Predictive Safeguarding: Can Data Help Prevent Harm Before It Happens in Adult Social Care?

A missed visit, an unexplained withdrawal of cash, a change in behaviour or an unusually high level of staff turnover may mean very little when viewed alone. Safeguarding concerns often become clearer only when separate events are connected: repeated medication omissions, increasing isolation, unexplained injuries, changing relationships, deteriorating self-care or several low-level incidents involving the same service, person or pattern of practice. By the time that pattern becomes obvious, harm may already have occurred.

This creates an important future direction for the Safeguarding Knowledge Hub: whether better use of information can strengthen prevention rather than merely improve retrospective investigation. Developments in safeguarding prevention and early intervention increasingly intersect with digital records, analytics and artificial intelligence. At the same time, digital safeguarding and technology-enabled harm remind providers that technology can create risks as well as reveal them.

Predictive safeguarding should therefore not mean an algorithm deciding that somebody is being abused, that a family member presents a threat or that a care worker is unsafe. Its more credible role is to identify patterns that warrant human attention earlier. This article examines how that model could develop in adult social care in England, how it interacts with Care Act safeguarding, CQC assurance and provider governance, and where safeguards are needed to protect autonomy, privacy and human judgement.

Safeguarding prevention begins before a safeguarding concern is raised

Safeguarding is sometimes operationally framed around what happens after a concern is recognised: immediate protection, reporting, referral, information sharing, enquiry, investigation and learning. Those processes matter, but prevention begins much earlier. It includes service culture, staffing stability, supervision, risk management, accessible communication, relationship-based practice and the ability to notice when something has changed.

Under the Care Act 2014 framework in England, safeguarding is concerned with protecting an adult's right to live in safety, free from abuse and neglect, while promoting wellbeing and taking account of the person's views, wishes, feelings and beliefs. Local authorities have statutory safeguarding responsibilities, including the section 42 enquiry duty where the relevant statutory conditions are met. Provider responsibilities sit alongside, rather than replace, those local authority functions.

Predictive approaches could strengthen the preventive end of this system. Instead of waiting for a threshold event, organisations could examine whether combinations of information suggest that risk is increasing. That might involve trends in incidents, complaints, staffing, missed calls, financial concerns, restrictive interventions, unexplained injuries or changes in an individual's behaviour and circumstances.

The distinction is fundamental. A predictive signal is not a safeguarding finding. It is an invitation to look more closely. Mature organisations would preserve that distinction throughout their policies, systems, training and governance rather than allowing an apparently precise risk score to acquire authority it does not deserve.

The strongest opportunity lies in connecting weak signals

Adult social care already generates substantial safeguarding-relevant information. The problem is often that evidence sits in different places. An incident may be recorded in one system, a complaint in another, staff absence elsewhere and concerns raised during supervision in narrative records. At individual service level, people may know fragments of the story without anyone seeing the pattern.

Predictive intelligence could help connect those weak signals. It might identify that one person has experienced several apparently unrelated incidents, that medication omissions are increasing on a particular shift, or that complaints about staff conduct coincide with high agency use and declining supervision completion. None of those relationships establishes abuse. They may, however, justify earlier managerial examination.

This makes information sharing, confidentiality and disclosure central to predictive safeguarding. Organisations cannot assume that safeguarding importance permits unrestricted collection or sharing of personal information. Data protection, confidentiality, purpose and proportionality remain relevant, alongside safeguarding duties and legitimate information-sharing requirements.

Providers can use the Digital Transformation Readiness Assessment to examine whether their information governance, data maturity, workforce capability, digital strategy and cyber resilience are sufficiently developed for more advanced analytical approaches. Predictive capability built on fragmented governance can magnify existing weaknesses rather than resolve them.

Prediction cannot replace professional curiosity

Safeguarding frequently depends on noticing something that does not fit. A person's explanation may change. A care worker may sense that someone is unusually anxious when a particular visitor is present. A family member may suddenly insist on speaking for a person who previously communicated independently. An apparently accidental injury may not fit the account of how it occurred.

