Agentic AI in Adult Social Care: Where Could Autonomous Digital Systems Safely Fit?

Imagine a homecare provider receiving an overnight notification that several visits may be affected by staff absence. Instead of simply alerting the coordinator, a digital system checks available staff, reviews travel constraints, identifies possible cover, prepares revised schedules and drafts messages for approval. A second system might monitor outstanding quality actions, gather relevant records and prompt managers when deadlines approach. Both go beyond displaying information: they begin to organise work.

This is the emerging territory of agentic artificial intelligence. Unlike conventional digital records or systems that respond to individual questions, agentic AI can pursue a defined objective through a sequence of actions, potentially interacting with multiple applications. For adult social care, the implications extend beyond technology procurement into the design of services, workforce responsibilities and organisational accountability.

The opportunities belong within the wider discussion of digital transformation in adult social care. However, the particular significance of AI and automation in care is that software may begin initiating activity rather than waiting for each individual instruction. That makes digital safeguarding and technology-enabled risk central to decisions about what these systems should be permitted to do.

The question is not whether providers should embrace or reject autonomy as a single proposition. It is where limited, observable and reversible digital action could improve services, and where decisions affecting people's rights, safety and relationships must remain under meaningful human control. This article focuses primarily on regulated adult social care in England, while recognising that the other UK nations operate under distinct regulatory frameworks.

What Makes Agentic AI Different From Existing Care Technology?

Most familiar digital care systems depend on people to initiate tasks. A care worker enters a record, a coordinator changes a rota, or a quality manager runs a report. Some systems automatically generate reminders or flag predetermined exceptions. Generative AI adds the ability to produce summaries, draft text or answer questions based on available information.

Agentic AI introduces a further capability: a system can be given an objective, determine intermediate tasks, use authorised digital tools and take actions towards completing that objective. An agent might retrieve records, compare information, generate correspondence, update a workflow and check whether a response has been received. Its behaviour may depend on changing information rather than following one fixed sequence.

These distinctions are important because autonomy is not binary. A system that drafts a rota for approval is materially different from one that publishes the rota, informs staff and reallocates visits without review. Similarly, identifying a potential discrepancy in a care record is different from amending the record or changing the person's support instructions.

Some products marketed as agents are conventional workflow automation with an AI interface. Others use language models to select actions dynamically. Neither label establishes safety or reliability. Providers need to understand the actual functions, permissions, data flows and failure modes of the particular system being considered.

The strongest early applications are likely to be those where the task is well bounded, the consequences of error are manageable, and an accountable person can inspect what happened. An autonomous system that organises routine administrative work may create genuine capacity. A system allowed to make unsupported judgements about risk, consent or care needs presents a fundamentally different proposition.

Where Could Autonomous Systems Add Operational Value?

Adult social care contains substantial coordination work that is essential but frequently fragmented. Staff availability, referrals, care reviews, equipment requests, quality actions, commissioner returns and training records may sit across separate systems. Managers spend time locating information, identifying omissions and asking colleagues to complete tasks.

Agentic systems could potentially reduce this friction by moving information between authorised workflows, preparing routine actions and drawing attention to exceptions. The most credible applications are not necessarily those involving direct care decisions. They may be the administrative processes surrounding care, where timely coordination creates more space for professional attention.

Potential uses include:

• Workflow coordination: preparing outstanding task lists, chasing approved administrative actions and routing work to responsible colleagues.
• Quality administration: assembling evidence for scheduled reviews and identifying overdue improvement actions.
• Workforce planning: producing draft cover options against agreed availability, competence and continuity constraints.
• Referral administration: checking whether required information has been received and preparing requests for missing documents.
• Service communication: drafting accessible updates for staff or families, subject to appropriate checks and permissions.

These possibilities connect with automation and operational workflow design, but benefits depend on process quality. Automating an unclear referral pathway can accelerate confusion. Giving an agent access to inconsistent workforce data can produce plausible but impractical cover arrangements. The process needs to be understood and improved before autonomy is introduced.

Providers should also distinguish productivity from outcomes. Reducing the time required to assemble an audit pack may be useful, but the more significant question is whether managers then spend additional time observing practice, speaking with people receiving support and resolving risks. Efficiency only becomes a care-quality benefit when the released capacity is directed towards meaningful work.

