Can AI Improve Mental Capacity Decision-Making Support Without Replacing Professional Judgment?

Artificial intelligence creates an unusual opportunity in mental capacity practice. Used carefully, it could help care teams bring together fragmented information, translate complex choices into more accessible formats, identify gaps in records and prompt practitioners to consider questions they might otherwise overlook. Used badly, the same technology could turn a deeply human, decision-specific process into an apparently objective score or recommendation that quietly displaces the person and the professional responsible for the decision.

For adult social care providers exploring AI through the Digital Transformation in Social Care Knowledge Hub, mental capacity is therefore a particularly important test of digital maturity. The issue is not simply whether an AI system can analyse information. It is whether technology strengthens mental capacity, consent and best-interests decision-making while preserving the assumptions, safeguards and human accountability required by law.

That distinction should shape any use of AI and automation in care. AI may support preparation, communication, evidence retrieval, quality assurance and reflection. It should not become an artificial decision-maker deciding whether somebody has capacity, what is in their best interests or which restrictions should be imposed. This article examines where the boundary could sit, how providers might govern emerging applications and what meaningful assurance would look like in practice.

Mental capacity is a human legal judgment, not a prediction problem

The Mental Capacity Act 2005 applies in England and Wales and begins from principles that are difficult to reconcile with simplistic automated classification. A person must be assumed to have capacity unless it is established otherwise. They must not be treated as unable to make a decision until all practicable steps to support them have been taken without success. An unwise decision does not itself demonstrate incapacity. Where a decision is made for someone who lacks capacity, it must be in their best interests and consideration must be given to whether the purpose can be achieved in a less restrictive way.

Capacity is also decision-specific and time-specific. Someone may be unable to make one complex financial decision while being fully able to decide what they want to eat, where they want to spend the afternoon or whether they want a particular visitor. Capacity may fluctuate. Communication difficulties do not themselves establish incapacity. A person's diagnosis, disability, age, appearance or behaviour cannot substitute for the required assessment.

This is precisely where an apparently sophisticated AI model could create false confidence. A system trained on historical care records might identify characteristics associated with previous findings of incapacity. It could potentially recognise patterns involving dementia, learning disability, communication difficulty, brain injury or mental illness. But statistical association is not the legal test. If those patterns influence a recommendation about an individual, the technology risks reproducing assumptions that the Mental Capacity Act was designed to prevent.

The central operational boundary should therefore be clear: AI can help practitioners work with information, but responsibility for applying the Mental Capacity Act remains human. The practitioner still needs to understand the particular decision, the relevant information, the person's ability to understand, retain, use or weigh that information and communicate their decision, and the relationship between any difficulty and an impairment or disturbance in the functioning of mind or brain.

The stronger opportunity is supported decision-making

The most constructive use of AI may occur before anyone concludes that a person cannot make a decision. The Act requires practicable support, yet in real services the quality of that support can vary considerably. Information may be too complicated, discussions may happen at the wrong time, communication knowledge may be scattered across records, or staff may not recognise that a different presentation could enable the person to decide for themselves.

AI-supported systems could potentially help practitioners identify communication preferences from existing records, generate an initial plain-language explanation, suggest visual or alternative formats, organise the relevant issues into smaller steps, or remind staff about conditions under which the person communicates most effectively. For somebody whose capacity fluctuates, technology might help a practitioner identify recorded patterns suggesting that a decision could appropriately be revisited at another time.

None of those functions establishes capacity. Their value lies in strengthening the process through which the person is supported to exercise it.

This is consistent with a broader commitment to accessible information and communication. A provider could ask a much more useful digital question than “Does this person have capacity?”: “What do we know about how this person best understands information and communicates decisions, and what reasonable steps have we not yet tried?”

That reframing matters. Technology designed around prediction may encourage classification. Technology designed around support can potentially increase autonomy.

Scenario: making a tenancy decision accessible

A woman with a learning disability living in supported accommodation is considering whether to move to a different flat. Staff initially record that she appears unable to understand the tenancy implications because discussions about rent, bills, location and support arrangements quickly become confusing. She repeatedly changes the subject and says she does not know what she wants.