These observations involve context, relationships and professional curiosity. Data can strengthen them but cannot fully reproduce them. An algorithm sees what has been recorded and what its design allows it to interpret. It may not see fear, coercion, subtle communication, cultural context or the significance of a change that is meaningful only because a worker knows the person well.

The operational risk is therefore two-sided. Staff may ignore a meaningful concern because a predictive system classifies risk as low, or they may treat a high-risk classification as proof that abuse is occurring. Both responses weaken safeguarding judgement.

Strong implementation establishes that frontline observations can override digital reassurance and that predictive alerts require proportionate human review. Supervision and safeguarding competency assessment can explore whether staff understand these boundaries. Training attendance alone cannot show that a worker can distinguish an automated warning from a safeguarding conclusion.

Scenario: financial changes that do not tell the whole story

A supported living provider uses analytical software to identify unusual combinations of safeguarding-relevant events. A man with a learning disability begins withdrawing substantially more cash than usual. Records also show that he has cancelled two planned activities and recently asked staff not to contact his sister about his finances. The system identifies a possible financial-abuse pattern and prompts managerial review.

A poor response would be to assume exploitation, restrict access to money or contact family automatically. Instead, a senior worker speaks with him privately using communication he understands. He explains that he has started a relationship and is saving cash to buy an expensive present. He does not want his sister involved because she disapproves of the relationship.

The conversation also reveals something requiring further exploration: his new partner has repeatedly asked him for money. The worker discusses his choices, what healthy relationships can look like and what would make him feel uncomfortable. His capacity is not questioned merely because others consider the spending unwise. With his agreement, support is adjusted so he can discuss financial decisions without staff controlling them.

The alert has been useful because it created an opportunity for conversation. It did not establish abuse. Subsequent safeguarding action would depend on what further information emerged and the circumstances involved. The person's autonomy remains central while the provider stays alert to coercion or exploitation.

Making Safeguarding Personal creates an essential boundary

Predictive safeguarding can easily become system-centred rather than person-centred. An organisation may become so focused on preventing an adverse event that it loses sight of what the person wants to happen. Safeguarding is not simply the elimination of all risk; adults retain rights, relationships, preferences and, where they have capacity for the relevant decision, the ability to make choices others may regard as unwise.

Making Safeguarding Personal provides an important counterweight to purely risk-driven models. The person's desired outcomes, communication needs and experience should influence how concerns are explored and what proportionate responses look like. Prediction should create an opportunity for better engagement, not justify bypassing it.

The Mental Capacity Act 2005 remains central in England where capacity and decision-making are relevant. Capacity is decision-specific and time-specific. An algorithm cannot infer incapacity from a diagnosis, pattern of risk, communication difference or repeated decisions that professionals dislike. Where there is reason to doubt capacity for a particular decision, the appropriate human process remains necessary.

The Positive Risk-Taking Planner can support structured consideration of autonomy, potential benefits, risks and safeguards where difficult decisions arise. Predictive information may contribute evidence, but the framework becomes credible only when information is interpreted alongside the person's wishes, rights and individual circumstances.

Predictive safeguarding can operate at person, service and organisational level

The term can imply prediction about individuals, yet some of the most valuable applications may concern organisations rather than people. A provider could examine whether certain combinations of workforce, quality and incident information precede safeguarding deterioration. Increasing agency dependence, management vacancies, overdue supervision, rising complaints and declining audit results may together indicate a service becoming less resilient.

This does not mean unstable staffing inevitably produces abuse or neglect. It means organisational conditions can change the probability that poor practice goes unnoticed or that staff become unable to provide consistent support. Predictive intelligence could direct quality resources towards services showing combinations of early warning indicators rather than waiting for a serious incident or inspection outcome.

At organisational level, potentially useful domains might include:

• changes in incident type, frequency or severity;
• complaints, whistleblowing and concerns raised by people or families;
• staff turnover, vacancies, sickness and agency dependence;
• supervision, competency and management-capacity indicators;
• restrictive interventions, medication errors and unexplained injuries; and
• overdue safeguarding, audit or improvement actions.

