IPM Take
The healthcare AI conversation is changing.
For years, the dominant question was whether artificial intelligence could perform a task. Now healthcare organisations have to decide whether a system should be bought, how it connects to existing workflows, who remains accountable for its decisions and what happens when its performance changes after deployment.
Appinventiv’s practical guide captures that shift well. It recommends starting with a defined high-value use case, preparing the data, building compliance in from the beginning, integrating directly into existing workflows and continuously monitoring performance after launch.
The important part is not the vendor playbook itself. It is that regulators, health systems and independent surveys are increasingly pointing in the same direction.
AI adoption is becoming less of a model-selection problem and more of a healthcare-governance problem.
Executive Summary
Generative AI adoption in healthcare is no longer confined to experimentation. In McKinsey’s fourth-quarter 2025 survey of US healthcare leaders, 50% said their organisations had implemented generative AI, compared with 25% in late 2023. More than 80% reported that at least one generative-AI use case had reached end users.
Physician use is growing too. An American Medical Association survey published in March 2026 found that 81% of surveyed physicians were using AI professionally, more than double the rate recorded when the AMA first surveyed physicians in 2023. Common uses included summarising medical evidence, documentation and other workflow support.
Appinventiv argues that healthcare AI projects should begin with specific problems rather than broad transformation programmes. Its guide divides current use cases across imaging, generative AI, predictive analytics, automation, conversational systems, drug discovery and clinical decision support, while recommending phased implementation around data governance, compliance and EHR integration.
The guide also provides indicative development costs of $40,000 to $90,000 for a proof of concept, $120,000 to $300,000 for a production MVP and $350,000 or more for enterprise deployment. These numbers are Appinventiv’s own planning estimates, not independent market benchmarks, and costs will vary substantially with data preparation, integration, validation, cybersecurity and regulatory requirements.
The regulatory direction is increasingly clear. FDA guidance emphasises lifecycle risk management for AI-enabled medical devices and now provides a final framework for predetermined change control plans, allowing planned model modifications to be assessed within an approved change strategy.
In the UK, the National Commission into the Regulation of AI in Healthcare has recommended staged authorisation, continuous real-world monitoring and clearer public information about AI-enabled medical devices. Those remain recommendations pending the government’s response, not yet a new statutory regulatory regime. The NHS Alliance has similarly argued that regulation alone will not be enough without workforce skills, implementation guidance, investment and evidence about which technologies work in practice.
Why it matters
- HTA bodies: AI evaluation cannot stop at accuracy. Assessments increasingly need to examine comparative clinical utility, workflow impact, human oversight, resource use, equity, downstream testing and performance after deployment.
- Payers: Administrative AI may produce relatively fast efficiency gains, but clinical AI carries different risk. Payment models should distinguish systems that automate low-risk workflows from technologies that influence diagnosis, treatment or access to care, and demand evidence proportional to those risks.
- Industry / innovation partners: The market is moving beyond impressive demonstrations. Products that integrate into EHRs, generate auditable evidence, support regulatory monitoring and fit existing clinical workflows may have a stronger route to adoption than technically superior models that remain difficult to deploy.
Healthcare AI is entering a less glamorous phase.
The demonstrations have worked. The procurement decisions are being made. Now organisations have to make the technology function inside healthcare.
That distinction matters because a model can be technically impressive and operationally useless.
Appinventiv’s current implementation guide argues that the largest costs and difficulties frequently sit around the model rather than inside it: cleaning and governing healthcare data, connecting to EHR systems, meeting privacy and regulatory requirements, building audit trails and monitoring the system after deployment.
This is consistent with what healthcare leaders themselves are reporting. McKinsey’s latest US survey suggests adoption has reached a more mature phase in which implementation, integration and return on investment are becoming as important as traditional concerns about AI safety. Interest is also moving toward agentic systems that can coordinate multiple tasks rather than simply generating text.
That shift raises the stakes.
An AI tool that drafts a note creates one set of risks. An AI agent that reads patient information, updates systems, books follow-up and escalates clinical findings creates a very different one.
Compliance cannot be bolted on after the model works
The commercial temptation is to build first and regulate later.
Healthcare is making that strategy increasingly difficult.
In the United States, the FDA’s digital-health framework now includes final guidance on predetermined change control plans for AI-enabled device software. The principle is important because AI products may change after initial authorisation. Developers can describe planned modifications, how those changes will be developed and validated, and how their effects will be assessed.
FDA guidance also increasingly frames AI-enabled medical devices through a total-product-lifecycle approach, including the data used to develop a system, validation, transparency, monitoring and performance following modifications.
The UK is moving toward a similar lifecycle philosophy. Its National Commission has proposed staged authorisations for some healthcare AI and continued real-world monitoring rather than relying solely on a single point of approval. The Commission gathered input from more than 12,000 people, and the government has yet to formally decide which recommendations it will implement.
That distinction between recommendation and law matters. AI governance is moving quickly, but organisations should not confuse emerging frameworks with requirements that are already legally binding.
The cheapest AI may be the one that never touches a clinical decision
The Appinventiv guide also makes a commercially useful distinction between clinical and administrative applications.
Documentation, scheduling, coding and workflow automation can offer comparatively clear measures of value, such as clinician time saved or reduced administrative effort. Clinical decision support, diagnostic AI and patient-facing agents have potentially greater clinical impact but also create higher requirements for evidence and oversight.
That helps explain why ambient documentation has become such an attractive entry point for health systems. The operational problem is obvious, outcomes are measurable and the clinician can remain responsible for reviewing the final note.
The calculus changes when AI starts determining who gets investigated, prioritised or treated.
At that point, model accuracy becomes only one component of performance. Organisations need to ask whether performance differs across patient groups, whether clinicians understand when the tool should be ignored, whether errors can be detected, and whether the AI creates downstream costs elsewhere in the system.
WHO’s longstanding governance framework makes the same broader point: health AI should protect human autonomy, safety, transparency, accountability, inclusiveness and equity, with continuing assessment during actual use.
The next stage of healthcare AI will therefore not be won solely by whoever builds the smartest model.
It will be won by organisations that can prove where AI belongs, integrate it without damaging workflows, monitor it after deployment and stop it when the evidence says they should.
Healthcare has already learned how to buy AI. The harder task is learning how to govern it.
