Clinical Competency in the Age of AI

Safe, equitable, and effective practice in AI-augmented healthcare

Shoshana Bloom Founder and Principal Consultant, Equiti Health Ltd
April 2026

Executive Summary

AI is being deployed into clinical practice faster than the competency structures needed to make that deployment safe.

This white paper reports the findings of a systematic narrative synthesis of 445 studies on AI-related clinical risk. It identifies four structural gaps in existing competency provision, proposes a five-domain framework to address them, and sets out the implications for workforce and education bodies, regulators, clinical educators, and clinical governance teams.

The Problem

When clinicians use AI tools repeatedly - over months and years - without structured competency to support that use, the evidence shows that clinical skills atrophy, reasoning patterns degrade, and professional accountability erodes. This is not a technology failure. It is a workforce competency failure. The tools are being deployed; the competencies to use them safely have not been defined, specified, or made available to clinicians.

The Evidence

A synthesis of 445 studies published 2021–2026, including the first multicentre real-world evidence of clinical skill attrition through AI use - a documented fall in unassisted diagnostic accuracy among clinicians working with AI assistance over three months. 73% of included studies were published in 2025 or 2026, reflecting the pace at which the risks are becoming visible.

Four gaps no existing framework addresses

Gap 01

Deskilling prevention

When clinicians rely on AI to perform tasks they once did themselves, they gradually lose the ability to perform those tasks independently. For trainees entering AI-intensive environments, the foundational skills may never develop at all. No existing framework makes preventing this an assessable clinical competency.

Gap 02

Algorithmic bias at point of care

AI tools trained on unrepresentative data produce systematically less accurate outputs for certain patient groups. Detecting this bias requires clinical skill exercised in the consultation room, not just awareness that bias can exist. No existing framework makes this a user-level clinical competency.

Gap 03

Governance literacy

Many health organisations lack documented clinical safety assurance for their deployed AI tools. Clinicians carry full personal professional accountability for decisions they inform - but no framework currently equips them to ask whether a tool is safety-assured, or what to do when it is not.

Gap 04

AI disclosure and shared decision-making

Patients have a right to know when AI has contributed to clinical decisions about their care. Supporting that conversation in accessible language - preserving the patient's ability to engage meaningfully - requires specific clinical skill. No existing framework addresses this.

What is proposed

A five-domain competency framework, differentiated across three career stages - Foundation, Practitioner, and Senior/Leadership - that addresses each of these gaps and other identified gaps directly. The framework establishes, for the first time, what clinicians at each stage of their career should be able to do when working with AI tools: not simply what they should know about AI, but how they should practise safely with it.

The framework is built on an architecture that connects individual clinical risk, systemic governance risk, and population-level health equity risk - and specifies the clinician's distinct role at each level.

This is the first framework to define these competencies in operational, assessable form - specifying not what clinicians should be aware of, but what they should be able to demonstrate at each career stage.

The Problem: AI in Clinical Practice Without the Competencies to Match

The policy discourse around clinical AI has been shaped largely by questions of performance: can AI match or exceed the diagnostic accuracy of a consultant radiologist? Can it identify deteriorating patients before the clinical team can? These are legitimate questions and the evidence on many of them is encouraging. They are not, however, the questions that determine whether AI-augmented healthcare is safe and equitable. The question that determines safety is different: what happens to the clinician who uses an AI tool repeatedly, over months and years, without actively maintaining their skills and protecting their independent clinical judgement?

Key Finding

The evidence shows that AI deployment without appropriate clinical competency in place does not supplement clinical reasoning. Over time, it erodes it.

In 2025, the first multicentre real-world evidence of this erosion was published. Budzyñ et al. reported a fall in unassisted adenoma detection rates from 28% to 22% among endoscopists working with AI assistance over three months. The skill had not been assessed during this period. It had not been exercised. In the presence of an AI tool that reliably performed the task, it had atrophied.

This finding is not isolated. A synthesis of 445 studies identifies seven distinct risk mechanisms through which AI deployment, without protective competency structures, degrades clinical capability, erodes governance, and concentrates harm in the communities least able to detect or challenge it. The workforce using these tools has received, at most, awareness-level orientation to AI as a technology. What AI does to clinical reasoning, professional judgement, and the patient-clinician relationship over time? On this, clinicians have received almost no training.

