What AI in Healthcare Is Really Becoming?

From agentic systems to workflow-native infrastructure, the winners always own the work flows.

It’s clear we’re still deep in the AI hype cycle. By 2025, things have shifted. The demos are tighter. The questions are smarter. And more importantly, the real value of AI is finally showing up in workflows, not just slide decks.

The story is no longer about who has the best GPT wrapper. It’s about who owns the work, and what that means for automation, defensibility, and scalability.

From Wrappers to Workhorses: The Rise of Agentic Systems

Late 2023 and most of 2024 were flooded with AI features. Smart summarizers, auto coders, and co-pilots all looked impressive, but a lot of them saw resistance to adoption due to a lack of integration.

That’s finally changing. In 2025, we’re seeing a clear shift toward agentic systems and the AI tools that not only summarize but also observe, reason, act, and improve embedded within the workflow. The best examples we’re seeing today are systems that can orchestrate real workflows, automating multi-step tasks like claims scrubbing, benefits verification, prior authorization packet preparation, or referral triage. At the same time, they’re building internal command centers that act as context-aware copilots. These aren’t just fancy chatbots. They’re pulling from all the messy internal data, such as EMRs, CRM notes, claims history, faxed records, and call transcripts, to surface what matters and help ops teams move the needle.

These systems have become part of the team. Instead of shaving minutes off a task, they’re wired directly into the operational stack and starting to absorb various functions.

But agentic capability alone doesn’t create value. What it does unlock is upside potential in terms of automation capacity, bandwidth, and speed. The real question is: where is the unmet need? Is it a pain point that past AI couldn’t fix well, or a chance to redesign the workflow entirely?

That’s where the next layer of defensibility is emerging.

Owning the Workflow = Owning the Moat

AI’s value is proportional to where it lives. A lot of teams are still building a co-pilot here and a tool there, but the stickiest ones are embedding AI deep into the workflow itself.

IMO, that rule of thumb hasn’t changed. What was true in the early ML days, when models had to be baked into care management or claims ops to matter, it is still true today with LLMs and agents — AI is only as valuable as the workflow it controls.

One example from 2023/2024 is the AI call center workflow.

Provider groups, payers, and pharma all operate complex call centers that handle everything from appointment reminders to benefit verification. At first, these looked like low-margin automation targets. But the smartest companies used them as an entry point.

Instead of stopping at AI-generated responses, they built orchestration layers that:

• Trigger downstream actions in EMRs, rev cycle, or CRM

• Extract structured data from unstructured calls

• Surface upstream operational bottlenecks

Call centers became the integration point for broader backend automation. That turned “voice AI” into infrastructure.

On the flip side, voice itself is a commodity. From an investor lens, we’re often evaluating the best team among dozens of call center plays. Each claims that they grow from $0 to $1M ARR in a few months, but under the hood, there’s often a big gap between booked ARR and contractual ARR. This largely happened because customers were also intrigued by the phenomenal AI capability and were testing a handful to a dozen products at first. Now we see the trend starting to fade because of fatigue. I’ve also heard from other investors about the same trend repetitively, when many are still wrapping their heads around which teams can execute beyond the wedge.

If you zoom out, it becomes a question of approach. Teams that take a workflow-first perspective by starting with a real operational gap can assemble commoditized apps like voice or chat around the core pain. It’s not about whether voice AI is “wrong,” but whether it’s rooted in something painful and durable.

That’s the approach we lean toward at CHAP Health: start with unmet need, build for the workflow, and expand outwards.

We’re seeing that play out across our portfolio:

• Laptis (laptis.io) works with behavioral treatment centers, healthcare systems, and payors to speed up the SUD patient referral process by using AI agents. But the real unlock isn’t communication automation but the referral infrastructure gap. Delays in behavioral care increase readmission risk and payer cost. On the patient receiving side, Laptis’s agentic layer acts on referrals, tracks intake progress, and closes the loop, which EMRs weren’t built to do.

• Tabflows (tabflows.com) is helping independent physician groups that sit outside Epic/Cerner ecosystems. They provide a glue layer that combines agentic task orchestration with a command center for coordinating administrative tasks as well as retrieving critical clinical information from partners, e.g., diagnostic labs and specialty clinics. For physician-entrepreneurs, it’s a way to access enterprise-grade automation without the overhead.

AI Is Cheaper to Build, and the Best Are Moving Fast

One of the biggest tailwinds in 2025 has proved that it’s never been easier and cheaper to build real AI infrastructure due to;

• Foundation models become increasingly cost-efficient

• Infrastructure is modular, e.g., LangChain, LlamaIndex, vector DBs, hosted inference endpoints

• Agentic orchestration can be built and tested in weeks instead of months

This has compressed build timelines and redefined what a seed-stage company can look like. Teams that used to take 12–18 months to reach real revenue are now showing up with working products, first customers, and pilot expansion paths within the first quarter!

And it’s not just about speed. It’s also about what’s now possible and where LLMs are leapfrogging traditional ML-based systems.

Take claims processing. Traditional approaches rely on statistical ML or hardcoded rules. They work until they don’t, i.e., when payer logic changes or policies get updated. This leads to leakage, denials, and escalating manual work.

With reasoning built in, LLMs can determine “what needs to be done” instead of guessing “what might go through.” That means identifying missing documentation, adjusting codes contextually, or escalating edge cases earlier. They’re compensating for the brittleness of the systems instead of just navigating around it.

This shift from probabilistic optimization to reasoning-based orchestration is unlocking a new wave of investment landscape. On one hand, private equity firms that own service-heavy platforms, e.g., billing groups, call centers, or staffing businesses, are picking up AI vendors to automate core processes and drive margin expansion. On the other hand, AI-native startups backed by venture capital are starting to acquire traditional service organizations. Not just for revenue or distribution, but because those service companies sit at the center of real workflows and often come with operational data that would be very costly to collect from scratch.

And it’s not just about revenue or relationships. When AI-first companies acquire traditional services, they’re buying longitudinal operational data, which will become the foundation for rich internal data lakes powering agents across the stack.

The upside is real: AI-native infra can materially expand EBITDA, and that uplift will increasingly be reflected in enterprise value multiples. Turning a low-multiple services business into a high-multiple AI ops platform is actually in motion.

We’re seeing that firsthand with our portfolio company Avant Health (avanthealth.ai). Avant is a full-stack, AI-native TPA that integrates directly into legacy insurance systems—covering everything from eligibility and claims to PBM, stop-loss, utilization management, and more. But the real value goes beyond systems integration. Avant enables a dramatically better member experience through benefit explanation and real-time network navigation, while also delivering rich, actionable employer insights into how insurance dollars are being spent. The transformation happens as Avant is moving insurance from a world of reactive plan design and fragmented care coordination toward a model of proactive, data-driven insurance guidance. It’s not just improving the customer experience. It is reframing what a modern TPA can be.

To summarize it, AI in healthcare isn’t cooling off. It is maturing. The gap is widening between teams building features and teams building infrastructure, and between those who start with shiny tech and those who start with painful workflows.

At CHAP Health, our bread and butter is pre-seed. We work with founders from zero to one, when the wedge is still forming, when the workflows are still messy, and when the AI isn’t just helping, it’s doing.

Qi

CHAP Health Ventures