Introducing AI into organisations requires more than tools and compliance: a vision that strengthens human judgement and builds trust. Three diagnoses, four design principles and concrete healthcare examples — for leadership teams who want AI to be not only adopted, but effective.
As of May 2026
Why a vision now?
In many organisations, AI adoption currently follows a familiar pattern: individual teams experiment with tools, parallel initiatives run without coordination, the supply of applications grows faster than the ability to use them meaningfully. There is also the feeling of falling behind compared to other companies or countries — triggering the reflex to move faster, without knowing exactly where.
The result is well known: AI-generated output that looks superficially competent but creates rework — described by BetterUp Labs and the Stanford Social Media Lab as Workslop. 40% of employees received such content in the past month; estimated productivity loss around USD 186 per person per month.[1] MIT data shows: approximately 95% of enterprise AI pilots deliver no measurable business value.[2]
Anyone who wants to lead with AI therefore needs two things: a clear view of what is happening — and a vision against which decisions can be measured. There is also an economic argument: as AI makes routine and standard services cheaper, the value of work shifts to where the human element is itself part of the service — judgement, attention and relationship. The behavioural economist Alex Imas (University of Chicago Booth) describes this as a transition to a relational economy.[14] This page provides a draft for both: a situational analysis and an orientation framework.
A compact test logic — not a compliance checklist, but four questions against which decisions can be measured:
Human judgement takes priority. Systems are designed to consider human preferences, correction and context, rather than optimising a rigid objective in isolation. Good AI supports decisions — it does not replace responsibility or professional judgement (Stuart Russell: provably beneficial AI).[12]
Explainability is more than a technical feature. The key question is institutional: who is accountable to whom? AI systems require clear responsibilities, documented decision paths and verifiable governance structures. Without an identifiable accountability chain, robust trust cannot emerge.[9][13]
The benchmark is not the volume of outputs generated, but the quality of the person–task relationship. Good AI reduces unnecessary burden, administrative friction and documentation workload — particularly in complex environments like healthcare. It creates room for what only humans can do — AI and people complement each other rather than compete. In economic terms: what cannot be automated — judgement, attention, relationship — gains in value (Imas). Systems that merely generate more reports or alerts fail this principle.[1][5][14]
Trustworthy AI requires distributed control rather than central dependencies. Oversight, training data, models and auditability should remain organisationally traceable and separate. Monopolising data, infrastructure and decisions creates dependency. Distributing control creates resilience, transparency and room to act.[3][4]
Three diagnoses — framed as questions
Before deciding which AI tools to introduce, it is worth answering three questions. They structure the discussion and protect against a purely competition-driven adoption — adopting tools simply because others already use them. Behind each question is the observation of leading thinkers; the conclusions, however, are operational and decision-relevant, not philosophical.
Yuval Noah Harari describes in Nexus (2024) a rupture in the history of media: earlier technologies distributed information — the printing press spread texts, radio transmitted voices, the internet made content accessible. But they did not decide what to say. Today's AI systems do exactly that: they generate content, make predictions, influence behaviour in real time. Harari calls this the transition from tool to agent.[3]
For organisations, this has a concrete implication: when an AI system pre-screens applications, outputs therapy suggestions, prepares credit decisions or drafts customer communications, it is co-deciding — even if a human ultimately "confirms". The question is not whether this is bad. The question is whether this co-decision was deliberately designed or crept in unnoticed.