445

studies synthesised, 73% were published 2025–26

existing frameworks reviewed - all have critical gaps in competencies

70%+

of UK health organisations lack documented clinical safety assurance for digital health technologies used

Key Finding

No reviewed framework connects individual clinical competency to the three levels of risk that AI deployment creates. Without that architecture, clinicians cannot distinguish which risks they are primarily responsible for mitigating, which require organisational action, and which require policy-level intervention. This is the analytical failure this framework is designed to address.

What the Evidence Shows: Three Levels of Risk

AI introduces clinical risk at three distinct levels. The distinction between them is not one of severity - it is one of origin. That distinction determines who is responsible for what, and why individual clinical competency alone cannot resolve risks that arise at the systemic or population level.

At Level 1, risks are generated within individual clinical encounters by the psychology, reasoning, and behaviour of the clinician. They accumulate invisibly, one consultation at a time, and are only detectable through the patient outcomes they eventually affect. Individual clinical competency is both the cause and the mitigation. At Level 2, risks arise from how institutions configure the conditions in which those encounters take place - whether AI tools in clinical use have been validated, monitored, and safety-assured. Individual clinicians are embedded in these systems but are not their primary agent. At Level 3, risks are not experienced in individual encounters at all. They emerge from the accumulation of individual and institutional failures across a system, concentrated in the communities least able to detect or challenge them.

The architecture below is interactive. Click any risk term to see how it works and why it matters.

Figure 1 · The Clinical AI Risk Architecture

Click any risk term to see its definition

Level 3 — Population Health Risk

Risks visible only in aggregate, concentrated in communities least able to detect or challenge them

Awareness + Advocacy

No individual clinician encounters these risks directly. They emerge from the accumulation of individual and institutional failures across a system — concentrated in the communities already most disadvantaged.

Compounded Discriminatory Bias AI-augmented Digital Exclusion Tiered Consent

Governance failure → amplifies L3 · Moral disengagement → erodes L2 oversight

Level 2 — Systemic Clinical Risk

Risks arising from how institutions build — or fail to build — safe AI deployment conditions

Awareness + Escalation

Individual clinicians cannot generate or resolve these risks. They arise from procurement, governance, and oversight failures. In many health systems, the majority of organisations lack documented clinical safety assurance for their deployed AI tools.

Governance Failure Human Oversight Erosion Algorithmic Bias × Governance Failure

Deskilling + automation bias feed L2 · Cognitive load intensifies all levels

Level 1 — Individual Clinical Risk

Risks generated in individual clinical encounters by the psychology, reasoning, and behaviour of the clinician

Full Ownership

Invisible to standard performance management. Accumulate one consultation at a time. Only detectable through the patient outcomes they eventually produce. Individual clinical competency is the primary and sufficient mitigation.

Deskilling Mis-skilling Never-skilling Automation Bias Cognitive Load Amplified Automation Bias Moral Disengagement Therapeutic Relationship Erosion Efficiency Rebound

Five competency domains — responses to this risk architecture

Technical AI Literacy

Clinical Judgement Preservation

Algorithmic Bias + Health Equity

AI Safety + Governance

Professional Agency

Bloom (2026) · Equiti Health · Systematic narrative synthesis of 445 studies, 2021–2026

The architecture above establishes something that sounds straightforward but has significant practical consequences. The three levels are not a spectrum of severity - they are three fundamentally different types of risk, with three different primary agents of mitigation. At Level 1, only the clinician can solve the problem: no governance process, procurement standard, or training programme can substitute for a clinician who exercises sound, independent judgement in the presence of AI. At Level 2, only the institution can solve the problem: individual clinicians cannot fix absent safety documentation or inadequate monitoring. At Level 3, only policy and commissioning can solve the problem: no individual clinician can address population-level harm concentration through their own practice alone.

This distinction matters because it determines what any competency framework must do. It must be honest about what individual competency cannot achieve, and clear and specific about what it must.

Key Finding

The evidence shows that AI deployment without appropriate clinical competency in place does not supplement clinical reasoning. Over time, it erodes it. But the erosion does not stop at the individual level. Individual capability failures feed systemic oversight failures. Systemic failures amplify harm in the communities already most exposed. The three levels compound.

How the risks compound: seven pathways

The most important finding from this analysis is not what the individual risks are - it is how they interact. There are seven specific pathways through which the risks at each level amplify each other. Understanding these pathways is essential because a competency framework that addresses any one risk in isolation, without addressing the mechanisms that connect it to others, will leave those connections unmitigated.