An AI-enabled documentation system does not assess her capacity. Instead, it retrieves information already recorded about her communication: she processes one issue at a time, understands photographs better than abstract descriptions and becomes anxious during lengthy formal meetings. It prompts the support worker to consider whether the information has been presented in a sufficiently accessible way.

Working with the woman, the team creates a visual comparison of the two homes, visits the proposed flat twice and separates the decision into understandable components. A housing officer explains the tenancy in simpler language, and an advocate is involved because the woman wants independent support. The discussion takes place in the morning, when she usually concentrates more effectively.

She is subsequently able to demonstrate sufficient understanding of the relevant information, consider important advantages and disadvantages and communicate a consistent choice.

The technology has contributed something valuable, but not by answering the legal question. It helped the team take practicable steps that enabled the person to make her own decision. That is a materially different model of AI-supported care.

AI could improve preparation without conducting the assessment

Capacity assessments can involve information distributed across care records, communication plans, professional reports, previous assessments, risk records and conversations with people who know the individual well. AI may eventually be particularly useful in organising this material before a practitioner reaches a decision.

For example, an authorised system operating within appropriate information-governance controls might identify previous communication strategies, locate evidence of fluctuating cognition, highlight contradictory records or summarise earlier decisions that appear relevant. It might prompt the assessor to clarify exactly what decision is being considered rather than relying on a generic statement that the person “lacks capacity”.

The distinction between retrieval and judgment is crucial. A summary generated from records may be incomplete or wrong. Earlier documentation may itself contain assumptions. A historical capacity assessment about one decision cannot establish incapacity for another. Information may have changed since the record was written.

AI output should therefore function as material for professional verification, not as an authoritative account of the person. Providers considering this type of capability can use the Digital Transformation Readiness Assessment to examine whether their data, workforce, governance and technology arrangements are sufficiently mature before introducing higher-risk AI-supported processes.

A service with inconsistent records, weak access controls and limited digital competence does not become safer because AI can summarise its data more quickly. It may simply generate unreliable conclusions at greater speed.

Professional judgment has to remain genuinely meaningful

It is easy to state that a human remains “in the loop”. The harder question is whether that human genuinely exercises independent judgment.

Automation bias can arise when people give disproportionate weight to a computer-generated recommendation, particularly where the system appears sophisticated or where staff are under time pressure. A practitioner presented with an AI-generated statement that somebody is “highly likely to lack capacity” may approach the subsequent conversation differently, even if they remain formally responsible for the decision.

That creates a governance problem. Human oversight cannot consist of clicking an approval box beneath a recommendation that the practitioner does not understand, cannot challenge or rarely contradicts.

Meaningful professional judgment requires the practitioner to be able to:

• understand what function the technology is performing and what it is not designed to determine;
• identify the evidence underlying any AI-supported output;
• recognise uncertainty, missing information and potential bias;
• disagree with or disregard an output without inappropriate organisational pressure;
• record their own reasoning independently; and
• escalate unexpected, unsafe or systematically misleading outputs.

This makes digital workforce competence a safeguarding and quality issue rather than simply a technology-adoption issue. Staff need enough understanding to question AI, not merely enough confidence to use it.

Best-interests decisions expose the limits of computational reasoning

Where a person lacks capacity for the particular decision, AI might again help organise information without determining the outcome. A system could retrieve recorded wishes and preferences, identify previous statements, collate relevant risk information or flag that consultation with particular people has not yet been documented. These functions could make evidence easier to find and reduce the risk that significant information is overlooked.

But a best-interests decision is not a calculation in which competing factors receive fixed numerical weights. The decision-maker must consider the relevant circumstances, the person's past and present wishes and feelings, beliefs and values, and the views of appropriate people, while avoiding assumptions based simply on characteristics or condition. The person should be involved as far as reasonably practicable.

The significance of those factors depends on the actual decision and the individual. A person's long-standing commitment to living near family cannot sensibly be converted into a generic preference score. Nor can an algorithm resolve a genuine disagreement between autonomy, safety, relationships, health and quality of life simply by ranking predicted outcomes.

