Is AI growing up?

Artificial intelligence has been characterized by cycles of hype and disillusionment, sparked by impressive breakthrough news in deep learning and, more recently, large language models (LLMs). These technologies have fueled expectations of autonomous driving, all-knowing personal assistants, and automating business decision-making. However, the real-world impact of AI has so far been limited to narrow use cases as diverse as image analysis, speech recognition, or prediction of protein folding.

Now, a new wave of AI applications is emerging, so-called “agentic AI.” BigTech and professional services firms expect significant advancements and new opportunities. But what exactly is meant by agentic AI, and how can healthcare and medtech organizations approach this new frontier? This primer breaks the buzz down to key concepts and practical relevance.

The AI evolution

Over the past decade, AI use cases have steadily expanded, particularly in the areas of data analysis, pattern recognition, and predictive modeling. The development of generative AI (genAI) solutions such as LLMs, widely recognized by the public release of ChatGPT, added semantic processing and language proficiency to this toolkit. These advancements have extended practical applications into areas such as coding assistance, chatbot functionality, semantic search, and standardization of unstructured data.

These applications work so well because of the architecture underlying LLMs (aptly named “transformers”), which excels in tasks like translation and sequence completion. Similarly in healthcare, the frontrunner use cases of LLMs are AI scribes, which can automate parts of the clinical documentation process. While these applications are limited to isolated tasks and clear relationships of inputs and outputs, the focus is now shifting to systems that can perform complex, multi-step processes autonomously.

Agentic AI represents the next stage in this evolution: systems capable of planning, decision-making, and executing tasks autonomously. Research firm Gartner identifies agentic AI as a top trend for 2025, and the well-known tech incubator and VC firm YCombinator predicts that agentic AI’s business impact will significantly surpass the software-as-a-service (SaaS) industry.

AI agent practical definition and examples

So what actually is an AI agent? Although definitions still vary, a commonly cited and practically useful framework from OpenAI uses the following core components:

• autonomous planning
• action-taking,
• and memory.

Others add the ability to perceive an environment, reason about it, and act accordingly, but this anthropomorphic view does not really further our understanding and can rather be misleading.

While the foundational research on agentic AI has been ongoing for decades with impressive successes in games such as Go, one shouldn’t think of AI agents as a super-intelligent black box that now takes on human work. It also shouldn’t simply be regarded as “genAI 2.0”. Instead, they represent a new level of integration between AI capabilities and real-world operational workflows.

The current wave of practical innovation stems from engineered software that utilizes LLMs as decision-making components with added layers of sensory inputs, memory and capability to interact with the world. As Oriol Vinyals from Google DeepMind put it, you can loosely compare the LLM with an electronic brain and the agentic software around it with a digital body.

But driving this “brain” is a carefully crafted software layer that processes inputs and outputs (for example actually triggers the chosen actions), stores and retrieves information from memory and invokes the LLM components. Such agents are not constrained to a single LLM model either, but can utilize and orchestrate several models based on different architectures or simply invoked with different prompts.

In summary, current AI agents are software systems designed by humans for a given application that utilize LLMs to perform tasks with some level of autonomy. One such example is the chat interface ChatGPT itself, which utilizes models (e.g. GPT-4o, but also DALL-E for image generation) under the hood, can invoke tools such as web search and can output results in various tailored formats. Another example is a synthetic customer support call center agent that can listen to and react in human speech, retrieve information from company storage and trigger workflow actions according to the specific customer context.

What does it mean for your healthcare business

So while the first LLM use cases primarily focused on individual steps like reviewing or summarizing text, generating ideas, or providing second opinions, the next stage involves using agentic systems to automate business processes.

By combining LLM capabilities with defined workflow rules, healthcare organizations can move beyond simple chatbots that respond to predefined queries and toward systems that automate end-to-end processes, such as coordinating patient pathway, managing supply chains or processing claims.

Agentic AI approaches are expected to drive the development of new tools and services by IT vendors and startups. It is safe to expect an influx of new offerings and diverse internal opportunities for improving operational efficiency exist in every organization.

While it’s easier than ever to run pilots, to effectively utilize these opportunities in scalable form, healthcare organizations should follow a structured approach:

1. engage stakeholders early and ensure buy-in
2. focus on high-value use cases and assess readiness and hurdles to integration, adoption and value creation
3. understand AI’s limitations, data quality and performance requirements for the task and define scope accordingly
4. accompany rollout and use with monitoring based on pre-defined metrics

This is an interdisciplinary process and requires the incorporation of business, technical, and human aspects within the organization, and should be managed by a person dedicated to bringing these aspects together for joint decision-making and governance.

Conclusion

Agentic AI is not magic, it’s a technical evolution of AI that holds significant potential for value creation of AI also in healthcare and medtech. By understanding the opportunities and limitations of this technology, focusing on high-impact use cases and ensuring alignment among stakeholders, organizations can benefit now from the next wave of AI innovation.

If your organization is exploring how to gain real-world value from AI, reach out to us at Synwisery for tailored approaches to AI adoption. Together with our partner humest GmbH and their human-centric framework for better AI acceptance, we are bridging the critical gap between technological advances and human adoption.