AI Adoption in Healthcare: Why Most Rollouts Fail — and How to Join the 5% That Win

Reading time: ~11 minutes. Last updated: June 2026.

Key takeaways

• 95% of enterprise generative-AI pilots deliver no measurable impact on the bottom line (MIT Project NANDA, The GenAI Divide: State of AI in Business 2025). McKinsey's State of AI 2025 confirms the picture: more than 80% of organizations see no material effect on EBIT.
• The reason rollouts fail is not the technology. MIT calls it a 'learning gap'; McKinsey puts it in one line: AI is 20% algorithms and 80% organizational rewiring.
• In healthcare the stakes go beyond ROI to patient safety and EU AI Act compliance, which treats most clinical AI systems as 'high-risk' and mandates human oversight.
• The key to the winning 5% is goal alignment across functions (R&D, IT, Regulatory, Medical Affairs, Quality) and psychological safety in teams — not another tool.
• One distinction reframes the whole topic: AI does not collaborate with people. Collaboration is a social process between humans. With an AI system you run an exchange and coordination. It sounds like a linguistic detail — it decides whether your transformation succeeds.

If you follow the headlines, you know the dissonance. On one side: AI designs drug candidates, reads X-rays, and writes clinical documentation. On the other: the hard data says almost no enterprise AI rollout produces a return.

That is not a paradox. It is a diagnosis. And the good news is that the problem blocking AI adoption in healthcare is solvable — because it is organizational, not technological. This article shows exactly where transformation breaks and what the organizations that succeed do differently.

What is AI adoption in healthcare, really?

AI adoption in healthcare is not buying a model. It is a durable change in processes, roles, and decision-making that lets AI tools genuinely move clinical, operational, and financial outcomes. That is the difference between 'we have a chatbot' and 'we cut the time from hypothesis to drug candidate.'

The clearest way to see it is at the level of value. There is individual AI — it saves a single person time (a faster email, a quicker article summary). And there is institutional AI — it redesigns the entire workflow so the organization scales revenue, quality, and speed. Most companies stop at the first level and then wonder why outcomes don't change. Because, as a core thesis of modern transformation puts it: productive people don't make a productive company — productive processes make a productive company.

Why do 95% of AI rollouts fail to move the needle?

Because companies 'swap the engine' but keep the old factory. They replace Excel with a chatbot for Excel, email with an email-drafting assistant — and expect a revolution. But a tool without a redesigned process is just a faster way of doing the same things.

The numbers are unambiguous and consistent across two independent studies:

• MIT Project NANDA (2025), The GenAI Divide: based on 300+ deployments, 52 interviews, and 153 executive surveys — only 5% of integrated AI pilots generate measurable value; the rest stall before they scale. The authors state plainly that the divide is driven not by model quality or regulation, but by approach.
• McKinsey, State of AI 2025 (November 2025): although roughly 80% of firms now use generative AI in at least one function, more than 80% see no material EBIT impact, and only about 39% attribute any earnings impact to AI (usually under 5%). The winning cohort is barely ~5–6% of organizations — and they consistently redesign their workflows.

This is not a data-science problem. It is an organizational one. And organizational problems are solved with people and processes — not another license.

AI doesn't collaborate. People do.

In marketing, 'human-AI collaboration' is a fashionable phrase. It sounds good, but it is imprecise — and that imprecision actively hurts rollouts. Collaboration is a social process. It happens between people, teams, and organizations. Its essence — as social psychologists Victor Wekselberg and Jacek Wasilewski show — is goal alignment across levels (individual, team, organization). It does not mean everyone holds one identical goal. It means individual, team, and organizational goals are consistent and non-conflicting.

So what do we actually do with AI? Two things that are easy to mistake for collaboration:

• Exchange (a transaction). You give the system context and a prompt; you get an output. Trust here rests on the expectation that the other party delivers what was agreed — not on shared goals. That is the definition of an exchange, not collaboration.
• Coordination. You decide who leads the task: whether a human leads the AI, a defined workflow leads, an agent plans the steps under supervision, or a swarm of specialized agents runs. This is a technical component of work — necessary, but not the essence of collaboration. A high standard of coordination does not imply a high standard of collaboration.

Why is this more than semantics? Because it shows up in results. A meta-analysis published in Nature Human Behaviour (Vaccaro et al., 2024) found that combinations of humans and AI generally did not beat the best of either alone — not the human alone, not the AI alone. You see augmentation of the human, but in most applications true 'synergy' is missing. In other words: dropping AI into a team does not create a magical, collaborating super-unit. What actually decides outcomes is the collaboration between the people who design, supervise, and interpret that system.

For healthcare, this isn't philosophy — it's a risk map. An algorithm may 'solve the problem correctly at the level of a cell but fail to solve it at the level of the whole biological system.' Clinical context, nuance, accountability for the patient, and trust are human competencies. AI amplifies them. It does not replace them.