The Positive Risk-Taking Planner can support teams in structuring consideration of autonomy, risk and least restrictive options without substituting a formula for accountable decision-making. The same principle should govern AI: structure can strengthen reasoning, but the responsibility for the decision remains with the appropriate human decision-maker.

Scenario: an AI summary misses what matters most

An older man with dementia is assessed as lacking capacity to decide whether to remain at home with an increased package of care or move permanently into residential care. An AI-supported record system summarises recent incidents: two falls, missed medication, night-time wandering and increasing calls to his daughter. Its generated overview understandably emphasises the risks associated with remaining at home.

The social worker reviewing the case notices that the summary contains little about the man's values. Earlier records show that he has repeatedly described his home and garden as central to his identity. His daughter explains that he cared for his wife there for many years and has consistently said he would accept substantial support to remain there. His reactions during visits also suggest that unfamiliar environments cause considerable distress.

The professional does not treat those wishes as automatically decisive, nor does she dismiss the risks. Instead, the multidisciplinary discussion considers whether additional night support, medication technology, falls interventions and family involvement could create a workable less restrictive alternative. The man's own responses are actively sought throughout.

The case illustrates an important limitation. AI can summarise what has been documented most frequently or most clearly, but that is not necessarily the information carrying greatest significance in a best-interests decision. Human judgment is needed not only to verify the data but to understand the person represented by it.

Bias could convert historical inequality into apparently objective evidence

AI systems learn from data, rules or combinations of both. In adult social care, historical information can contain unequal patterns of assessment, diagnostic assumptions, inaccessible communication, cultural misunderstanding and inconsistent recording. An AI system can reproduce those patterns without any intention to discriminate.

Consider records in which people with profound communication differences have historically been described more frequently as lacking capacity because practitioners did not have access to appropriate communication expertise. A model trained on those records might learn that communication characteristics predict incapacity. The apparent statistical relationship could then reinforce the original practice weakness.

Similar risks arise where cultural expressions, neurodivergent communication, mental illness, acquired brain injury or dementia are interpreted through data created for other purposes. AI-generated inferences can appear neutral because they are computational, but neutrality of presentation is not evidence of fairness.

This makes testing across different groups important, while itself raising information-governance questions about the use of sensitive personal information. Providers should understand what data a system processes, what inferences it makes, how performance has been tested and whether particular groups experience materially different outcomes.

The relevant principle is broader than technical model accuracy. A system could be statistically accurate overall while still being inappropriate for mental capacity practice. Providers need to consider equality, accessibility, explainability and whether technology risks undermining the presumption of capacity.

Information governance becomes more demanding when AI interprets care records

Mental capacity work frequently involves highly sensitive information: health conditions, cognitive functioning, family relationships, safeguarding concerns, behaviour, communication, finances and intimate details of daily life. Introducing AI changes the processing of that information even where the underlying care records already exist.

Providers therefore need to understand more than whether a supplier describes its product as secure. They need clarity about what information enters the system, why it is processed, where it is processed, whether it is retained, who can access it, whether it is used to improve or train models, what contractual controls apply and how outputs become part of the care record.

Health information and other special category data attract additional protections under UK data protection law. AI may also generate new inferences about an individual from existing information. Where processing is likely to create high risks to people's rights and freedoms, data protection impact assessment becomes particularly important.

The regulatory position around automated decision-making has also evolved following the Data (Use and Access) Act 2025. Providers should not rely on outdated summaries of Article 22 or assume that the existence of nominal human involvement resolves every issue. Restrictions remain particularly important where special category information is involved, and meaningful human involvement, transparency and safeguards require substantive consideration.

For social care, the safest operational principle is stronger than merely asking what automation is legally possible. Decisions concerning capacity, best interests and restrictions should remain visibly attributable to accountable human decision-makers. Digital records and information governance should make that accountability clearer rather than obscure it behind system-generated text.

CQC assurance would depend on practice, not the presence of AI

CQC does not require providers to use AI for mental capacity practice, and adopting an AI product would not itself demonstrate regulatory quality. The relevant questions arise through the quality of consent, person-centred care, safeguarding, governance, workforce competence, records and the way technology affects people's experiences.