Take the most clinically significant: deskilling and automation bias form a self-reinforcing loop. As AI performs more clinical tasks reliably, clinicians have fewer opportunities to exercise the skills those tasks require. As skills atrophy, the ability to critically evaluate AI outputs diminishes. As critical evaluation diminishes, automation bias increases. As automation bias increases, skills atrophy further. This loop operates invisibly - neither the skill loss nor the bias increase is detectable through standard clinical performance management. It only becomes visible through the patient outcomes it eventually produces.

Cognitive load intensifies every pathway involving automation bias. The conditions of clinical deployment - time pressure, alert volume, fatigue - are precisely the conditions under which automation bias is highest. AI tools assessed as clinically safe under controlled evaluation may carry significantly elevated risk in the high-load environments where they are most commonly used. Figure 2 maps all seven pathways. Click any pathway to see the specific factors that combine and the mechanism through which they interact.

Figure 2 · Seven Compounding Risk Pathways

Click each pathway to expand

Individual factors Systemic / structural factors

Deskilling ↔ Automation Bias — self-reinforcing loop

Deskilling↔Automation Bias

Level 1

▼

Deskilling reduces the clinical reasoning baseline from which AI outputs can be evaluated — which increases automation bias. Automation bias then reduces the independent judgement that would maintain skills, accelerating deskilling further. Each mechanism amplifies the other.

A landmark randomised study demonstrated that physician plus AI arms performed worse than AI operating alone, because AI assistance without adequate critical appraisal actively degrades performance relative to either agent independently.

Cognitive Load Amplified Automation Bias

Cognitive Overload+Automation Bias→Evaluative Suppression

Level 1

▼

High cognitive load — time pressure, alert volume, fatigue — reduces the deliberate analytical processing that mitigates automation bias. Alert override rates of 90–96% in deployed environments document systematic evaluative suppression at scale.

AI tools assessed as safe under controlled conditions carry significantly higher risk in high-load environments such as emergency medicine and overnight primary care.

Burnout and the Efficiency Rebound Effect

Burnout+Efficiency Rebound→Elevated Cog. Load

Level 1

▼

Burnout impairs the cognitive and emotional resources required for safe AI-augmented practice. The Efficiency Rebound compounds this: AI lowers the marginal cost of performing clinical tasks, generating additional demand that absorbs efficiency gains — increasing rather than reducing workload.

AI deployment should not be assumed to reduce the cognitive load conditions that amplify automation bias. The same tools that generate efficiency gains may simultaneously elevate imaging volume, multiply alert burden, and accelerate burnout.

Algorithmic Bias × Governance Failure

Algorithmic Bias×Governance Failure→Undetected Discrimination

Levels 2–3

▼

Algorithmic bias produces systematically less accurate outputs for demographic groups underrepresented in training data. Governance failure — absent post-deployment monitoring, absent equity audit, absent clinician bias literacy — means the bias perpetuates rather than being detected and corrected.

The interaction is multiplicative: each factor enables the other to persist. Bias without governance goes undetected. Governance without bias literacy fails to act on what is detected.

Compounded Discriminatory Bias

Algorithmic Bias+CL Automation Bias+Governance Weakness→Harm Concentration

All levels

▼

Three mechanisms converge into a single pattern of harm concentration. Algorithmic bias produces less accurate outputs for underrepresented groups. The tendency to defer to AI without critical appraisal is highest among non-specialists most commonly deployed in under-resourced settings. Governance weakness is typically most acute in precisely these settings.

The result: the populations most at risk from biased AI outputs are served by the clinicians least equipped to detect that bias, in the settings least able to monitor it.

Moral Disengagement → Human Oversight Erosion

Moral Disengagement(at scale)→Oversight Erosion

Levels 1–2

▼

Structural conditions — legally ambiguous accountability frameworks, asymmetric professional risk of overriding AI, high output volume, and institutional framing — drive individual clinicians toward progressive withdrawal from personal accountability for AI-assisted decisions (moral disengagement).

When this occurs at scale, the cumulative effect is human oversight erosion: the human clinical oversight that governance frameworks depend on as a fundamental safeguard is weakened or absent. AI governance is premised on human oversight functioning — this is the mechanism by which it can fail from within.