That a downstream "confirmation" alone is not sufficient is empirically documented: in clinical decision-support systems, automation bias leads professionals to revise correct judgements of their own in favour of incorrect system suggestions — human oversight then becomes a formality rather than genuine control.[16]
Operational lever: Before every AI deployment, name clearly where the system only provides information — and where it effectively makes decisions. Both are legitimate, but must be decided consciously.Mustafa Suleyman, co-founder of DeepMind and CEO of Microsoft AI, observes in The Coming Wave (2023) a recurring pattern: powerful technologies diffuse in waves — and beyond a certain level of maturity are no longer containable by individual providers, states or markets. Entry barriers fall, the same tools become available everywhere. Suleyman calls this the containment problem: how does one keep a technology controllable that simultaneously brings enormous benefits and carries serious risks?[4]
For leadership teams in companies or public bodies, this creates a concrete tension. On the one hand: the competition is using the tools — those who don't keep up fall behind. On the other: if everyone uses the same pre-trained models, part of one's own differentiation disappears. Suleyman's answer is not abstention, but a narrow path: between reflexive adoption and reflexive rejection. Concretely: which tools are strategic (own data, own logic, own governance) and which are commodity (interchangeable, no competitive advantage)?
That technological progress does not automatically generate broadly shared benefit, but depends on how control is distributed, is also the core of Acemoglu & Johnson's Power and Progress: whoever determines the direction and use of a technology decides how its returns are distributed. For organisations this means not surrendering data sovereignty and governance to individual providers out of convenience.[18]
Operational lever: Separate strategic from commodity AI applications. For strategic applications, consciously design data sovereignty, model selection and governance — do not delegate them.Hartmut Rosa (Jena) describes modern organisations as acceleration societies: those who become faster must become faster still to hold their place.[5] Byung-Chul Han adds the burnout society: digital work environments generate an unbounded self-exploitation.[6] Both findings may seem abstract — but they have a very concrete consequence for AI deployments.
In accelerating systems, productivity gains are not converted into time but into compression. Those who write faster write more. Those who report faster report more often. Precisely this effect appears empirically in AI use: auto-summaries lower the threshold for meetings, AI emails create expectations of more responses, generated reports make reports the norm. The Workslop phenomenon is not just a quality problem — it is a symptom that AI without a vision lowers the threshold for work rather than reducing it.[1]
Organisational research has empirically described the same mechanism for digital tools: mobile email gave knowledge workers more individual flexibility, but at the same time raised the collective expectation of constant availability and reduced the ability to disconnect — more autonomy individually, more compression overall (the autonomy paradox).[17]
Operational lever: Before every AI deployment, explicitly name which activity should decrease. If the answer is missing, the AI will reliably generate more work elsewhere — even if it is measurably "efficient".European Approach
In the global debate, the European approach is often described as a brake. This reading is too narrow. Europe has built something in recent years that exists in no other major economic area in this form: a coherent value framework for AI that brings together regulation, ethical principles and practical guidelines. For organisations operating in the EU, this is less a burden than an orientation.
Those who align with this framework build AI applications that are auditable, explainable and accountable — and thereby accessible to a market in which trust is scarce and trustworthiness is becoming a resource. This is not a moral bonus, but a strategic asset.
Three pillars support this framework:
Risk-based classification (Regulation 2024/1689), complemented by GDPR, MDR/IVDR and the European Health Data Space (Regulation (EU) 2025/327) in the healthcare context. For high-risk systems, binding requirements apply: risk management, data quality, transparency, human oversight.[7][8]
The Ethics Guidelines for Trustworthy AI (EU High-Level Expert Group, 2019) consolidate seven practical requirements — from human oversight through robustness to accountability — and are today the reference point for the AI Act.[9]
Julian Nida-Rümelin and Nathalie Weidenfeld articulate an explicitly European position: machines are tools — they can fulfil tasks but cannot bear authorship, dignity or responsibility. Anthropomorphising machines obscures the assignment of accountability that is constitutive of the rule of law and democracy.[10][11]
Healthcare as litmus test
The vision remains abstract until tested against concrete applications. Healthcare is the sharpest test case: here the difference between human-centred and output-centred AI deployment is particularly clear, because the consequences directly affect people — patients, nurses, physicians.
The following overview shows seven typical application areas. In each row, the same technology and use case is meant — the difference lies only in what it is deployed for and which measure of success applies.