The purpose is not to construct a league table of supposedly dangerous services. It is to recognise where several pressures are moving in the wrong direction and decide whether additional support, audit, leadership attention or safeguarding scrutiny is justified. That approach can strengthen safeguarding audit, assurance and board oversight by moving governance from isolated metrics towards connected intelligence.

Scenario: the risk sits in the service, not one member of staff

A multi-service provider's monthly dashboard shows no single safeguarding indicator outside its normal tolerance for one residential service. There have been two medication errors, one complaint about delayed personal care, several minor falls and increased agency use. Staff supervision is also beginning to fall behind. Viewed separately, each issue appears manageable.

An analytical model identifies that the combination resembles patterns seen before periods of quality deterioration elsewhere in the organisation. The service is flagged for human review rather than automatically classified as unsafe.

The operational director and quality lead visit the service. Conversations with people receiving support and staff reveal that an experienced deputy recently left and the Registered Manager is covering both rota gaps and management duties. Permanent staff are spending less time mentoring unfamiliar agency workers, and daily handovers have become rushed. There is no evidence that one worker is deliberately causing harm, but conditions for neglect and missed care are becoming more plausible.

The provider responds by strengthening management cover, reducing avoidable administrative demands, prioritising recruitment and restoring structured handovers and supervision. Existing incidents are reviewed for any safeguarding implications, with appropriate referrals made where necessary. Senior leaders monitor whether staffing stability, people's experiences and quality indicators improve.

Predictive safeguarding has therefore supported prevention without inventing an allegation. It has identified an organisational vulnerability early enough for leadership action.

Bias could turn safeguarding intelligence into discriminatory suspicion

Predictive systems learn from existing information. Safeguarding data, however, do not represent an objective record of all harm. They represent harm that was noticed, recorded, reported, referred and categorised. Differences in reporting practice, access to services, professional judgement and organisational culture can therefore become embedded within datasets.

If historical data show more recorded incidents for a particular population, service type or demographic group, a model may learn that association without understanding why it exists. Higher reporting could reflect greater underlying risk, stronger recognition, greater service contact or different thresholds. Conversely, populations experiencing under-reporting may appear artificially low risk.

This creates an equality and human-rights issue as well as a technical one. Predictive systems should not become mechanisms through which disabled people, people from minority communities or people with unconventional lifestyles are subjected to disproportionate monitoring because historic patterns have been translated into automated suspicion.

Human review therefore needs to ask why an alert has arisen and whether the variables used are legitimate. Organisations should examine differences in false alerts, missed concerns and interventions between groups where data allow. Where a model cannot be meaningfully explained or tested, leaders should be cautious about allowing it to influence consequential safeguarding decisions.

Digital surveillance can itself become a safeguarding concern

Technology capable of predicting harm may depend on increasingly detailed information about people's lives. Sensors can record movement. Telecare can detect routines. Digital care systems can analyse behaviour. Financial technologies can identify transactions. Location systems can show where someone has travelled. Each may support legitimate care or safety objectives, but together they can create an extensive architecture of surveillance.

The fact that monitoring could prevent harm does not automatically make it proportionate. A person living in supported housing does not lose their right to privacy because technology makes continuous observation possible. Providers need clarity about purpose, lawful processing, access, retention, security and the extent to which people understand the monitoring taking place.

Consent also requires careful handling. Agreement given because someone believes monitoring is a condition of receiving support is different from meaningful choice. Where capacity is relevant, decision-making requires the appropriate individual consideration rather than blanket assumptions based on disability or diagnosis.

This is where positive risk-taking and least restrictive practice become especially important. The safest technical configuration is not necessarily the most rights-respecting support arrangement. A mature provider considers whether monitoring is necessary, whether less intrusive alternatives exist, what benefit it produces and whether the arrangement remains proportionate as circumstances change.

Information sharing remains a human governance decision

Predictive safeguarding becomes more powerful as information from different organisations is connected, but information sharing cannot be treated as a purely technical integration problem. Social care providers, local authorities, NHS organisations, police and other partners hold information for different purposes and under different responsibilities.

Safeguarding can provide compelling reasons to share relevant information, particularly where serious risk exists, but organisations still need proportionate decision-making. Staff should understand when consent is relevant, when information may need to be shared without consent, who can authorise decisions and how reasoning is recorded. Complex cases may require appropriate professional or information-governance advice.