Consent Gaps and Equity Disadvantage

Low Literacy + Digital Exclusion+Absent Disclosure→Tiered Consent

Levels 1 + 3

▼

Patients frequently receive AI-influenced care without disclosure or meaningful consent. Low health literacy and digital exclusion create a tiered consent hierarchy: patients least able to engage with AI disclosure are those who most need its protection.

The same communities most at risk from biased AI outputs (I5) are least equipped to exercise informed preferences in response to them. This pathway connects the individual clinical encounter to the population-level equity risk at Level 3.

Why the pathways must be understood together

These seven pathways are not independent. The deskilling–automation bias loop (I1) intensifies under cognitive load (I2), which is worsened by efficiency rebound (I3). Moral disengagement at the individual level (I6) degrades the governance oversight that should catch systemic failure (I4). Governance failure at Level 2 amplifies the concentration of harm at Level 3, where it falls on the communities identified in I5 and I7. A framework that addresses any one of these risks in isolation leaves the others unmitigated.

Bloom (2026) · Equiti Health · Systematic narrative synthesis of 445 studies, 2021–2026

Figure 2 is the analytical foundation for the competency framework that follows. Each pathway represents a specific clinical or structural failure; each domain in the framework in Section 4 is a direct response to one or more of those failures. The seven pathways are not a list of problems - they are a map of a single interconnected risk system. The framework must be equally integrated to address them.

What Existing Frameworks Provide - and Where They Fall Short

Twenty-three existing AI competency and capability frameworks were reviewed as part of this research, spanning undergraduate medical education, postgraduate clinical development, workforce capability, and international consensus. The full analysis is in the appendix to the primary research document.

The strongest framework reviewed was produced through a two-round expert Delphi study involving over 200 participants from 79 countries. It defines digital health competencies across four domains and includes 19 specific learning outcomes for AI in healthcare - the most detailed AI literacy specification of any reviewed framework. It has been adapted for clinical training in several countries and represents the most rigorous platform that medical education has so far produced for clinical AI competency development.

Despite its strength, what the full analysis shows is that this platform - across all 23 reviewed frameworks - does not reach the operational safety risks that AI deployment creates at the point of care. In every reviewed framework, the competencies most critical for front-line patient safety are limited to awareness statements. Clinicians are expected to understand what AI is. They are not equipped to practise safely with it.

Key Finding

Across all 23 reviewed frameworks, the competencies that matter most for front-line patient safety are assigned to commissioners, technical specialists, and governance leads - not to the clinicians using AI tools and bearing personal accountability for the decisions those tools inform. The workforce is increasingly AI-informed. It is not AI-competent in the safety-critical sense that practice now requires.

The analysis identifies four specific structural gaps - four categories of clinical risk with no competency response in any reviewed framework. They are presented below. Click each one to see what is missing, why it matters clinically, and how this framework addresses it.

Four Structural Gaps in Existing Frameworks

Click each gap to expand

Deskilling, mis-skilling, and never-skilling are not addressed as assessable clinical safety competencies

Level 1 — individual risk

▼

What the gap is

When clinicians use AI tools to perform clinical tasks, the skill required to perform those tasks independently can atrophy through disuse. When AI produces plausible but incorrect outputs that are accepted uncritically, incorrect reasoning patterns can become embedded. For trainees entering AI-intensive environments, foundational skills may never develop at all. These three pathways are distinct, cumulative, and clinically serious. No reviewed framework makes preventing them an operational, assessable competency.

Why it matters

The first multicentre real-world evidence documented unassisted diagnostic accuracy falling measurably over just three months of AI-assisted practice — during a period when no one was monitoring the skill. If clinicians cannot perform an assessment without AI assistance, they cannot evaluate AI outputs against an independent baseline, detect when AI is wrong, or practise safely when AI fails.

This framework defines all three pathways as distinct risks, specifies Minimum Clinical Competency Thresholds as a governance requirement, and makes their maintenance an assessable competency at Practitioner and Senior career stages.

Algorithmic bias detection is not a point-of-care clinical skill for front-line users

Level 3 — population risk

▼

What the gap is

In every reviewed framework, algorithmic bias identification and mitigation is assigned to procurement, deployment, and technical development roles. Clinical users are expected to be aware that bias can exist. None makes detecting demographic performance anomalies in specific deployed tools, for specific patient groups, in specific clinical contexts, a named clinical competency for front-line users.