Stop asking 'how should our people collaborate with AI.' Start asking 'how should our people collaborate with each other — in an environment where part of the work is exchange with AI and the coordination of agents?'

Five barriers that stall AI in healthcare

In a regulated sector, where an innovation must pass through R&D, IT, Regulatory, Quality, and Medical, these barriers are especially costly.

• Goal misalignment. Data Science is measured on model accuracy and speed, while Regulatory and Medical Affairs are measured on clinical risk and quality. The result: R&D deploys an algorithm that shortens molecule identification, and Regulatory rejects it for lack of transparency (the 'black box'). Without goal alignment across functions, the project dies at the proof-of-concept stage.
• Algorithm aversion vs. automation bias. Clinicians can be skeptical of AI recommendations — especially after one visible error (algorithm aversion). On the other side lurks automation bias: uncritical trust in the machine and the dismissal of one's own expertise. In medicine, both extremes can cost a patient's health.
• Lack of psychological safety. When people are afraid to admit they don't understand a tool, or that the system got something wrong, the organization loses its most valuable feedback and model errors stay hidden. Amy Edmondson of Harvard reminds us that interpersonal risk becomes business risk.
• Data — the foundation no one sees. No data, no AI — and data is today the most common and most underestimated cause of failure. Roughly 80% of data scientists' time goes not to models but to cleaning and integrating data, and about 60% of AI projects never reach production, mainly due to data quality, availability, and ownership. This too is a collaboration problem: between the people who enter data and the people who build models.
• AI doesn't reduce work — it intensifies it. An eight-month ethnographic study at a technology company (Harvard Business Review, Ranganathan and Ye, February 2026) found that workers voluntarily took on more tasks, worked faster, and worked longer, because AI made 'doing more' feel easy and rewarding. The short-term productivity surge turned into overload, cognitive fatigue, and burnout. In healthcare, where staff burnout is already a crisis, this barrier cannot be ignored. The remedy is not employee 'self-regulation' but a set of organizational norms for using AI — an 'AI practice.'

Four value levers for AI in healthcare

Every AI project should map to at least one of four levers. It's a simple filter that separates initiatives with real value from 'toys.'

• Revenue and innovation. Drug discovery and design, precision medicine, new service models. This is where AI-native CROs operate (Selvita with its CADD unit, Ardigen) alongside global pharma partnerships with AI-native firms.
• Cost. Automating knowledge work in the back office, handling tickets, extracting data from documents (OCR + LLM), classifying correspondence. Less routine, lower cost to serve.
• Risk and compliance. Compliance, quality, safety, monitoring of production processes, documenting decisions for regulatory audit. In the medical sector this lever often decides whether a model reaches the market at all.
• Speed. Shortening the decision cycle, time-to-market, and time-to-insight. GenAI for preparing documentation, synthesizing literature, and generating and evaluating hypotheses.

The EU AI Act and healthcare — what actually changed in 2026

If you're planning AI in medicine, regulation isn't an add-on — it's part of the design.

The EU AI Act (Regulation 2024/1689) has been in force since 1 August 2024 and applies in phases. Many medical AI systems qualify as 'high-risk,' which means hard requirements: risk management, data quality, documentation, transparency and — crucially — human oversight.

An important update worth knowing: under the Digital Omnibus on AI package, EU institutions reached a provisional political agreement in May 2026 to defer the deadlines for high-risk systems. For stand-alone systems under Annex III, obligations are set to apply from 2 December 2027 (instead of 2 August 2026), and for AI embedded in regulated products (including medical devices, Annex I) from 2 August 2028. Note: until the text is formally adopted and published in the Official Journal, the original deadlines technically still apply — so treat these dates as direction, not certainty.

What does this mean strategically? You got time — not an excuse. The human-oversight and decision-documentation requirements reward organizations that are already building collaboration processes among experts for reviewing AI outputs. That is exactly the same organizational capability that drives ROI. Compliance and effective adoption are, in practice, the same investment.

How to design AI adoption that delivers: 6 principles

These principles follow directly from what separates the winning 5% from everyone else.

• Narrow the scope. When in doubt, narrow. One concrete pain point, one process, a short cycle. Small scope wins, because it proves value quickly and can actually be redesigned.
• Redesign the process, not just the tool. The question is not 'which model do we use,' but 'what should the workflow look like after AI.' Skip this, and you're left with a faster email instead of scalable revenue.
• Start with data and data ownership. In 2026, an AI strategy without a data strategy doesn't exist. Decide who owns data quality and availability — at once a technical question and a question of collaboration across functions.
• Establish goal alignment across silos. Map the goals of R&D, IT, Regulatory, Medical, and Quality so they are consistent. That, not another status meeting, determines whether a project moves from pilot to production.
• Build the human into the loop (human-in-the-loop) and add observability. Autonomy without observability is a risk — in medicine, a double one. Mechanisms to raise concerns, to iteratively retrain models on expert feedback, and to transparently document decisions build trust and satisfy the oversight requirement.
• Create an 'AI practice' — norms, not self-regulation. Since AI intensifies work, set the rules: when to pause, how to sequence tasks, how to verify outputs. This protects the team from burnout and quality loss.