Designing Boundaries Around Digital Autonomy

A practical way to assess agentic AI is to classify proposed actions by consequence rather than technical sophistication. The same agent may safely undertake one task while requiring explicit authorisation for another. Permissions should reflect the consequences of an incorrect action, the sensitivity of the information involved and the availability of effective reversal.

Three operational boundaries are particularly useful. The first covers low-consequence administrative activity, such as sorting non-sensitive tasks or preparing an internal summary. The second covers actions that affect service operations, such as proposing rota changes or drafting information for external recipients. The third covers decisions directly affecting care, rights, safety, finances or access to services.

For the first category, narrowly scoped automation may be reasonable where activity is logged and exceptions are controlled. In the second, approval before execution will often be appropriate, especially where errors could affect continuity or confidentiality. In the third, AI should normally support evidence gathering or option development rather than exercise independent decision-making authority.

The distinction is not simply between a human being present and absent. A manager clicking approval without sufficient information may provide little meaningful oversight. Effective authorisation requires the ability to understand the proposed action, challenge its assumptions, inspect relevant evidence and reject it without pressure.

The Digital Transformation Readiness Assessment can help leadership teams structure an initial review of digital capability, governance and resilience before considering more autonomous applications. Such assessment supports organisational planning; it does not validate a particular AI system or certify its safety.

Human Accountability Cannot Be Delegated to Software

For regulated services in England, the Health and Social Care Act 2008 (Regulated Activities) Regulations 2014 remain relevant regardless of whether a task is undertaken through paper records, conventional software or an AI-enabled workflow. Regulation 9 concerns person-centred care, Regulation 11 addresses consent, Regulation 12 safe care and treatment, Regulation 17 good governance and Regulation 18 staffing. Their application depends on the service and circumstances, not on a special exemption for digital systems.

An AI agent is not a Registered Manager, a regulated professional or a substitute for the provider's legal responsibilities. Where a system produces an incorrect staffing recommendation or mishandles sensitive information, the organisation still needs to explain how the system was selected, controlled, monitored and corrected.

Responsibility should therefore be allocated before deployment. Operational managers need authority over service workflows. Information governance leads need visibility of data processing and access arrangements. Clinical or specialist practitioners should oversee relevant professional boundaries. Directors and boards need assurance that the overall risk remains proportionate to the expected benefit.

CQC's assessment approach may examine how providers deliver safe, effective, responsive and well-led care. AI-related evidence may arise through people's experiences, staff accounts, records, incidents and leadership oversight. There is no single agentic-AI document that guarantees regulatory compliance. The relevant question is whether the provider can demonstrate effective governance and safe practice.

That includes the capacity to stop a system. If managers cannot suspend an automated workflow promptly, identify what it has changed or restore a safe manual process, the organisation has not retained effective control merely because a person originally approved the purchase.

Safeguarding, Consent and the Limits of Automated Judgement

Safeguarding illustrates why apparently useful digital intelligence needs careful boundaries. An agent could identify repeated missed visits, unusual changes in care records or unresolved incident actions. It could bring information together more quickly than a manager reviewing several disconnected systems. Earlier visibility may support prevention, particularly where concerns develop gradually rather than through one obvious incident.

However, a pattern is not a safeguarding conclusion. Changes in behaviour, missed appointments or differences between recorded and expected activity may have several explanations. An automated alert may identify something worth exploring, but it cannot establish the person's wishes, understand every contextual factor or determine whether abuse or neglect has occurred.

Where a concern suggests immediate danger, existing emergency and safeguarding arrangements remain applicable. Where the Care Act 2014 section 42 criteria are met, the relevant local authority has responsibilities concerning safeguarding enquiries. An AI system should not independently decide that a concern falls below a referral threshold or close a potential safeguarding matter without appropriate human consideration.

The same principle applies to mental capacity. Under the Mental Capacity Act 2005, capacity is decision-specific and time-specific, and a person must be supported to make their own decision wherever practicable. An AI-generated assessment summary cannot replace the required decision-making process. Nor should a system infer incapacity from a diagnosis, communication difference or historical record.

The interaction with mental capacity, consent and best-interests decisions is particularly important where agents are designed to recommend restrictions, alter support arrangements or communicate decisions to families. Human review must consider the person's communication needs, available support, wishes and relevant legal safeguards.

