| Target architecture and care record integration | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Informed | Consulted |
| DPIA, AI policy and DSPT | Consulted | Consulted | Consulted | Informed | Informed | Consulted | Consulted | Accountable and Responsible |
| Safeguarding signs and pattern rules | Consulted | Consulted | Consulted | Informed | Responsible | Consulted | Accountable | Informed |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Informed | Consulted |
| Pilot testing in real services | Accountable | Responsible | Consulted | Responsible | Responsible | Responsible | Consulted | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted | Consulted |
| Run, monitor and improve | Accountable | Responsible | Consulted | Consulted | Responsible | Consulted | Consulted | Consulted |
R ResponsibleA AccountableC ConsultedI Informed· not involved
Option A
If there's no architect on the programme, I'd own the design end to end: voice capture, the link to your care record supplier, the pattern checks, the evidence pack and the gates, working closely with your safeguarding lead and quality team. I'd report to your programme board and your nominated individual.
Option B
If your partner, your care record supplier or your own team already has an architect, I'd take on the speech, pattern review and evidence pack under their lead, and give them a second senior opinion on integration, security and data protection.
Goes wrong
Drafts add detail the carer never said, and a wrong note becomes the record.
How the design avoids it
Drafts are built only from the carer's words and the care plan, guardrails block additions, and the carer confirms every note.
Goes wrong
Managers get a flood of flags and stop reading them.
How the design avoids it
Start with a few patterns agreed with managers, tune thresholds in shadow mode, and measure how many flags are useful.
Goes wrong
A safeguarding sign is missed because the AI didn't flag it, and people trust it too much.
How the design avoids it
The AI adds to, never replaces, the usual routes for raising concerns, and the safeguarding lead samples unflagged notes.
Goes wrong
Carers feel watched, and pattern review becomes a performance tool.
How the design avoids it
A written policy that it supports care, not discipline, agreed with staff before the pilot.
Goes wrong
The design assumes the care record supplier allows writing drafts back.
How the design avoids it
Agree the supplier route in discovery, and design a read-only fallback.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Tag evidence to the five key questions, which stay, so re-mapping to the final key lines of enquiry is a configuration change, not a rebuild.
Keep the log of drafts, flags and managers' decisions in a form you could show the ICO.
Build the business case on note quality and time given back to care, not on reducing staff.
Keep records structured and exportable from your care record, so new expectations on data sharing or reporting are an update, not a migration.
Expect your IT and care record suppliers' duties to change, and review contracts and supplier assurance alongside your DSPT.
Good answers are specific. Vague answers to any of these are worth following up.
Whether you need someone to lead the design or to support the architect you already have, let's talk it through.