Why it matters

A landmark study demonstrated that a widely deployed commercial health algorithm systematically underestimated the clinical needs of Black patients at equivalent risk scores. Knowing that bias can exist in principle does not equip a clinician to detect it in the output they are reviewing for the patient in front of them. That requires a specific, learnable clinical skill.

This framework makes algorithmic bias detection an explicit Practitioner-level competency: the ability to identify equity red flags at the point of care, for specific patients, and to escalate through formal governance channels.

Governance literacy is not established as a patient safety requirement for front-line clinicians

Level 2 — systemic risk

▼

What the gap is

Across all reviewed frameworks, governance is consistently treated as the responsibility of commissioning and technical roles. No framework establishes that every front-line clinician using AI tools has a professional obligation to understand whether those tools are safety-assured, and a right and responsibility to act when they are not.

Why it matters

A significant proportion of clinical organisations deploy AI tools without documented clinical safety assurance. Clinicians in those environments carry full personal accountability for AI-assisted decisions without the institutional infrastructure that should underpin them. This is a live patient safety risk in clinical environments today.

This framework makes governance literacy, gap recognition, and the use of escalation pathways an explicit clinical competency — not an institutional procedure.

AI disclosure, consent facilitation, and AI-augmented shared decision-making are absent from all frameworks

Levels 1 + 3

▼

What the gap is

Patients have a right to know when AI has contributed to clinical decisions about their care. Supporting that conversation in accessible language — in a way that preserves the patient's ability to engage meaningfully with the information and their own preferences — requires specific clinical skill. No reviewed framework addresses AI disclosure, consent facilitation, or AI-augmented shared decision-making as explicit clinical competencies.

Why it matters

AI-augmented shared decision-making operates differently from standard shared decision-making. The clinician must hold the AI output, their own independent clinical judgement, and the patient's values in productive tension — and ensure the patient can meaningfully engage with all three. The ability to participate meaningfully in AI-augmented consultations is not equally distributed, which connects this gap directly to the population equity risks at Level 3.

This framework addresses AI disclosure and AI-augmented shared decision-making as specific clinical competencies across all three career stages, and connects them explicitly to the population equity dimension they carry.

These four gaps are not shortcomings of individual frameworks - they are the collective limits of what existing frameworks were designed to address. The frameworks reviewed were built for different purposes: AI literacy for graduates, digital capability for the workforce, governance standards for deployers. What they do not provide is what this framework is designed to supply: a structured account of what practising clinicians must be able to do to practise safely with AI tools in clinical settings.

The Framework: Five Domains, Three Career Stages

The competency framework proposed in this research is a direct response to the risk architecture in Section 2 and the gaps in Section 3. It does not add AI awareness to existing frameworks. It defines, for the first time, what clinicians at different stages of their career should be able to do when working with AI tools in clinical practice - not what they should know about AI in the abstract, but how they should act when AI is present in the clinical encounter.

The framework has five domains. Each domain addresses a specific failure identified in Sections 2 and 3.

Domain 1 - Technical AI Literacy is the enabling prerequisite: without it, none of the others can function.

Domain 2 - Clinical Judgement Preservation is the primary response to Level 1 risk, exercised in every AI-supported clinical encounter. It directly addresses the deskilling loop and automation bias identified in pathways I1 and I2.

Domain 3 - Algorithmic Bias and Health Equity addresses Level 3 population-level risk, exercised at the level of the individual consultation - the only point where population-level harm can be caught or mitigated in practice.

Domain 4 - AI Safety and Governance is the primary response to Level 2 systemic risk, providing front-line clinicians with the governance literacy that existing frameworks assign exclusively to commissioners and deployment leads.

Domain 5 - Professional Agency and Person-Centred Care operates across both Level 1 and Level 3, addressing the moral disengagement pathway (I6) and the consent and equity pathway (I7).

The framework is differentiated across three career stages. The differentiation is not simply about depth of knowledge - it is about the nature of the obligation. Foundation clinicians (students and trainees) need to understand these risks and begin developing the habits that protect against them. Practitioners (qualified clinicians) need to apply those habits consistently and independently in their own clinical practice. Senior clinicians and leaders need to create the organisational conditions, governance structures, and training environments that make safe AI-augmented practice possible for the clinicians they lead.

Figure 3 shows what each career stage should be able to do within each domain. The bolded statements in each cell are the most clinically consequential competency at that stage - the one whose absence most directly enables the risks described in Sections 2 and 3.