Consent to care, agreement to use a particular technology and the lawful basis for processing personal information are related but distinct questions. A provider should not assume that consent to receive care automatically authorises the use of personal information in an autonomous AI service.

Operational Scenario: A Homecare Agent Preparing Emergency Cover

A domiciliary care provider supports an older person who receives two daily visits. The morning call includes assistance with personal care, breakfast and medicines support within the agreed care plan. During an overnight staffing disruption, a scheduling agent identifies that the allocated care worker will be unavailable and begins searching for alternative cover.

The system reviews authorised staff availability, geographical routes, recorded competencies and the person's established preferences. It produces two possible arrangements. One would preserve the usual visit time but use an unfamiliar worker. The other would involve a familiar worker arriving later than the agreed window. Rather than automatically choosing the option with the shortest travel time, the agent presents both to the coordinator and identifies the potential impact on the person's routine.

The coordinator checks whether the recorded competency information is current, contacts the person using their preferred communication method and confirms the relevant medicines-support requirements. The person explains that a later visit would create difficulty because of a planned appointment. The coordinator selects the earlier arrangement, ensures an appropriate handover and authorises the change.

The agent then updates the scheduling system and prepares staff notifications within its approved permissions. The coordinator checks that the messages have been issued and that the visit is subsequently confirmed. Any unresolved visit remains visible for escalation rather than being marked as covered merely because a worker was assigned.

At the next operational review, the Registered Manager examines whether the agent's proposals preserved continuity, whether competency data were accurate and whether people were involved in changes affecting them. The benefit is faster preparation of options, not automated determination of what matters most to the person.

Information Governance, Privacy and Cybersecurity

Agentic systems can create information governance risks beyond those associated with a conventional digital care record. A single agent may retrieve information from several applications, combine it into a new output, send it to another system and retain a record of its reasoning or activity. The greater the range of permissions, the greater the potential impact of a mistake or compromise.

UK GDPR and the Data Protection Act 2018 apply where personal data are processed. Health and care information commonly includes special category data, requiring both an appropriate Article 6 lawful basis and an Article 9 condition. Providers need to establish their responsibilities, including whether suppliers act as processors or independent controllers for particular activities.

A data protection impact assessment will be required where processing is likely to result in high risk to individuals' rights and freedoms. Given the sensitivity of care information, the scale and nature of access and the potential for consequential automated activity, providers should assess this carefully before deployment. A supplier's general privacy statement is not a substitute for understanding the provider's actual processing arrangements.

Agent permissions should follow the principle of least privilege. A system preparing a quality report may need read access to selected records, but it does not necessarily need permission to edit those records, contact external parties or access every person's full care history. Separation of duties matters: the ability to generate a recommendation should not automatically confer the ability to execute it.

Autonomous agents also introduce distinctive security concerns. Malicious or misleading instructions embedded in retrieved documents, emails or other data may attempt to redirect an agent's behaviour. This is sometimes described as prompt injection. The risk is particularly significant when a system can use tools, disclose information or modify records after reading untrusted material.

Providers should examine cybersecurity and digital resilience alongside supplier controls, authentication, access logging, incident response and system recovery. Sensitive information should not be transferred into unapproved consumer AI services simply because an employee finds them convenient.

There is also a distinction between automated assistance and solely automated decisions with legal or similarly significant effects. Article 22 of UK GDPR, as amended by the Data (Use and Access) Act 2025, has a specific framework for significant solely automated decisions, including additional safeguards and restrictions where special category data are involved. The legal position should be assessed against the actual decision and processing arrangements rather than assuming that every AI-generated recommendation is prohibited or that nominal human involvement always removes the issue.

Operational Scenario: An Agent Reviewing Supported Living Records

A supported living organisation introduces an agent that checks daily records for incomplete entries, unresolved actions and inconsistencies between scheduled and recorded support. The intention is to reduce the administrative time spent identifying documentation gaps across several services.

During a trial, the agent flags a person's repeated refusal of a planned community activity as a possible deterioration in engagement. It proposes an additional review and drafts wording suggesting that staff should encourage attendance more consistently.

The service manager recognises that this interpretation is incomplete. The person has recently expressed a preference for quieter activities and has been exploring different community opportunities. Their decision does not, by itself, indicate a problem requiring corrective intervention. The manager speaks with the person, checks their communication preferences and reviews the existing support plan.

