Beyond AI Chatbots: Building the Infrastructure Healthcare Actually Needs with Ashley Cairns

In a recent episode of Digital Health Interviews, we spoke with Ashley Cairns - founder and CEO of Keywell, data scientist, and healthcare analytics expert whose career has spanned nearly two decades across Medicaid programs, UnitedHealth Group, Optum, and enterprise healthcare analytics. Having worked inside both public institutions and some of the largest healthcare organizations in the United States, Cairns has spent years studying how health plans use data, where operational bottlenecks emerge, and why so many promising technologies fail to scale.

While much of today's AI conversation revolves around increasingly capable language models, Cairns believes the real opportunity lies elsewhere. Healthcare doesn't lack intelligent models. It lacks the operational foundation that allows organizations to deploy them safely, integrate them with existing infrastructure, and adapt as the technology evolves. In her view, the biggest winners of the AI era won't necessarily build the smartest models; they'll build the systems that make those models usable inside one of the world's most complex industries.

Understanding Healthcare Means Understanding Incentives

Before founding Keywell, Cairns spent nearly seventeen years working in healthcare data science and analytics. Her experience ranges from Medicaid fraud detection and Affordable Care Act implementation to payer analytics, enterprise data products, and predictive modeling. That background gave her a unique perspective on a lesson many first-time healthcare founders only learn after entering the industry: success depends as much on understanding organizational incentives as it does on solving clinical problems.

According to Cairns, many startups enter healthcare with compelling technologies and meaningful missions, promising better outcomes and improved patient experiences. While those goals are undeniably important, they often fail to answer the question that healthcare buyers are actually asking: What problem am I personally trying to solve? Every department inside a health plan operates under different priorities, budgets, and performance metrics. Without understanding those internal incentives and without demonstrating measurable business value, even the most innovative solutions struggle to gain traction.

She points to another reality that surprises many newcomers: reimbursement still shapes innovation. If there is no reimbursement pathway or billing mechanism, introducing new technology becomes significantly more difficult. Healthcare innovation, in other words, is never just about technology. It is also about economics.

AI Is Changing the Economics of Healthcare

One particularly interesting part of the conversation focuses on an unintended consequence of AI adoption.

As documentation tools become increasingly sophisticated, providers are becoming better at capturing every billable detail during patient encounters. AI-powered scribes can identify appropriate procedure codes with greater accuracy and consistency than manual documentation, resulting in more complete reimbursement. While that improves documentation quality, it also increases overall healthcare spending when identical clinical services receive higher reimbursement simply because coding has become more precise.

Cairns believes this creates a new kind of technological arms race. Providers are using AI to optimize billing, while health plans will inevitably develop their own AI systems to verify coding accuracy and challenge inappropriate claims. Rather than reducing administrative complexity, AI could temporarily intensify it as competing algorithms evaluate one another's decisions.

Over time, she expects this dynamic to accelerate broader shifts toward value-based care and episode-based payment models, where reimbursement depends less on individual billing codes and more on patient outcomes. Until then, however, healthcare organizations may experience a period where AI improves efficiency while simultaneously increasing costs.

The Missing Layer Between AI and Healthcare

The creation of Keywell emerged directly from Cairns' experience helping organizations implement data-driven solutions long before generative AI entered the mainstream. When large language models suddenly became capable of automating tasks that previously required data scientists and analysts, it forced her team to rethink not only their products but their entire business model.

What they discovered was that most healthcare organizations were not actually struggling to access AI models. Instead, they lacked everything surrounding those models. Governance frameworks, security controls, compliance monitoring, infrastructure integration, permissions management, workflow orchestration, and long-term maintenance all represented significant barriers to adoption.

Rather than positioning Keywell as another AI platform, Cairns describes it as AI scaffolding - a foundational layer that sits on top of an organization's existing infrastructure and allows health plans to build AI capabilities without replacing their current systems. The term "scaffolding" is intentional. Like construction scaffolding, it provides a stable structure that supports rapid development while allowing organizations to retain ownership of what they ultimately build.

Healthcare Doesn't Need Another SaaS Platform

One of Cairns' more provocative arguments concerns the future of traditional SaaS products.

Healthcare organizations already manage hundreds of specialized software vendors, each solving one narrow operational problem. Prior authorization, appeals processing, member outreach, care coordination, and population health management often require separate platforms that rarely integrate smoothly with one another.

Keywell's long-term vision is different. Instead of selling another point solution, the company aims to provide infrastructure that allows organizations to automate many of these workflows within their own environment. Prior authorization, appeals management, patient 360 initiatives, and population health analytics become examples of capabilities built on the same underlying foundation rather than isolated applications purchased from different vendors.