A provider might demonstrate sophisticated digital infrastructure yet have weak Mental Capacity Act practice if staff routinely accept generated recommendations, capacity assessments are generic or people are not adequately supported to decide. Conversely, a provider using relatively simple technology may demonstrate strong practice through careful decision-specific assessment, accessible communication, defensible reasoning and meaningful involvement.

CQC assurance becomes stronger where different evidence sources tell a coherent story. Records may show how the person was supported; observation may demonstrate whether staff actually communicate accordingly; audits may identify recurring weaknesses; people and families may describe whether they feel involved; supervision may test practitioners' reasoning; governance reports may show whether concerns lead to improvement.

Providers can use the CQC Evidence Gap Analyzer to examine whether assertions about digital and mental capacity practice are supported by sufficiently varied evidence. This is particularly relevant where an organisation claims that AI has improved decision-making. The meaningful evidence is not the number of AI prompts generated but whether people receive better support, reasoning is stronger and inappropriate restrictions are avoided.

This aligns with CQC-related digital records and data assurance: technology should strengthen the reliability and accessibility of evidence without being mistaken for the quality of care itself.

Governance should define what AI is never authorised to decide

High-quality AI governance is not achieved by adding a general statement to an information-governance policy. Mental capacity requires explicit boundaries because the consequences of misplaced automation can affect where somebody lives, what care they receive, their finances, relationships, liberty and ability to take ordinary risks.

A provider introducing AI-supported functions should know which uses are authorised, which require additional approval and which are prohibited. A summarisation tool used to retrieve relevant information presents a different level of risk from a system generating a probability that a person lacks capacity. Technology proposing accessible wording differs from software recommending a best-interests outcome.

The provider should also define responsibility when an AI output is wrong. Accountability cannot disappear into a chain involving the frontline worker, Registered Manager, provider, software supplier and model developer. Operational managers need to know who verifies outputs, digital and information-governance leads need visibility of processing and supplier risk, and senior leadership needs assurance that high-risk uses are controlled.

The Governance Maturity Assessment offers a practical way to examine whether decision rights, delegated authority, escalation and board oversight are sufficiently developed for emerging technology. In this context, mature risk management and compliance should include the consequences of technology working as designed as well as the possibility of technical failure.

Scenario: a provider discovers automation bias through audit

A multi-service provider pilots an AI assistant that reviews capacity documentation and identifies possible omissions. The approved purpose is quality support: it can highlight where the specific decision is unclear, where practicable support is not recorded or where a best-interests record appears incomplete.

After three months, audit data shows something unexpected. Assessments receiving a high-risk AI flag are considerably more likely to conclude that the person lacks capacity, even though the tool was not designed to recommend an outcome. Interviews with staff suggest that some practitioners interpret multiple warnings as an indication that the case is inherently more likely to involve incapacity.

The provider pauses that element of the pilot rather than assuming more training alone will solve the issue. Quality and digital leads review the interface, wording and workflow. People with lived experience are involved in considering how the tool frames uncertainty. The supplier changes prompts so that they focus on support and missing evidence rather than risk scores, and practitioners receive case-based supervision about automation bias.

The pilot resumes with closer outcome monitoring.

The important governance evidence is not that the original system passed procurement checks. It is that the provider detected an unintended behavioural effect, understood its implications for people's rights, intervened and tested whether redesign changed practice. That is what responsible innovation looks like.

People should understand when AI is influencing their care

Transparency is especially important where technology operates behind professional processes. A person may reasonably believe that a social worker, nurse or care manager is considering their circumstances without realising that software has summarised their history, identified patterns or generated suggested wording that influences the discussion.

Providers need proportionate ways of explaining AI-supported processing that people can actually understand. A lengthy privacy notice is not equivalent to meaningful transparency. Information may need to be available in Easy Read, visual, audio or other accessible formats, and staff need enough understanding to answer questions rather than simply referring people to a supplier's technical documentation.

People should also have routes to challenge information. An AI-generated summary can become influential even when it contains an error. If inaccurate text is copied into a capacity assessment and subsequently repeated through later records, a single technological mistake can acquire the appearance of established fact.

This strengthens the case for co-production, choice and control in digital development. People drawing on support should not be involved only after systems have been purchased. Their perspectives can help providers understand whether explanations are meaningful, whether interfaces feel intrusive and whether proposed efficiencies actually improve the experience of making decisions.