The provider retains the useful administrative function but changes the agent's permitted behaviour. It can identify missing records and request human review of potentially significant patterns, but it cannot independently classify a person's choices as poor outcomes or amend person-centred goals. The organisation also tests whether its alerting rules disadvantage people who communicate differently or pursue unconventional routines.

In subsequent supervision, staff discuss the distinction between recording someone's choice and treating departure from an expected routine as a risk. The quality lead samples both flagged and unflagged records to assess whether the system is producing useful information or generating misleading confidence.

This example demonstrates the relationship between person-centred technology and professional interpretation. Digital consistency is valuable, but standardised expectations should not displace individuality, autonomy or positive risk-taking.

Workforce Competence, Professional Judgement and Digital Dependence

Agentic AI changes the work of managers and frontline teams even when it does not directly interact with people receiving support. Employees may increasingly review generated outputs, investigate exceptions, authorise proposed actions and identify when information is incomplete. These activities require judgement rather than simple familiarity with a software interface.

A care coordinator who can approve an agent-generated rota needs to understand what information the system has considered and what it may have omitted. A quality officer reviewing an automated incident summary needs to recognise unsupported conclusions. A Registered Manager needs to know when to suspend a workflow and escalate concerns about its reliability.

This creates a new dimension of digital workforce competence. Training should address practical use, confidentiality, error recognition, escalation, professional boundaries and the limits of automation. Completion of an online learning module is insufficient evidence that someone can safely oversee a consequential automated process.

Competence can be explored through observed exercises, supervised trials, realistic case discussions and reviews of decisions made using the system. Staff should be able to explain why they rejected an AI recommendation, not merely demonstrate that they know which button to press.

Organisations also need to consider workload displacement. A tool that saves ten minutes generating a report but creates twenty minutes of checking and correction may offer little benefit. Conversely, an agent that removes repetitive administrative steps may enable experienced workers to spend more time on supervision, relationship-based support and improvement activity.

There is a risk of automation bias, where employees accept a system's recommendation because it appears authoritative. The opposite problem is alert fatigue: staff become accustomed to excessive warnings and overlook a genuinely important exception. Both should be examined through workforce assurance, practical supervision and system performance reviews.

Providers should protect the ability to work safely when digital systems are unavailable. A workforce that has become dependent on automated scheduling, record retrieval or task allocation may struggle during a cyber incident or prolonged outage unless alternative procedures remain understood and tested.

Commissioning, Procurement and Supplier Accountability

Agentic AI will increasingly become a procurement and contract-management question. Providers may encounter these capabilities within existing care-management software rather than purchasing a separate AI product. A routine supplier upgrade could therefore introduce new automated functions, permissions or data flows that were not considered during the original procurement.

Before enabling such features, providers should understand the supplier's intended use, technical limitations, subcontracting arrangements, security controls, audit capabilities and approach to significant changes. Contractual responsibilities need to cover incident reporting, access to relevant records, data return, service continuity and exit arrangements.

The relationship between digital procurement and contract management and care quality becomes especially important where one system connects staffing, care records and quality reporting. A contractual service-level guarantee for software availability does not establish that automated recommendations are appropriate for individual care situations.

Local authority and NHS commissioners may reasonably seek assurance about material digital risks within commissioned services, although contractual requirements vary. Providers should be able to explain what the technology does, how people remain protected and how service outcomes are monitored. Commissioners should avoid assuming that claimed administrative savings automatically translate into lower safe delivery costs.

Where providers propose AI-enabled efficiency during tendering or service redesign, the Commissioner Evidence Builder can support the organisation of relevant performance, governance and outcome evidence. Claims about savings, continuity or improved responsiveness still require real operational validation.

Commissioning teams also have an interest in avoiding digital exclusion. People using direct payments, smaller providers and organisations serving rural or specialist communities may have different levels of digital infrastructure. The introduction of agentic functions should not make access to good-quality support dependent on a provider's ability to purchase the most advanced software.

Operational Scenario: Autonomous Quality Monitoring Across Multiple Services

A provider operating several residential and community services is considering an AI agent to support its quality team. The organisation already maintains incident logs, complaints records, audit results and improvement plans, but information is held in different applications. Quality managers spend considerable time compiling reports before they can examine emerging patterns.

The proposed agent would retrieve authorised information, identify overdue actions, compare trends and prepare draft summaries for operational review. It would also alert the responsible manager where a previously agreed corrective action appears to remain incomplete.