For Cairns, the shift resembles a broader transformation already happening across enterprise software. Organizations are moving away from application-centric strategies toward data-centric architectures, where shared infrastructure enables new use cases to be added far more quickly over time.

Why Betting on One AI Model Is the Wrong Strategy

If there is one area where Cairns refuses to make predictions, it is the race among frontier AI models.

OpenAI, Anthropic, Google, Meta, and healthcare-specific models continue improving at extraordinary speed, with benchmark rankings changing almost weekly. Rather than committing to a single provider, Keywell has adopted a deliberately flexible strategy. Different models excel at different tasks. Smaller models may be sufficient for classification or routine automation, while larger frontier models remain better suited for complex reasoning. Specialized healthcare models may outperform general-purpose systems in specific clinical applications.

The real engineering challenge, she argues, is not selecting today's best model. It is creating an environment where models can be replaced tomorrow without disrupting production systems.

That problem remains largely unsolved across the industry. Organizations invest months validating a workflow using one model, only to find that a faster and less expensive alternative appears shortly afterward. Switching models is rarely as simple as changing an API. Every replacement requires new validation, testing, monitoring, and governance before it can safely handle real-world healthcare operations.

To address this, Keywell emphasizes testing environments that allow organizations to compare models using their own production data rather than relying solely on public benchmarks. Real healthcare workflows, Cairns argues, provide far more meaningful evidence than generalized performance scores.

AI Governance Is Becoming the Next Competitive Advantage

Healthcare AI discussions often focus on HIPAA compliance, but Cairns believes the conversation needs to become much more nuanced.

HIPAA itself was written decades before large language models existed and therefore offers little direct guidance on many of today's AI-specific questions. The underlying principles - privacy, auditability, access controls, and responsible handling of protected health information - remain highly relevant, but applying those principles to modern AI systems requires significant interpretation.

One area where Keywell invests heavily is auditability. Every interaction involving protected patient information should be traceable at the individual user level, just as traditional electronic health record systems maintain detailed audit logs. Permissions governing data access must remain separate from permissions controlling AI agents, ensuring that organizations always know who accessed information, when they accessed it, and how it was used.

Cairns also notes that many organizations have adopted blanket prohibitions against training AI models on patient data, even though HIPAA itself does not explicitly forbid every form of model training. In many cases, the restrictions arise from contractual concerns and uncertainty about future regulatory interpretation rather than existing law. Until clearer guidance emerges, organizations are choosing caution over experimentation.

Regulation Shouldn't Outpace Innovation

When asked whether healthcare needs entirely new AI regulations, Cairns takes a measured position.

She expects guidance to evolve, but warns that overly prescriptive regulation could quickly become obsolete in a field advancing as rapidly as artificial intelligence. Healthcare has already experienced unintended consequences from earlier regulations that made data sharing unnecessarily difficult, even for patients attempting to access their own medical records. Repeating that pattern could unintentionally slow valuable innovation.

Instead of relying exclusively on formal legislation, she sees an increasingly important role for industry organizations, technical standards, and professional guidance developed by experts who understand both healthcare operations and emerging AI technologies. That collaborative approach may prove more adaptable than regulations attempting to define rules for technologies that change every few months.

Building for the Healthcare That Doesn't Exist Yet

Toward the end of the interview, Cairns offers advice that reflects perhaps the most strategic theme of the entire conversation.

Healthcare founders should not build products solely for today's technology landscape.

They should build for the landscape that will exist two years from now.

Software development is becoming dramatically less expensive as AI accelerates engineering productivity, while the capabilities of intelligent systems continue expanding at an extraordinary pace. Products designed around today's technical limitations may become obsolete surprisingly quickly.

For that reason, Cairns encourages founders to evaluate every idea against two simultaneous trends: software creation becoming increasingly automated, and AI becoming increasingly capable. Companies that succeed will be those building durable infrastructure rather than temporary features that frontier models will eventually absorb.

Infrastructure, Not Intelligence, Will Shape the Next Wave of Healthcare AI

Ashley Cairns offers a perspective that stands apart from much of today's AI conversation. Rather than asking which model is smartest or which chatbot performs best, she asks a more fundamental question: what will allow healthcare organizations to continue benefiting from AI no matter how the technology changes?

Her answer is infrastructure.

Models will improve. Benchmarks will shift. New frontier systems will emerge every few months. But healthcare organizations will continue facing the same operational challenges - integrating data, maintaining compliance, validating outputs, protecting patient privacy, and adapting workflows without disrupting care.

The companies that solve those foundational problems may never receive as much attention as the model creators themselves. Yet they may ultimately play an even more important role in determining how artificial intelligence transforms healthcare over the coming decade.