Each of these is fundamentally an organizational and human decision, not a technical one. That is why the leader's role shifts from 'maker' to 'manager-orchestrator': you no longer manage just people and processes, but people, agents, and automations — and the decision interval shrinks from hours to minutes.

Poland as a proving ground for AI in life sciences

This isn't a story of 'catching up.' Poland is the sixth-largest pharmaceutical market in Europe and the largest in Central and Eastern Europe, with strong institutional support and a maturing innovation ecosystem:

• The Medical Research Agency (ABM) runs a multi-year national plan to develop the biomedical sector, with a significant share of funds going to companies that use AI.
• The National Centre for Research and Development (NCBR) funds innovation and participates in European programs (including Horizon Europe).
• Local players are deploying AI in practice: Polpharma (digital twins in manufacturing), Adamed (machine learning in molecule design), Selvita and Ardigen (AI-driven drug discovery), Infermedica (medical triage used by, among others, PZU Zdrowie).

The gap? Polish organizations still lack the practices and tools to run complex projects at the intersection of biology, medicine, and data science — where generic tools (MS Teams, Jira) can't handle the specifics of a regulated industry. Whoever builds that organizational capability first will win in a market where technology isn't scarce — the ability to absorb it is.

FAQ: AI adoption in healthcare

Will AI replace doctors and scientists?

Not for the foreseeable future. The evidence (including the Nature Human Behaviour meta-analysis, 2024) shows AI works best as an amplifier of the human, not a replacement. Clinical context, accountability, and patient trust remain human domains.

Why didn't our AI pilot make it to production?

Most often for three reasons: no redesigned process, data problems (quality, availability, ownership), and goal misalignment across functions. These are organizational barriers, not technological ones — and they can be removed.

Is 'human-AI collaboration' a good term?

It's a convenient shorthand but imprecise. Collaboration is a social process between humans, grounded in goal alignment. With an AI system you run an exchange and coordination. That distinction helps you design roles correctly: AI is a tool, and it's the people around it who collaborate.

What does the EU AI Act mean for my medical project?

Most likely a 'high-risk' classification, with requirements for human oversight, risk management, and documentation. Deadlines for high-risk systems are being pushed (directionally to 2 December 2027 for Annex III and 2 August 2028 for medical devices), but preparation should be underway now.

Where exactly should we start?

With one narrow process tied to a clear value lever (revenue, cost, risk, or speed), with clean data, and with mapped, aligned goals across functions. Prove value in weeks, not quarters.

Will AI reduce my staff's workload?

Not necessarily — research shows that without deliberate norms, AI tends to intensify work. That's why you need an organizational 'AI practice': rules for verification, pauses, and task sequencing that guard against burnout.

Conclusion: it's time for architects, not tool fans

The era of 'lone geniuses' and isolated algorithms in healthcare is ending. Machines analyze billions of data points, but it is people who understand clinical complexity, create context, and build trust. The future of the sector lies not in blind faith in black boxes, but in the smart pairing of human expertise with compute — and in the collaboration between people that holds the whole thing together.

The question for leaders is no longer 'should we deploy AI.' It's: are you ready to become an architect of an organization where technology is a tool that amplifies teams — and where outcomes are decided by goal alignment and human trust?

Sources

• MIT Project NANDA (2025). The GenAI Divide: State of AI in Business 2025.
• McKinsey & Company (2025). The State of AI in 2025: Agents, innovation, and transformation.
• Ranganathan, A., Ye, X. M. (2026). AI Doesn't Reduce Work — It Intensifies It. Harvard Business Review, February 2026.
• Vaccaro, M., Almaatouq, A., Malone, T. (2024). When combinations of humans and AI are useful. Nature Human Behaviour.
• Wekselberg, V., Wasilewski, J. Cooperation, Collaboration, Coordination, Groupthink (on collaboration as a social process and the alignment of goals).
• Edmondson, A. C. — research on psychological safety (Harvard Business School).
• European Commission / Council of the EU / European Parliament (2024–2026). EU AI Act (Regulation 2024/1689) and the Digital Omnibus on AI — May 2026 provisional agreement to defer high-risk deadlines.
• PAIH / European Parliament Briefing (2025–2026) — the position of the Polish pharmaceutical market in Europe.