Both columns show the same technology. What is desirable is their effective interplay: people with their strengths — judgement, experience, attentiveness — and AI with its new capabilities to handle routine and complexity. The human-centred deployment (left column) deliberately designs this interplay: AI takes on the recurring work, while people retain judgement and responsibility and gain room for what only they can do. This applies in healthcare — to physicians, nurses and patients — just as it does to employees in companies. The effect reinforces itself economically: the more AI takes over what is standardisable, the more valuable the human capabilities that cannot be automated become.[14][15] The output-centred deployment (right column) misses this interplay — it generates more activity without making the work better.
| Application area | Human-centred — relieves, strengthens judgement | Output-centred — compresses, replaces judgement |
|---|---|---|
| Nursing documentation | Speech recognition documents visits in the background → more time with patients; documentation becomes a by-product of care | Auto-reports create expectation of more reports → documentation burden grows, nursing time shrinks |
| Diagnosis support | Differential diagnosis suggestions sharpen clinical hypotheses, expand the search space, surface blind spots | "Fastest result" becomes a KPI → hypothesis space narrows, confirmation bias is amplified |
| Patient communication | Translation and explanation in patient language, comprehension of findings improved, questions deliberately encouraged | Chatbots lower the consultation threshold → enquiry volume rises, structural overload grows |
| Radiology / imaging | Second opinion and flagging of uncertain findings for specialist review — AI as safety net | Triage without final physician responsibility → finding treated as "confirmed by AI", accountability diffuses |
| Clinical trials | Patient matching and automated literature screening save research time; expert assessment remains with the study team | Endpoint selection and study design by model preference → bias enters the study design |
| Administration / coding | DRG coding is suggested, final review by human — reduction of repetitive tasks | Automated approvals / rejections without individual review — GDPR Art. 22 as unclear boundary |
| Therapy recommendation | Guideline-compliant suggestion, physician adapts to individual patient, recommendation documented | "Recommendation" becomes de facto binding → liability shifts unclearly, clinical autonomy declines |
The behavioural economist Alex Imas (University of Chicago Booth) provides an economic rationale for the healthcare vision. His thesis: when AI makes standard and routine services almost arbitrarily reproducible, the human element becomes the scarce good. The value of work shifts into a "relational" economy — to where the person is itself part of the service. He explicitly names care, medicine, therapy and support as core sectors.[14]
Concretely, three human capabilities gain importance — not despite but because AI takes over routine:
Weighing decisions under uncertainty, taking responsibility for individual cases and navigating institutional rules — what holds complex care together.
Presence, listening and recognising what is not in the record. In care this is not a "soft factor" but part of the service itself.
AI systems can be reliable, but not trustworthy: trust rests on goodwill and motives that a machine lacks — the trustful relationship with patients remains bound to people.[19]
Strategic framing: the more routine services become automatable, the more this human core becomes the carrier of value (cf. Imas[14]).
Strategic conclusion: a human-centred AI deployment that frees up these capabilities rather than displacing them — letting people and AI interact effectively — is the more robust path not only ethically but also economically. This applies to patient-centred care as well as to knowledge work in companies.[14][15]
Three questions before deployment
Those who can answer these three questions substantively have already considered essential requirements of the EU AI Act in the design — and at the same time have an internal argument for or against the respective project. The questions work equally for clinics, GP practices, pharma, MedTech and standard enterprise deployments.
If the second half cannot be answered clearly, the vision is missing. AI deployments without explicit reduction targets reliably generate additional work elsewhere.
Once an AI recommendation becomes de facto binding without this having been decided, the accountability chain is broken. This is not only a liability problem, but a governance problem.
Vendor lock-in rarely arises through a conscious decision — it arises through the convenience argument of individual procurement decisions. Those who decouple data, model and auditability retain room to manoeuvre.
Where does your organisation stand? In 10–15 minutes to a well-founded assessment of your Digital Maturity & AI Readiness — based on validated frameworks.
References
Unless otherwise stated, accessed: April 2026.