Figure 3 · Five-Domain Clinical AI Competency Framework

Click any domain to expand competency requirements across career stages

① Foundation

Students & trainees

② Practitioner

Qualified clinicians

③ Senior / Leadership

Leaders & educators

Technical AI Literacy and Evaluation

All levels — enabling domainUnderstanding AI performance, failure modes, and validation evidence

▼

① Foundation

Can name AI types and limitations including supervised learning, large language models, and clinical decision support.

Can interpret basic performance metrics including sensitivity, specificity, and AUC.

Understands that confidence scores reflect training distributions, not clinical certainty.

Knows that AI tools are validated on specific populations and can recognise that validation may not reflect local demographics.

② Practitioner

Can evaluate validation evidence against local patient demographics and clinical context.

Able to identify LLM-specific failure modes including hallucination, prompt sensitivity, and uncertainty misrepresentation.

Can critically appraise commercial AI claims against peer-reviewed evidence and recognise that a tool validated in one setting may degrade in another.

③ Senior / Leadership

Can lead AI procurement evaluation using structured evidence appraisal frameworks.

Able to set institutional evidence standards for AI tool adoption.

Understands the regulatory classification of Software as a Medical Device and its implications.

Can assess whether a tool's validation population is sufficiently representative for the local setting before endorsing clinical adoption.

Clinical Judgement Preservation and Critical AI Appraisal

Level 1 — primary responseDeskilling prevention, verify-first discipline, cognitive load management

▼

① Foundation

Understands deskilling, mis-skilling, and never-skilling as distinct and cumulative risks to their own developing competency.

Aware of the concept of Minimum Clinical Competency Thresholds and why they matter.

Can apply verify-first thinking — able to form an independent clinical view before reviewing AI output, in supervised settings.

② Practitioner

Can apply verify-first discipline consistently and documents independent clinical reasoning before reviewing AI recommendation.

Actively monitors own skill maintenance across domains where AI assistance is available.

Can recognise when time pressure, fatigue, or alert volume are increasing uncritical AI deference — and is able to act on that recognition in the moment.

③ Senior / Leadership

Able to ensure Minimum Clinical Competency Thresholds are defined and maintained for their teams and specialty.

Can create and protect unassisted practice environments for structured skill maintenance.

Can lead audit of team AI acceptance rates as a governance signal for automation bias. Able to advocate for deskilling prevention in training curricula and revalidation standards.

Algorithmic Bias Awareness and Health Equity

Level 3 — primary | Level 1 individualEquity red flags at point of care, compounded discriminatory bias

▼

① Foundation

Can name types of algorithmic bias including training data bias, label bias, and demographic underrepresentation.

Understands that overall accuracy statistics do not guarantee accuracy across demographic subgroups.

Knows to ask about the training population of AI tools they are asked to use — and able to articulate why this matters for patient safety.

② Practitioner

Can identify equity red flags in AI outputs at the point of care: unexpected or unexplained findings in patients from underrepresented groups.

Knows the training populations of deployed AI tools in their clinical environment.

Able to report equity concerns through formal clinical governance channels rather than informally adapting practice to compensate.

③ Senior / Leadership

Can initiate local equity audits of deployed AI tools: able to review output patterns across patient demographic groups and escalate anomalies.

Reviews validation evidence for demographic stratification as a condition of endorsing AI tools.

Able to advocate for equity-stratified procurement requirements at institutional, system, and policy level.

AI Safety, Governance and Professional Accountability

Level 2 — primary responseGovernance literacy, escalation pathways, professional accountability

▼

① Foundation

Understands that clinical AI tools should be regulated and safety-assured before clinical deployment.

Aware of the existence of applicable clinical safety standards and what they require.

Knows they have a right and a professional responsibility to ask about the governance status of AI tools they are asked to use. Understands that personal professional accountability is not transferred by tool certification or deployment approval.

② Practitioner

Can understand and apply the applicable governance framework for AI tools in clinical use, including relevant regulatory standards.

Able to identify absent safety assurance and treats that absence as a clinical risk.

Knows and can use the clinical safety escalation pathway — and understands the professional obligation to do so.

③ Senior / Leadership

Can lead institutional governance processes for AI tools: able to ensure safety documentation is current, complete, and regularly reviewed.

Able to manage post-deployment surveillance programmes including equity-stratified outcome monitoring.

