AI Exercise for Older Adults: Coaching Within Clinical Limits

AI Exercise for Older Adults in the UK: Coaching Within Clinical Limits

AI can enlarge an exercise demonstration, suggest an easier variation or remind someone about a chosen walk. It cannot guarantee that an exercise is safe, diagnose falls risk from a phone camera or prescribe a clinical programme without assessment.

Older adults are not one risk group. A person may be a competitive athlete, returning after surgery, living with frailty, taking medicines that affect balance or using mobility equipment. The responsible service starts with the person’s goals and the limits of its intended purpose.

This guide is current to 31 July 2026. The UK Chief Medical Officers’ activity guidance is UK-wide, while NHS pathways, falls services and professional arrangements vary across England, Scotland, Wales and Northern Ireland. NICE guidance has a defined application. Confirm the nation, care setting and accountable clinician or provider. This is general operating guidance, not medical or exercise advice.

Name the service and the claims

Three products can look similar but require different evidence and control:

Write an intended-purpose statement covering users, excluded users, setting, target outcome, who selects exercises, necessary supervision and emergency limitations.

The MHRA’s guidance on intended purpose for software as a medical device explains that purpose determines risk and evidence. Marketing a camera as detecting falls risk or prescribing treatment can create a medical purpose even if the interface says “wellness.”

Do not call a product a physiotherapist. The HCPC’s standards for physiotherapists cover safe and effective professional practice, assessment and exercise or movement skills. An algorithm may support a registered professional; it does not hold registration, form a therapeutic relationship or accept professional accountability.

Record every content version and clinical approver. Do not let a general-purpose model invent exercises, contraindications or recovery advice during a session.

Start from current UK movement guidance

The UK Chief Medical Officers’ physical activity guidelines were updated on 10 July 2026. For older adults, they emphasise daily activity, building gradually, strength, balance and flexibility on at least two days a week, an aim of 150 minutes of moderate aerobic activity across the week, and breaking up prolonged sedentary time where possible.

Those are population guidelines, not an automatic weekly prescription. The report recommends initial trained supervision for some higher-risk people.

An activity app should ask the person to choose a meaningful goal—walking to a shop, gardening, standing for a task, attending a class or maintaining strength—then offer a small step. Allow rest, lower intensity and non-standing alternatives. “Some is better than none” is a safer starting point than shaming a person for missing 150 minutes.

Track what the person chose, completed and found difficult. Do not reward risky intensity or streaks that discourage rest during illness.

Include clear stop guidance in the approved content and an accessible route to a professional when symptoms, recent change, injury or confidence make self-directed exercise unsuitable. The app should not diagnose the symptom or tell someone to ignore it.

For mainstream coaching outside a falls pathway, see AI in UK gyms and performance coaching.

Keep falls prevention inside the assessed pathway

NICE’s falls guideline NG249 distinguishes people who need comprehensive falls assessment, gait and balance assessment or general activity advice. It recommends comprehensive assessment and management for specified groups, including some people with frailty, injury, loss of consciousness, inability to get up or recurrent falls.

A model should not collapse this into one risk score. Falls can involve medicines, blood pressure, vision, cognition, continence, footwear, hazards and underlying conditions. A gait clip cannot evaluate all of them.

Within an established falls service, encode the assessed programme:

• exercise name and purpose;
• starting level and equipment;
• supervision and support;
• approved progression or regression;
• frequency and review point;
• symptoms or events that stop the session;
• contact and escalation; and
• responsible professional.

NICE defines supervised exercise as regular contact that can reassess performance, correct technique, progress or regress exercise and motivate the person; not every exercise must be observed. Technology can make that contact easier, but should not silently progress difficulty between reviews.

After a fall, do not issue a generic “balance workout.” Follow the person’s agreed plan for injury, loss of consciousness, inability to rise or other urgent concerns. Record the event and make it easy to contact the service or emergency route.

Measure falls and near-falls carefully. Increased activity can change exposure, and self-report may be incomplete. Track confidence, functional measures chosen by the service, programme adherence, adverse events and referrals, with an appropriate comparison and duration.

Use camera feedback as a prompt, not a diagnosis

Phone video can estimate joint points or count a defined repetition under favourable conditions. Clothing, walking aid, camera angle, lighting, occlusion, tremor and non-standard movement can produce confident errors.

Use it only for a bounded task, such as “the full body was not visible” or “this repetition may need review.” Do not infer neurological disease, frailty, pain or fall probability unless the product has the required intended purpose, evidence and professional pathway.

Validate with people who reflect the intended population, including different bodies, skin tones, mobility aids, clothing, home layouts and movement strategies. Do not score a safe adaptive technique as failure merely because training video showed one idealised form.

Report:

• unusable-session rate;
• repetition count error;
• false unsafe-form alerts and missed agreed hazards;
• performance with chairs, frames and sticks;
• user comprehension and distress; and
• reviewer correction time.

Keep an “unable to assess” state. If the camera falls, freezes or sees only part of the person, pause feedback instead of guessing. A person must be able to exercise without recording where the service can safely support that choice.