Commissioners and system partners will need assurance about digital decision support

AI-supported mental capacity practice is unlikely to remain solely an internal provider issue. Local authorities, NHS organisations and other commissioners may encounter AI-generated material within assessments, reviews, safeguarding work and multidisciplinary decision-making. Different organisations may also use different systems, creating questions about provenance and accountability.

Commissioners do not need to prescribe one technology model, but they have legitimate interests in whether providers manage digital risk, protect information and maintain lawful person-centred practice. Procurement and contract monitoring may increasingly need to distinguish useful innovation from automation that weakens accountability.

A provider should therefore be able to explain what the system does, what information it uses, where human judgment occurs, how staff competence is assured, how errors are reported and what evidence demonstrates that the technology improves rather than distorts practice. The Commissioner Evidence Builder can help organisations structure this type of assurance where digital practice forms part of contractual or partnership discussions.

Integrated working adds another challenge. A capacity decision may involve care staff, social workers, clinicians, advocates and family members. AI cannot resolve differences in professional roles or legal authority. Teams still need clarity about who is the relevant decision-maker, what information can appropriately be shared and when disagreement requires escalation, independent advocacy, legal advice or ultimately consideration by the Court of Protection.

Workforce competence must include the ability to challenge technology

Introducing AI changes the competence expected of managers and frontline teams. Staff do not need to become data scientists, but they do need to understand the limitations of the systems they use. In mental capacity practice, that includes recognising hallucinated or unsupported content, distinguishing summaries from verified facts and avoiding the assumption that computational language is inherently objective.

Training attendance alone would provide weak assurance. Providers should be interested in whether practitioners can apply the Mental Capacity Act correctly when AI output is plausible but misleading. Supervision and case discussion can test whether staff challenge recommendations, return to original evidence and recognise when specialist input is needed.

Registered Managers have an important operational role but should not carry the entire governance burden. Digital leads may oversee system configuration and supplier relationships; information-governance specialists address data processing; quality teams review practice; safeguarding leads examine rights and risk; senior leaders determine organisational boundaries for higher-risk uses. The Registered Manager needs enough visibility to understand how the technology affects the service and to escalate concerns.

This is where workforce assurance should move beyond confirming that employees completed an AI module. Mature assurance examines actual reasoning, practice variation, confidence to challenge technology and whether digital tools are changing professional behaviour in unintended ways.

Scenario: AI helps prepare a complex decision but does not resolve disagreement

A man with an acquired brain injury wants to resume travelling independently by bus. His support team is concerned because he has recently become disorientated and has stepped into traffic on two occasions. His family believes independent travel should stop. He strongly disagrees and says the restrictions are destroying an important part of his life.

An AI-supported system helps the practitioner collate recent incident records, previous occupational therapy recommendations, the man's stated goals and information about successful journeys. It also identifies inconsistencies: some records describe him as lacking capacity to travel independently, while others record no formal decision-specific assessment.

The technology has exposed a problem but cannot resolve it.

The team undertakes a proper capacity assessment relating to the specific decision, using accessible information and practical discussion of the relevant risks. The man is supported to explain how he understands those risks and what strategies he would use. Occupational therapy input helps explore alternatives, including route training, a simplified journey, location technology that he is comfortable using and graduated practice.

Whether he ultimately has capacity for the relevant decision remains a matter for lawful human assessment. If he does, an unwise or risk-bearing choice cannot simply be converted into incapacity. If he does not, any decision made on his behalf still requires an individual best-interests process and consideration of less restrictive options.

AI has improved evidence retrieval and exposed weak recording. Professional judgment, supported decision-making and rights-based practice determine what happens next.

Board assurance should test outcomes and unintended effects

Boards and directors do not need to review individual capacity assessments routinely, but they should understand whether the organisation is introducing AI into high-impact decision processes and what controls accompany that use. A generic innovation update stating that a pilot is “performing well” provides little assurance.

Leadership information could instead examine where AI is deployed, reported errors, overrides, complaints, equality concerns, staff feedback, data incidents and findings from Mental Capacity Act audits. Particularly important would be evidence of whether patterns of capacity outcomes change after technology is introduced.