Multi-agency safeguarding becomes stronger when information enables collective understanding rather than simply being transferred between inboxes. Predictive systems might help identify relationships across organisations, but responsibility for interpreting and responding to those relationships remains with accountable people and agencies.

The same principle applies to local authority safeguarding processes. Provider analytics can support recognition and referral, but they do not determine whether statutory enquiry duties are engaged or replace local safeguarding arrangements. Where immediate protection is necessary, staff should follow established safeguarding and emergency pathways rather than waiting for analytical certainty.

Scenario: repeated low-level incidents across organisational boundaries

An older man receives homecare and regular support from relatives. Over several months, different organisations hold apparently minor pieces of information. The homecare provider records that food is sometimes unavailable. A community professional notes that he appears anxious about money. The provider also records two occasions when a relative cancels visits at short notice and one comment from the man that he is worried about being a burden.

No individual observation proves neglect or financial abuse. A shared safeguarding intelligence process identifies the accumulation of concerns and prompts proportionate review. The homecare provider's safeguarding lead examines its own records and ensures the man has an opportunity to speak privately. He explains that a relative has taken control of his bank card and that he does not know how much money remains in his account.

The provider follows the relevant safeguarding process and shares necessary information through established arrangements. The man's wishes and communication are central to subsequent discussions, alongside consideration of immediate financial protection and any capacity issues relevant to particular decisions.

The important feature is not that an algorithm has “predicted abuse”. The system has connected information that was previously too dispersed to reveal the pattern clearly. Human conversation then produces the information needed for proportionate safeguarding action.

CQC assurance will depend on practice, not predictive sophistication

For CQC-regulated adult social care services in England, predictive safeguarding would sit within existing expectations around safeguarding, risk, person-centred care, governance, staffing and learning. Purchasing analytical technology does not itself demonstrate that people are safer.

CQC assurance is strengthened where policy, frontline practice and leadership oversight align. Staff should know how to recognise and report concerns even when technology identifies nothing. Managers should understand what predictive alerts mean and what response is expected. People should be able to describe whether they feel safe and listened to. Incidents, complaints, safeguarding records and governance information should show that concerns lead to appropriate action and learning.

The CQC Evidence Gap Analyzer can help providers examine whether their evidence base connects safeguarding arrangements with implementation and outcomes. This is particularly relevant where digital systems create extensive audit trails: a record showing that an alert was opened does not establish that the subsequent judgement or response was appropriate.

Under CQC-related risk, safeguarding and restrictive practice, evidence may need to be triangulated across people's experiences, staff competence, records, incidents and leadership oversight. Predictive technology becomes one component of that evidence architecture, not a substitute for it.

Frontline competence remains the decisive safeguard

Predictive systems can help prioritise attention, but staff still need safeguarding competence. Workers require the confidence to recognise abuse and neglect, listen to people, preserve evidence where relevant, report concerns and act immediately where someone is in danger. They also need to understand coercion, capacity, consent, professional boundaries and the ways in which harm may present differently between individuals.

This makes safeguarding training and competency more important rather than less important in a data-rich environment. Staff should understand that a low-risk score never cancels a direct concern. Equally, a high-risk score does not authorise them to investigate beyond their role or confront an alleged perpetrator in ways that could increase danger.

Competence can be tested through supervision, observation, case discussion, reflective practice and examination of actual responses to concerns. Managers can explore whether staff understand escalation routes and whether they feel psychologically safe to challenge colleagues or leaders. A provider with sophisticated analytics but a culture in which workers fear speaking up remains vulnerable.

Registered Managers therefore need assurance about both sides of the system: whether technology is operating as intended and whether people remain capable of acting independently of it. Digital confidence should include knowing when the technology may be wrong.

Boards need to understand what the model is missing

Board assurance around predictive safeguarding should not become a presentation of increasingly sophisticated risk scores. Directors and trustees need to understand the assumptions behind the information they receive, what data are absent and where uncertainty remains. A dashboard can make weak information appear more authoritative simply because it is presented consistently.