Can advocate for governance standards that protect front-line clinicians operating in settings where institutional infrastructure is absent or inadequate.

Professional Agency and AI-Augmented Person-Centred Care

Level 1 + Level 3Moral accountability, AI disclosure, shared decision-making

▼

① Foundation

Able to maintain personal accountability for all AI-assisted clinical decisions as a non-negotiable professional obligation.

Can explain AI involvement in care to patients in accessible, non-technical language.

Aware of moral disengagement as a professional risk and able to recognise the structural conditions — not personal failing — that make it more likely.

② Practitioner

Can recognise personal moral disengagement triggers and actively counter them.

Able to disclose AI involvement to patients where it is material to clinical decisions, in language accessible to that patient.

Can integrate AI information into shared decision-making while keeping patient preferences, values, and understanding primary.

Applies heightened attention and effort for patients with lower health literacy or digital access — and can adapt the consultation accordingly.

③ Senior / Leadership

Can lead organisational culture and policy supporting professional moral accountability in AI-augmented practice.

Able to develop and implement AI disclosure standards for their institution or specialty.

Can monitor AI acceptance rates as a governance signal for moral disengagement at team level.

Able to advocate for equitable access to meaningful AI-augmented consultations for patients facing health literacy and digital exclusion barriers.

Bloom (2026) · Equiti Health · Systematic narrative synthesis of 445 studies, 2021–2026

Design principles

The framework is evidence-based: every domain and competency is traceable to the included evidence base of 445 studies. It is proportionate: the depth of competency specification reflects the depth of evidence available in each domain. It is career-stage differentiated: obligations differ systematically between foundation, practitioner, and senior stages, reflecting progression of responsibility rather than educational level alone. And it is actionable: each competency specifies what a clinician should be able to do - not simply what they should be aware of.

Key Finding

No domain operates in isolation. The risks are interconnected, and the competency architecture must be equally integrated. A clinician with strong technical literacy but no ability to detect algorithmic bias in the patient in front of them has an incomplete competency profile. The framework is designed to be used as a whole.

The current version represents the evidence-informed architecture. Full competency specifications - including assessment criteria, progression indicators, and institutional accountability requirements for each career stage - will be developed in consultation with practising clinicians across professional groups and published as the companion implementation document.

What This Means for Practice, Education, and Policy

The following implications are presented as propositions for engagement. Implementation guidance will be developed in consultation with clinical stakeholders and published as a companion document. The framework is designed for international application; where specific regulatory or governance frameworks are referenced, regional equivalents apply.

For Workforce and Education Bodies

Existing clinical AI capability frameworks provide the right foundation for the practising workforce. For undergraduate medical education, recent international consensus frameworks on digital health competencies - including work from the Medical Schools Council and Health Data Research UK - provide the complementary platform from which graduating doctors enter practice. The competencies in this framework are designed to build on both.

The priority extensions required across all workforce frameworks are: deskilling prevention specified as an assessable clinical competency with defined minimum competency thresholds; algorithmic bias detection operationalised as a point-of-care, user-level clinical skill; front-line governance literacy established as a patient safety requirement across clinical professional groups; and Domain 5 - Professional Agency and AI-Augmented Person-Centred Care - as a structured competency domain. Career-stage differentiation should reflect clinical responsibility and autonomy, with particular attention to the never-skilling risk for practitioners training in AI-intensive environments.

For Regulators and Professional Bodies

The currently ambiguous professional accountability landscape for AI-assisted clinical decisions is creating structural drivers of moral disengagement. Regulatory clarity on what accountability for AI-assisted decisions requires - including the conditions under which following an AI recommendation does and does not satisfy professional accountability standards - is a patient safety matter, not only a legal one.

The asymmetric professional risk currently associated with overriding AI recommendations, relative to following them, requires clarification. Where regulatory guidance cannot address this asymmetry, professional bodies can establish the practice standards that fill the gap. Delay in doing so is not a neutral position: it leaves the structural conditions for moral disengagement in place.

For Clinical Educators and Training Leads

Minimum Clinical Competency Thresholds - defined floors of unassisted clinical competency below which AI exposure should not commence, and which must be actively maintained throughout AI-augmented practice - should be considered for inclusion in training curricula and placement governance frameworks. Protected unassisted practice environments are not optional adjuncts to AI training. They are its prerequisite.