If one service suddenly records substantially more findings of incapacity after adopting AI support, that should trigger enquiry even if documentation quality appears to have improved. Equally, an increase in recorded supported decision-making may be positive, but boards should ask whether people's experiences confirm the apparent improvement.

The governance question is therefore not “How accurate is our AI?” in isolation. It is “What is this technology doing to decisions, professional behaviour and people's rights?”

This is consistent with stronger quality assurance and governance. Evidence should connect digital performance with human outcomes rather than treating system metrics as a separate technical domain.

The future could move from assessment automation towards decision accessibility

The most valuable future direction may be different from the one implied by much discussion of AI. Rather than attempting to automate judgments about incapacity, innovation could concentrate on making decisions more accessible to people.

Generative systems may become better at adapting explanations to different communication requirements. Multimodal tools could combine text, symbols, images and speech. With appropriate controls, AI might help practitioners explore different ways of explaining consequences or identify where information remains too abstract. Translation and communication technologies could widen participation where language or communication differences currently create barriers.

AI may also strengthen quality assurance. Systems could identify records where capacity language is generic, where the decision is not clearly specified, where a finding appears to rely heavily on diagnosis or where best-interests documentation lacks evidence of the person's wishes. Used as an organisational learning tool, this could help quality teams focus human review where it is most valuable.

These possibilities remain emerging. Their value will depend on accuracy, accessibility, evidence, information governance, procurement quality and how people themselves experience the technology. AI-generated accessible information also requires verification: simplifying information incorrectly can be as problematic as presenting it in an inaccessible form.

Over time, the strongest model may therefore be one of augmented supported decision-making. Technology helps people understand, helps practitioners retrieve and test evidence, and helps organisations identify weak practice. Humans retain responsibility for interpretation, judgment, relationships and accountability.

A practical boundary for responsible AI in mental capacity practice

The distinction between acceptable support and inappropriate substitution can be expressed through a simple governance principle: the closer an AI function comes to determining a person's legal status, rights or restrictions, the stronger the case for keeping that function outside automated decision-making altogether.

Lower-risk applications may include helping produce accessible drafts, locating relevant records, checking whether documentation addresses required considerations and identifying possible inconsistencies for human review. Risk increases where technology predicts incapacity, recommends best-interests outcomes, ranks people by decision-making ability or proposes restrictions.

Providers should not assume that adding a human approval stage automatically makes a high-risk system appropriate. The design question comes first: should the technology be generating that recommendation at all?

This also provides a useful procurement test. Suppliers should be able to describe intended use, limitations, data sources, validation, security, explainability and arrangements for errors and change. Providers should be wary of systems whose marketing promises exceed the evidence available or whose outputs cannot be meaningfully interrogated.

Digital audit and assurance should continue after implementation. Models, interfaces, provider workflows and staff behaviour can change. A safe pilot does not prove that a system will remain safe as it scales across different services or populations.

Conclusion

AI could make a meaningful contribution to mental capacity practice in adult social care, but its strongest role is unlikely to be deciding whether somebody has capacity. The Mental Capacity Act 2005 requires something fundamentally more individual than prediction: a presumption of capacity, practicable support to enable decision-making, attention to the specific decision and time, accountable assessment, and where necessary a person-centred best-interests process that considers wishes, feelings, values and less restrictive alternatives.

Technology can strengthen that framework. It can help retrieve evidence, identify gaps, improve accessibility, expose inconsistent recording and support organisational learning. It can also introduce bias, amplify poor historical practice, create misleading summaries and encourage professionals to defer to outputs that appear more objective than they really are.

The dividing line is therefore professional and organisational judgment. Providers need clear boundaries for AI use, meaningful human oversight, strong information governance, digitally competent staff and evidence that innovation improves people's ability to participate rather than quietly narrowing it.

The future opportunity is not an algorithm that decides who can decide. It is technology that helps more people make their own decisions, helps practitioners reason more carefully when they cannot, and gives leaders better evidence about whether rights-based practice is genuinely happening. That is a more demanding ambition for AI, but it is also far more consistent with the purpose of adult social care.