The Governance Maturity Assessment can support examination of risk ownership, delegated authority, escalation and organisational assurance. For predictive safeguarding, governance should establish who can change thresholds, who reviews model performance, how significant errors are escalated and when concerns become organisational risks.

Boards may benefit from a limited set of connected assurance questions:

• Are predictive alerts leading to proportionate review and action?
• Where are staff overriding the system, and what does that reveal?
• Are false alerts or missed concerns concentrated in particular services or groups?
• Do people experience monitoring as supportive, intrusive or restrictive?
• Are recurring safeguarding themes connected with workforce or quality pressures?
• Can leaders demonstrate that completed actions produced sustained improvement?

This supports stronger decision-making and escalation. The board does not need access to every safeguarding record. It needs sufficient intelligence to understand whether controls are effective, whether significant risks are being surfaced and whether executive action follows when patterns indicate wider organisational vulnerability.

Commissioners will need to distinguish intelligence from profiling

Commissioners may increasingly encounter predictive safeguarding within provider assurance, contract monitoring and service design. Used carefully, aggregated information could identify pressures affecting whole markets: workforce instability, recurring service failures, increasing safeguarding concerns or geographic gaps in preventive support.

The danger is that predictive information becomes a simplistic provider risk score. Historical safeguarding numbers can be misleading without context. A provider with a strong reporting culture may record more concerns than an organisation where staff under-report. Higher numbers do not automatically indicate weaker safeguarding, just as low numbers do not establish safety.

Commissioner assurance should therefore consider reporting culture, seriousness, themes, response quality, action closure and evidence of learning alongside volume. Where algorithms influence contract scrutiny or intervention, providers should be able to understand the evidence being considered and correct inaccurate information through appropriate processes.

The Commissioner Evidence Builder can support providers in structuring evidence across contractual commitments, performance, safeguarding and improvement. Predictive intelligence is strongest when it prompts informed dialogue and proportionate assurance rather than creating opaque classifications that neither commissioner nor provider can meaningfully interrogate.

Learning systems could become predictive rather than purely retrospective

Safeguarding learning commonly follows incidents, enquiries, complaints or investigations. Organisations examine what happened, why controls did not prevent it and what should change. Predictive safeguarding creates the possibility of using that accumulated learning prospectively.

If repeated reviews show that certain organisational conditions commonly precede harm, quality systems can monitor those conditions earlier. For example, recurring safeguarding events might be associated with weak handovers, unstable leadership, deteriorating staffing continuity or failures to complete agreed actions. The lesson is not that any one factor predicts abuse with certainty; it is that combinations of vulnerabilities deserve attention.

This strengthens the relationship between safeguarding and root cause analysis and thematic learning. Organisations can move beyond identifying immediate causes towards understanding recurring system conditions. Learning then influences workforce planning, supervision, digital design, commissioning discussions and investment rather than remaining within individual incident action plans.

Quality teams can use the Quality Dashboard Builder to connect safeguarding themes with relevant workforce, quality and outcome measures. The purpose is not to create more indicators. It is to make relationships between existing information visible enough for leaders to decide where deeper enquiry is warranted.

Scenario: the model creates a safeguarding risk of its own

A homecare organisation introduces an automated risk model using incident records, missed calls, complaints and selected care-record information. Within several months, managers notice that people living alone with cognitive impairment are disproportionately classified as high risk. The system responds by recommending more frequent welfare checks.

At first, the recommendations appear cautious and protective. A quality review, however, finds that some people are receiving additional unplanned monitoring despite no clear evidence that this improves their safety. One woman says she feels staff are constantly checking on her and that the additional visits make her feel less independent.

The provider pauses automatic use of the recommendation and reviews the model with its supplier, safeguarding lead and information-governance function. It examines whether living alone and cognitive impairment have been weighted in ways that convert ordinary characteristics into assumptions about vulnerability. People's experiences are included in the review.

The organisation introduces stronger human authorisation before monitoring can increase, clearer documentation of purpose and periodic review of whether additional observation remains necessary. It also tests whether similar bias affects other groups.