Never use remote video as proof that someone is independent or safe at home. It shows a narrow moment and can create false reassurance for relatives or services.

Personalise with consent and explainable rules

Useful personalisation comes from declared capacity, preference and the professional plan—not inferred age stereotypes.

Store:

• activities the person enjoys;
• mobility and communication support they choose to share;
• available space and equipment;
• preferred session length and reminders;
• approved difficulty range;
• language, captions and display needs; and
• people authorised to support the account.

Explain a suggestion in ordinary language: “You chose seated movement today” or “Your clinician approved this easier variation.” Avoid “the AI detected decline” when the evidence is missed sessions or uncertain video.

Allow the user to reject, correct and delete preferences. Do not reduce options permanently after one tired session. Review after illness, hospital attendance, medicine change, fall, new pain or equipment change through the agreed pathway.

Design for large text, clear audio, captions, high contrast, simple navigation, slow demonstrations and enough time. Provide printable or telephone-supported alternatives. Digital access, confidence and dexterity should not determine access to falls prevention.

Make social matching optional and safeguarded

A group or exercise partner can make activity enjoyable, but similarity of a score does not establish compatibility or safety. Do not promise to “combat isolation” from a match.

Let people choose group, one-to-one, known-contact or no-social modes. Match only on attributes they agree to share, such as activity type, approximate pace, format and general availability. Do not disclose address, diagnosis, exact age, fall history or live location.

Verify organisers appropriately, moderate messaging and provide report, block and exit controls. Prevent financial solicitation and requests to move immediately to unmonitored channels. Publish safeguarding routes and staff response times.

For residential or home-care contexts, coordinate with the accountable provider and the person’s plan. A virtual partner is not supervision, a welfare check or emergency response. For the wider setting, see AI and technology in UK elderly care.

Protect health, video and caregiver data

Gait, fall history, exercise tolerance and inferences about physical condition can be health data. The ICO’s special-category data guidance explains that health data includes information revealing a person’s health state.

Identify controller roles among app provider, clinic, care provider, employer or family account. Establish lawful basis and special-category condition, give a usable privacy notice and complete the necessary impact assessment.

Default to on-device or short-lived pose processing where feasible. Do not retain room video just because storage is cheap. Separate health records from social profiles and analytics. A family member should see only what the person has authorised, with capacity and safeguarding handled under the relevant framework.

Do not sell movement or vulnerability data, use it for insurance or advertising, or repurpose it to assess care staff. Give access, correction, deletion or restriction routes where applicable, and preserve clinical records according to policy.

Apply the NCSC’s secure AI development guidance: strong authentication, least privilege, protected model updates, encrypted transfer, dependency control and incident response. Test account recovery that does not lock out a user with limited digital access.

Build physical and operational failure modes

Before each session, show space, footwear, equipment and support checks appropriate to the approved content. Demonstrations should not obscure a trip hazard or require a phone to be positioned dangerously.

Define offline behaviour. If connectivity or camera feedback fails, either continue with a pre-approved non-camera version or stop safely; never change to a new exercise. Keep emergency contacts available outside the model.

Record adverse events, near-falls, pain-related stops, dizziness, confusion, safeguarding reports and delayed professional contact. Review them by content and model version.

A measurable 90-day pilot

Pilot a low-risk feature, such as delivering clinician-approved seated strength and balance reminders to one community cohort. Do not begin with automated falls diagnosis or unsupervised progression.

Days 1–30 — define and baseline

• write intended purpose, exclusions, supervision and escalation;
• confirm professional, device, privacy and safeguarding responsibilities;
• approve fixed content and accessible alternatives;
• baseline uptake, completion, support calls, confidence and adverse events; and
• test devices, aids, languages and home conditions.

Days 31–60 — shadow personalisation

• generate recommendations for professional review only;
• sample accepted, rejected and “unable to assess” outputs;
• test camera occlusion, low light, lost connection and account recovery;
• review performance across relevant user groups; and
• rehearse fall, symptom, safeguarding and data-incident pathways.

Days 61–90 — limited release

• let users choose among approved options within their plan;
• prohibit autonomous progression and diagnostic claims;
• review adverse events and unresolved support daily;
• audit caregiver and staff access weekly; and
• obtain clinical, safeguarding, accessibility, privacy and security sign-off.

Release only when 100% of delivered exercises come from the approved plan and version, at least 95% of users can start and stop independently or with intended support, no progression occurs without required review, adverse-event rate does not exceed baseline tolerance, camera uncertainty triggers “unable to assess,” and every escalation and offline drill completes.

Pause after a fall or injury plausibly influenced by advice, delayed urgent assessment, exercise outside the plan, missing accessibility route, coercive social contact, unauthorised health or room-video access, failed escalation or unapproved content/model update. Revalidate after health pathway, population, exercise library, device, evidence, law or intended-purpose change.

The practical verdict

AI can make an approved activity easier to understand and adapt. It cannot certify safety or replace the assessment behind a falls or rehabilitation programme.

Start small, respect the person’s own goals and keep progression, uncertainty and escalation visible. Movement can support health; responsible coaching never turns that truth into a guarantee.