No statutory decision
Aria never determines eligibility, capacity, best interests, deprivation of liberty, safeguarding outcomes, Continuing Healthcare eligibility or charges.
Safeguards and governance
Aria prepares. People decide. Always.
These are the rules we hold ourselves to in a statutory service: what the model may do, what it may never do, when it must stop, and how a council can prove all of that to an inspector.
Prototype status · the practice framework is drafted and will be owned by a registered social worker before any pilot goes live
Six rules that do not bend
A named human owns every output
Every draft, referral and summary carries the name of the practitioner accountable for it, and the audit trail shows who changed what.
Stop and hand over
Risk, safeguarding, liberty and capacity triggers interrupt the flow immediately and route to a human — with everything Aria heard attached.
Transparency by default
People are told they are talking to an AI in the first message, and can ask for a human at any point without giving a reason.
Data minimisation
We collect what the conversation needs and nothing more. Production runs under a completed DPIA, with UK data residency.
Publish the failures
Model incidents are logged, contained and reported to the practice lead and programme board. The pause control is documented, not theoretical.
Escalation
Triggers that always hand over to a person
These are non-negotiable. When one appears, Aria stops, tells the person what is happening, and creates an immediate notification with an owner.
| Trigger | Aria does | Route and timing |
|---|---|---|
| Disclosure or suspicion of harm, neglect, abuse or exploitation | Stops the flow, captures the concern verbatim, tells the person what happens next | Safeguarding lead, immediate (notification N3) |
| Immediate danger to life | Instructs the person to call 999 and stays with the conversation | Emergency services, immediate |
| Deprivation-of-liberty indicators | Records AGNI-consistent evidence over time; never concludes | DoLS / LPS team, immediate (N9) |
| Question about a person's mental capacity | Captures supported decision-making and routes onward | Qualified MCA assessor (N4) |
| Carer in crisis, or unable to continue | Flags urgency, offers contingency and immediate contact | Carer services and duty lead, immediate (N18) |
| Self-harm or suicidal thoughts | Provides crisis numbers, flags for urgent human contact | Duty lead, immediate, with crisis signposting |
| Any request for a human being | Offers a callback or a direct line, no reason required | Duty team, immediate |
| Model incident: an output that is wrong or unsafe | Incident record and containment; pause available | Carely.care practice lead and PSW, within one hour (N19) |
Oversight
Who is accountable for the model
The practice framework is the document that governs Aria's behaviour: conversation standards, required questions, prohibited outputs, escalation rules, language and accessibility standards, and the QA sampling regime.
It is written to be owned by the council's Principal Social Worker, not by us. In a pilot, the PSW can inspect it, amend it, and audit against it — and the framework changes when they say so, not when it suits us.
- A registered social worker owns practice quality (Carely.care practice lead)
- The council's PSW holds veto over practice-affecting changes
- Quality assurance samples Aria's drafts monthly, with published findings
- Every AI output is attributable, versioned and retained for audit
- An independent review of the model's impact is commissioned at pilot end
Incident and pause controls
- Pause in one step. A named council officer can suspend Aria's assessment drafting without a change request.
- Graceful degradation. If the model is paused, the front door reverts to structured forms and human triage — the service continues.
- Containment. Incidents are triaged within one hour, with affected outputs identified and re-reviewed by a person.
- No silent learning. Nothing a person tells Aria in a live service is used to train a model.
- Full audit. Inputs, outputs, versions and human overrides are retained so an inspector can reconstruct any decision.
Compliance posture
What we align to, and where we actually stand
We would rather state the position accurately than claim certifications we do not yet hold.
| Standard or framework | What it means here | Status |
|---|---|---|
| UK GDPR and Data Protection Act 2018 | Lawful basis, minimisation, retention and rights for everything the front door handles | Design complete; DPIA completed before any live pilot data |
| DHSC guidance on AI in adult social care | Safeguards, transparency and human oversight expectations | Aligned by design; published in the practice framework |
| Mental Capacity Act 2005 and the AGNI judgment | Supported decision-making, multifactorial evidence, no determination by the model | Reflected in flows, escalation rules and the MCA/DoL pathway |
| Care Act 2014 | s.4 information, s.9 assessment, s.10 carers, s.13 eligibility, s.42 safeguarding | Mapped in the practice framework |
| NHS Data Security and Protection Toolkit | Security and information governance evidence for health-adjacent flows | Assessment planned as part of pilot onboarding |
| Cyber Essentials Plus | Independently tested controls for the service and its supply chain | Targeted before handling live council data |
| WCAG 2.2 AA | Accessibility of every public and professional surface | Built to AA, including keyboard operation and reduced-motion support |
| CQC local authority assessment framework | Evidence that supports the council's own assessment, not a substitute for it | Designed to produce audit-ready evidence as a by-product |
Data protection
What we do with information
In the prototype: nothing you type into Aria is stored or transmitted. The demonstration runs entirely in your browser with synthetic data.
In this website's forms: we store your name, email, organisation, area and message so we can reply to you, and we notify the founding team. We do not sell it, share it, or add you to a marketing list. See the privacy notice.
In a live service: a DPIA is completed before go-live, data stays in the UK, retention follows the council's own schedule, and every AI output is attributable and auditable. Nothing said to Aria in a live service is used to train a model.
Interrogate us
Ask the hard governance questions before you consider a pilot. We would rather answer them now than after a procurement.