During testing, the agent identifies an apparent increase in medicines-related incidents at one service. The quality lead investigates and discovers that the increase partly reflects improved reporting following a staff development programme. The system has correctly identified a numerical change but cannot independently determine whether underlying safety has deteriorated.

The provider therefore requires every significant trend to be interpreted alongside incident severity, reporting culture, observation, staff feedback and people's experiences. The agent can prepare evidence and identify exceptions, but it cannot independently close an improvement action or certify that a service has become safer.

The organisation uses the Quality Dashboard Builder to structure the information reaching senior leaders. Board reports distinguish automated findings from verified management conclusions and identify actions that remain subject to human confirmation.

Over several months, the provider compares the agent's findings with independent quality reviews. It measures whether significant concerns are identified earlier, whether unnecessary alerts have increased and whether improvement actions are genuinely completed. The value lies in strengthening the quality team's capacity for interpretation and intervention, rather than replacing its judgement.

Building an Assurance Model That Tests Actions, Not Just Outputs

Traditional software assurance often focuses on availability, data security, functionality and user satisfaction. Agentic systems require additional attention because their actions may vary according to the information encountered and the tools available. Testing whether an agent can produce a convincing report is not the same as establishing that it will behave safely when records conflict or circumstances change.

A proportionate assurance model begins with an inventory of authorised agents and their functions. Each should have a named organisational owner, a defined purpose, specified access permissions and clear boundaries around what it can execute independently. Where a system can take external action, providers need to understand how authorisation, monitoring and reversal operate.

Testing should include ordinary workflows and difficult cases: incomplete records, contradictory instructions, unexpected requests, unavailable systems and attempts to obtain information outside permitted access. The organisation should also test what happens when an agent cannot complete a task. A safe system should expose uncertainty and failure rather than create the appearance that work has been completed.

Four forms of evidence are especially valuable:

• Permission evidence: records showing which actions an agent can perform, who authorised those permissions and when they were reviewed.
• Execution evidence: reliable logs identifying actions attempted, actions completed, human approvals and exceptions.
• Outcome evidence: comparison of service performance, errors, staff workload and people's experiences before and after introduction.
• Recovery evidence: demonstrations that workflows can be suspended, errors corrected and essential services continued safely.

These controls connect with digital audit and assurance, but documentation alone is insufficient. The strongest evidence includes observed testing, sampled transactions and confirmation from staff and people receiving support that the system is operating as intended.

The provider's internal controls and assurance arrangements should also address changes after deployment. Supplier updates, new integrations, altered prompts and expanded permissions may materially change behaviour. An agent approved for one limited purpose should not gradually acquire wider authority without renewed consideration.

The Governance Maturity Assessment offers a way to examine whether accountability, escalation and oversight arrangements are sufficiently developed to support such decisions. Boards should receive evidence about exceptions and actual consequences, not only the number of tasks automated.

Operational Scenario: An Agent Supporting Referral Coordination

A specialist supported living provider receives referrals from several local authorities. Each referral may involve assessments, funding discussions, housing considerations, staffing requirements and information from health professionals. Delays frequently arise because essential documents are missing or clarification requests remain unanswered.

The provider trials an agent that checks incoming referral information against an agreed administrative framework. It identifies missing attachments, prepares questions for the referrals coordinator and records outstanding responses. It can organise information and draft correspondence, but it has no authority to accept or reject a referral or determine whether the provider can safely meet someone's needs.

One referral concerns a person with an acquired brain injury who wishes to move closer to family. The agent identifies gaps in the available information about communication support and community access. Its draft questions focus on those gaps rather than making assumptions about the person's suitability for the service.

The referrals coordinator reviews the questions, contacts the commissioner and arranges a conversation involving the person and relevant professionals. The provider's assessment considers staffing competence, environmental requirements, funding and the person's own priorities. A manager makes the eventual service decision through the established governance process.

The trial reduces the time spent locating missing documents, but the provider also checks whether the system creates unnecessary information requests or disadvantages people whose circumstances do not fit standard referral categories. Feedback from commissioners and prospective residents informs revisions.

The distinction is fundamental: the agent supports the administration of assessment, while qualified and accountable people retain responsibility for interpreting needs, considering reasonable adjustments and making consequential decisions.