The never-skilling risk is most acute in high-AI training environments. Curriculum governance in these settings requires active, not passive, skill-floor protection. Clinical educators bear a specific responsibility for this: they are the architects of the conditions in which the next generation of clinicians will develop their clinical baseline, in environments where AI is present from the beginning.

For Clinical Governance and Patient Safety Teams

The evidence shows that in many health systems, a significant proportion of clinical organisations are deploying AI tools without documented clinical safety assurance - without formal validation against local patient populations, without defined post-deployment monitoring, and without a responsible individual accountable for clinical performance. Clinicians in those organisations are carrying professional accountability for AI-assisted decisions without the infrastructure that should underpin them.

Governance remediation is urgent wherever it is required. Where it is underway, clinicians require explicit guidance on what their professional obligations are in the interim - they cannot wait for governance to be complete before they encounter patients. Where remediation has not begun, the conditions for post-deployment surveillance, equity monitoring, and incident reporting for AI-related events do not exist in a sufficiently robust form to detect the harms that may already be occurring.

Next Steps

This white paper establishes the evidence base and framework architecture. The next phase of this work will develop the full competency specifications, assessment criteria, and implementation tools that translate the framework into practical guidance for health organisations, professional bodies, clinical educators, and workforce development leads.

The companion implementation document will provide career-stage-specific competency items in assessable form, assessment frameworks and indicators for each domain, guidance for organisations seeking to adopt the framework, and tools for clinical educators integrating AI safety competency into training programmes and curricula.

This work will be developed with input from practising clinicians across professional groups and career stages, and will be published in consultation with relevant health system, regulatory, and educational stakeholders. The research framework on which this white paper is based is available in full for those wishing to engage with the underpinning evidence.

Working on AI transformation in your health organisation?

Equiti Health supports health organisations, workforce bodies, and clinical educators to implement evidence-based AI competency frameworks, develop clinical AI governance infrastructure, and design equitable AI deployment strategies. If you would like to discuss how this framework applies to your context, or to find out more about Equiti Health's AI transformation services, get in touch.

Contact Equiti Health →

equitihealth.co.uk · hello@equitihealth.co.uk

About This Research

This white paper reports the findings of a systematic narrative synthesis of the clinical AI risk and competency literature, conducted January–April 2026. The evidence base comprises 445 studies: 28 anchor studies identified through targeted multi-database search, and 417 records retrieved through systematic PubMed search across ten domain clusters. In addition, 23 current clinical AI competency frameworks were identified and reviewed.

The narrative synthesis methodology followed Popay et al. (2006) guidance. Full methodological detail is set out in the primary research document. AI assistance (Claude, Anthropic) was used throughout the research and writing process; all analytical judgements and editorial decisions were made by the author.

Shoshana Bloom, Founder and Principal Consultant, Equiti Health Ltd. Published by Equiti Health Ltd, April 2026. This white paper may be cited with appropriate attribution: Bloom S (2026). Clinical Competency in the Age of AI: Safe, equitable, and effective practice in AI-augmented healthcare. Equiti Health Ltd.

Key References

Abdulnour et al. (2025). Educational Strategies for Clinical Supervision of AI Use. NEJM, 393(8):786–797.
Budzyñ et al. (2025). Endoscopist Deskilling Risk After AI in Colonoscopy. Lancet Gastroenterol Hepatol, 10:896–903.
Car et al. (2025). The Digital Health Competencies in Medical Education Framework (DECODE). JAMA Netw Open, 8(1):e2453131.
Goddard et al. (2012). Automation Bias: A Systematic Review. JAMIA, 19:121–127.
Graafsma et al. (2024). AI to Optimise Medication Alerts. JAMIA, 31(6):1411–1422.
Jabbour et al. (2023). Measuring the Impact of AI in Diagnosis. JAMA, 330(23):2275–2284.
Medical Schools Council and HDR UK (2025). Data Science in the Medical Curriculum.
NHS HEE (2023). AI and Digital Healthcare Technologies Capability Framework.
Obermeyer et al. (2019). Dissecting Racial Bias in a Health Algorithm. Science, 366(6464):447–453.
Oskrochi (2025). Digital Health Technology Compliance in NHS Trusts. JMIR.
Reid and Mateen (2025). The Jevons Paradox in Global Health. Lancet Digital Health.
Popay et al. (2006). Guidance on Narrative Synthesis in Systematic Reviews. ESRC.