The case illustrates why digital audit and assurance must examine consequences rather than merely system performance. A technically functioning model can still produce poor care if its outputs create disproportionate restrictions. Predictive safeguarding therefore needs safeguarding from its own unintended effects.

The future is likely to be predictive assurance rather than automated safeguarding

Over the next several years, adult social care providers are likely to have access to increasingly sophisticated ways of analysing digital care records, workforce information, incidents, complaints and operational data. AI may make it easier to identify relationships that conventional dashboards overlook. Remote monitoring and greater interoperability could add further sources of intelligence.

The most credible direction is predictive assurance rather than automated safeguarding. Systems may help organisations decide where human attention is most needed, identify services showing emerging instability or recognise combinations of events around an individual that merit conversation. Decisions with significant consequences should remain subject to accountable human judgement.

This model also requires restraint. Not every available dataset should be combined simply because technology permits it. Organisations need to understand the purpose of analysis, whether information is sufficiently reliable, how long it is required and what happens when the system produces an alert. Cybersecurity and business continuity become safeguarding issues where essential protective processes depend on connected infrastructure.

People drawing on care and support should also influence the design of these systems. Co-production can explore which forms of monitoring feel supportive, what explanations people need, how they can challenge inaccurate information and what boundaries they expect around privacy. A predictive system designed solely by technology, risk and compliance teams is unlikely to capture the full meaning of safety in people's lives.

The strongest future systems may therefore be those that become better at uncertainty rather than pretending to eliminate it. They will identify patterns, show why attention may be needed and make it easier for people to investigate intelligently. They will not confuse statistical association with evidence that abuse has occurred.

Predictive safeguarding changes what good evidence looks like

As safeguarding becomes more data-informed, assurance needs to distinguish four levels of evidence. First, an organisation may show that information was collected. Second, it may demonstrate that a pattern was identified. Third, it may evidence that somebody reviewed and acted on that intelligence appropriately. The strongest level shows whether the action improved safety, respected the person's wishes and remained effective over time.

This distinction prevents digital activity from becoming a proxy for safeguarding quality. Thousands of automated alerts demonstrate system activity, not protection from harm. High referral volumes demonstrate reporting, not necessarily prevention. Completed actions demonstrate process, not necessarily improved experience.

A mature evidence architecture therefore triangulates records, direct observation, people's feedback, safeguarding outcomes, complaints, workforce intelligence and leadership oversight. Where an algorithm contributes to a decision, the organisation should retain enough information to understand how that contribution affected practice. Where it was wrong, learning should reach system design as well as individual staff.

This is particularly important where predictive models change over time. Supplier updates, altered data sources or changes in service populations may affect performance. Digital systems should therefore remain within continuous quality assurance rather than being treated as finished implementations once procurement and training are complete.

Conclusion

Predictive safeguarding could strengthen adult social care in England by making emerging patterns of harm, neglect and organisational vulnerability visible earlier. Its greatest contribution is unlikely to be predicting with certainty that abuse will occur. It is more likely to lie in connecting weak signals that would otherwise remain separated across care records, incidents, complaints, workforce information and organisational boundaries.

That opportunity requires careful limits. Safeguarding is fundamentally concerned with people, rights, relationships and proportionate protection. A risk score cannot establish abuse, determine capacity, replace professional curiosity or decide what outcome a person should want. Nor should prevention become a justification for continuous surveillance, unnecessary restriction or discriminatory profiling.

Strong implementation therefore depends on the relationship between technology and human systems. Frontline workers need safeguarding competence and confidence to challenge digital outputs. Registered Managers need clear operational controls. Safeguarding and quality leads need connected evidence. Directors and boards need assurance about bias, exceptions and outcomes. Local authorities, commissioners and system partners need intelligence that supports proportionate action without confusing prediction with proof.

The strongest future model is not automated safeguarding. It is earlier, more intelligent and more person-centred prevention: using data to ask better questions sooner, while ensuring that accountable humans remain responsible for listening, investigating, deciding and learning. Predictive safeguarding becomes valuable when it expands the opportunity to prevent harm without narrowing the rights and freedoms that safeguarding exists to protect.