Co-Production and the Experience of People Receiving Support

Agentic AI may operate largely behind the scenes, but its consequences can be visible in people's daily lives. A scheduling agent may influence who visits and when. A records agent may affect how someone's choices are represented. A referral agent may shape the questions asked before a service begins.

Meaningful co-production therefore needs to go beyond asking whether people like a digital application. Individuals should have opportunities to influence which processes are automated, which decisions require personal involvement and how errors can be challenged. Accessible explanations are particularly important where technology affects people with communication differences, sensory impairments or cognitive disabilities.

The principles of co-production, choice and control remain relevant even where no one interacts directly with the software. A person may reasonably prefer familiar staff, wish to understand why a visit has changed or object to certain information being shared across systems. Such preferences should be considered within the applicable legal and operational framework.

Providers also need to consider the possibility of indirect discrimination. An agent trained or configured around typical patterns may produce less suitable recommendations for people whose routines, communication or cultural needs differ. Human review should be capable of identifying these problems rather than merely approving outputs that appear statistically consistent.

Complaints, advocacy, safeguarding concerns and ordinary conversations provide important intelligence about unintended consequences. The absence of formal complaints does not establish that people understand or are comfortable with automated activity affecting their support.

What Could Agentic AI Look Like Over the Next Five Years?

Agentic AI in adult social care remains an emerging capability rather than an established universal operating model. Some organisations already use automation, predictive analytics and generative AI features, but these should not be confused with fully autonomous, multi-system agents operating independently across care processes.

Over the next several years, a plausible development is the introduction of narrowly scoped agents within existing care-management, workforce and quality platforms. Rather than purchasing one autonomous system to run an organisation, providers may encounter separate capabilities for scheduling, administrative coordination, evidence retrieval and task management.

Interoperability will influence what becomes practical. Agents working across disconnected applications may introduce duplicated information, inconsistent permissions and unreliable results. Better interoperability and system integration could improve coordination, but it will also increase the importance of data governance and clear accountability between suppliers.

A second likely development is greater attention to continuous assurance. Providers may move from testing software mainly at procurement towards monitoring automated actions, exception patterns and system changes throughout its operational life. Independent testing and more mature supplier assurance could become important purchasing considerations, although specific future requirements should not be assumed before they are established.

There may also be opportunities for agentic systems to support workforce and capacity planning. For example, an agent could assemble alternative staffing scenarios using available data, while managers assess the implications for continuity, competence and affordability. Scenario modelling would remain a planning aid rather than a substitute for actual service assessment.

The more ambitious possibility is coordinated digital assistance across provider, local authority and NHS interfaces. Such systems might help identify outstanding discharge information, follow up agreed equipment requests or support administrative continuity between organisations. These developments would depend on lawful information sharing, compatible systems, contractual arrangements and reliable human oversight.

The strongest long-term opportunity is not unrestricted autonomy. It is a carefully governed division of labour in which technology handles defined coordination tasks while practitioners retain time and authority for relationships, interpretation, safeguarding and person-centred decisions.

Conclusion

Agentic AI raises a more demanding question for adult social care than whether artificial intelligence can produce useful information. It asks whether digital systems should be allowed to act, how far that authority should extend and who remains responsible when an action affects a person's support.

For providers in England, the most credible early applications are likely to involve bounded administrative coordination, evidence preparation and workflow support. These activities can create value where processes are understood, information is reliable and the consequences of error remain manageable. Greater autonomy over staffing changes, communications or operational decisions requires correspondingly stronger safeguards. Decisions involving consent, safeguarding, restrictions, care needs and individual rights demand meaningful professional and personal involvement.

The regulatory and governance foundations are familiar even if the technology is new. Providers remain accountable for safe care, person-centred practice, information governance, competent staffing and effective oversight. Registered Managers need practical visibility of how systems influence their services, while directors and boards need evidence that permissions, performance, exceptions and recovery arrangements are controlled.

Implementation will also depend on workforce confidence and the involvement of people drawing on care and support. An efficient automated process that weakens choice, obscures accountability or creates additional risk cannot be judged successful simply because it saves administrative time.

The future of agentic AI in adult social care should therefore be defined by proportionate autonomy: digital systems undertaking useful, limited actions within clear boundaries, with human responsibility remaining visible, informed and effective. The measure of progress will be whether those arrangements improve continuity, professional capacity, safety and the everyday experiences of people receiving support.