When technology gets in the way

There are many reasons for slow adoption of the stethoscope — among which are lack of formal education, hesitancy of the patient and physician to have an instrument placed between them, and lack of opportunities for continuing education after leaving medical school.

Reinhart (2020)

The stethoscope was resisted not because clinicians were ignorant but because it placed something between the clinician and the patient.

Frontier AI capabilities

• Long-context windows (1M tokens): Ingest and reason across very large datasets
• Multimodal: Interpret figures, scans, audio, and video in real time
• Deep research: Conduct multi-source evidence synthesis
• Tool use: Search databases, run code, read PDFs, navigate the web, build prototypes
• Domain reasoning: Postgraduate level across most disciplines (expertise on demand)
• Agentic workflows: Plan and execute autonomously over house; task horizon doubling every ~7 months (Kwa et al., 2025)

AI in clinical practice

• Diagnosis and triage: matching or exceeding specialist performance across multiple clinical domains (Tu et al., 2024)
• Reducing expert-novice gaps: consistent high performance regardless of clinician experience level (Dell'Acqua et al., 2023)
• Complementary error profiles: AI errors and human errors are different errors (Lenskjold et al., 2023)
• Communication: generating responses that patients rate as more helpful and empathetic than physician responses (Ayers et al., 2023)

AI in physiotherapy practice

• MSK diagnosis: LLMs performing comparably to specialist clinicians on clinical reasoning tasks (Hao et al., 2025)
• Movement analysis: gait analysis and categorisation from mobile phone video (Reddy, 2025)
• Documentation: AI scribing increasing direct patient interaction time by 24%, appointments 8% shorter (NHS TORTUS trial, 2025)
• Rehabilitation: AI-generated personalised programmes adapted from patient history (Michou, 2026)
• Remote physiotherapy: AI-guided remote patient management with ~85% patient satisfaction (Flok Health)

Sanctuary strategies

• Denial: "AI can't replicate the therapeutic relationship"
  Fails to engage with the evidence

• Retreat: "Focus on what only humans can do"
  Defensive, and an ever-decreasing surface

• Restriction: "Don't let clinicians use it in practice"
  They will use it because it adds value

• Resignation: "Our roles will inevitably diminish"
  Learned helplessness dressed as realism

All position you as a spectator of something happening to you.

What AI systems lack

• Stateless: language models have no memory between sessions; you are a stranger every time, regardless of how many conversations you've had
• Static: knowledge is frozen at training; no ongoing awareness of your field, your specialisation, or developments since the training cutoff
• Contextless: no understanding of your expertise, clinical interests, patient population, or professional reasoning patterns
• Context is the bottleneck: model intelligence may be less important than context

Context sovereignty

Context sovereignty is a goal: practitioners control their professional context, shaping AI behaviour toward personally meaningful outcomes (Rowe & Lynch, 2025)

• Your CPD, specialisations, and clinical interests structured for AI
• The system accumulates context through interaction over time
• Your context works with any AI provider

What value do I bring? What value will I always bring?

What are agents?

• Memory: retains context across time: your history, preferences, clinical patterns, previous decisions
• Planning: sets goals and sequences actions to achieve them, not just responding to prompts
• Action: interacts with external systems: retrieves records, drafts notes, books follow-ups
• Reflection: evaluates its own outputs and adjusts

From a tool you use to a system acting on your behalf.

Distributed cognition

• Clinical reasoning has never been a purely individual act (expert performance has never only been about what's in your head)
• Expertise is distributed across people (MDT, supervision, peer consultation), artefacts (notes, imaging, guidelines), and environments (clinical setting, protocols, institutional knowledge)
• Managing this distributed cognitive system:
  • knowing when to trust each component
  • how to integrate conflicting inputs
  • how to take responsibility for the synthesis
• AI changes the composition of that system, not its nature

Professional context today

• What does an AI know about your specialist area and clinical interests?
• What does it know about the patient populations you work with?
• What does it know about your CPD trajectory and the gaps you're working on?
• What would it need to know to give you genuinely useful clinical support rather than generic outputs?

Patients are already using AI

• Patients routinely use AI to research diagnoses, interpret symptoms, prepare questions, and seek second opinions — before the appointment
• Some arrive having already formed a view of their situation
  • Reminder: Patients rated AI responses as more helpful and more empathetic than physician responses (Ayers et al., 2023)
• The instinctive clinical response treats this as a problem to manage, rather than an opportunity to embrace

These are patients exercising agency over their own health.

From passive observation to active support

Conventional approach: react when patients arrive with AI-generated information

Emerging approach: support better AI use before the encounter

• Provide patients with structured prompts: "Before your next appointment, ask your AI: what questions should I be asking about my progress with [condition]? What should I understand about the next phase of treatment?"
• This is health literacy for the AI age — patient activation extended into a new domain

Patient context sovereignty

While you're developing your professional AI ecosystem, patients will be developing relationships with their personal health AI that you can't assume you will have access to.

• Patients increasingly have persistent AI agents tracking health history, symptoms, concerns, and goals
• These agents accumulate context over time — just as yours does
• Key implication: the patient may arrive with richer AI-accessible context about themselves than you have in your clinical record

Multi-stakeholder relationships

The multidisciplinary team is already a negotiation between different professional contexts, expertise sets, and institutional positions.

• All stakeholders bring their own agents into the MDT
• Each agent carries its principal's values and priorities into the encounter
• The encounter is therefore a negotiation between context-rich positions

This doesn't replace the therapeutic relationship. It changes what the relationship is.

Putting it into practice

AI works from context and your context is the difference between generic outputs and meaningful support.

What you can do now:

• Articulate your professional context explicitly: your specialisation, patient population, CPD priorities, clinical reasoning patterns
• Give patients structured prompts for pre-session AI preparation
• Bring evaluative judgement to AI outputs — not deference, but engaged, contextual assessment