Skip to main content

Responsible AI

AI should support professional judgement, not replace it

Decisions at the Adult Social Care front door affect people's lives. EquiCare uses AI for the gathering, organising and drafting, and leaves every decision with qualified practitioners.

Clear boundaries

What AI does in EquiCare

Live

  • Converts speech to text, live and from recordings
  • Translates between English and 9 other supported languages
  • Organises what people say around Care Act outcomes
  • Suggests a triage outcome and an urgency, with reasons
  • Flags information that may indicate a safeguarding concern
  • Drafts summaries, records and reports for review
  • Suggests up to five local organisations for signposting
  • Runs the AI voice line for busy and out-of-hours periods

What AI does not do

  • Decide whether someone needs a needs assessment
  • Decide Care Act eligibility
  • Make or close safeguarding decisions
  • Decide what care and support someone receives
  • Send an outcome to a person without a practitioner's decision
  • Write to the formal record without practitioner approval

Safeguarding: a flag is not a decision

EquiCare can flag information that may indicate a safeguarding concern and explain why. The aim is for the right practitioner to see it sooner. The practitioner decides what it means and follows the council's safeguarding procedures.

AI can miss things and can flag things that turn out not to matter. That is why every request is reviewed by a person, not only the flagged ones. Flagged requests are prioritised so the most urgent are seen first.

Out of hours: if a caller to the AI voice line uses words that suggest risk, such as self-harm, neglect or immediate danger, the request is flagged as high priority and the detection is logged for safeguarding audit. A practitioner confirms the details of every voice-line request before it reaches triage.

The AI voice line is not an emergency service and flags are not monitored overnight. The voice line tells callers in immediate danger to ring 999, and out-of-hours requests, including flagged ones, are reviewed when the team is next working, usually the next morning. Councils keep their existing out-of-hours and Emergency Duty Team arrangements.

Which AI services, and what happens to the data

Speech-to-text, translation and language-model processing are provided solely by Microsoft through Azure AI Services and Azure OpenAI, hosted in Azure UK South.

EquiCare's AI features run on Microsoft Azure OpenAI and Azure AI Foundry. Details of the specific models and deployments are available to IG teams on request.

No data is shared with OpenAI or any other model provider, and no council data is used to train models.

EquiCare keeps recordings and transcripts according to each council's retention policy. If a council hasn't set its own retention, EquiCare keeps case records (assessments and notes) for 7 years, and call recordings and transcripts for 12 months. Each council can change these within platform-enforced limits, and audit logs are kept indefinitely in immutable storage. Call recordings from the voice line are stored in Microsoft Azure in the UK. Microsoft processes AI inputs and outputs transiently, apart from abuse-monitoring logs held for up to 30 days within Azure.

Before an EquiCare Notes recording starts, the person hears an automated message explaining the recording and why. Each recording is linked to its consent record.

An auditable process

EquiCare logs key activity, including sign-ins, record access, record creation and changes, AI-generated outputs, practitioner review and approval, and administrative changes. On request, we provide IG teams with EquiCare's audit-event catalogue. It lists every event the platform logs, including sign-ins, record access and changes, recording playback, exports, admin and permission changes, and retention and deletion actions. For each event it shows what is captured (who, what, when and which record) and confirms that audit logs are tamper-proof and never deleted.

AI features are governed for each council through EquiCare's audited Policy Pack: for example, the AI voice bot path is off by default and can only be switched on with JGME's written approval, and triage suggestions cannot currently be switched off per council while keeping intake.

We completed our AI assurance pack in October 2026. A council-ready DPIA template is in development. Until it is ready, we work directly with your IG, DPO and Caldicott Guardian on your own DPIA.

How we check the AI

Formal accuracy and bias testing of triage suggestions began in October 2026 and is due to complete before the pilot goes live. It uses synthetic test cases labelled by qualified social workers, with pass marks agreed in advance, and includes paired cases to check for bias by language, writing style, age and disability.

Results will be shared with the pilot council. Testing will be repeated after any change to the model or triage rules, and quarterly once live. During the first weeks of the pilot, triage suggestions will be compared with practitioner decisions before they are relied on.

Explicit crisis or safeguarding language automatically sets priority to High, with an urgent safeguarding recommendation that cannot be downgraded. Safeguarding concerns the AI identifies at high severity do the same. The system does not automatically refer to a safeguarding team; every result is reviewed by a practitioner, who makes the decision.

All AI output is advisory and labelled as such. Every suggestion is reviewed by a practitioner, who makes the final decision. AI performance will be reviewed against practitioner decisions during the pilot and quarterly after that. The audit-event catalogue shows how practitioner decisions are recorded.

Request an AI assurance discussion