A data science team can build a sentiment analysis tool for patient call centers using an LLM API for a few hundred dollars a month in token costs. A vendor will sell you something similar for tens of thousands annually. And increasingly, your EHR vendor is giving your organization tools to build custom AI workflows without writing code or hiring a startup.

So which path is right? The answer depends on more than cost. It depends on your team, your tolerance for risk, and how much governance you’re ready to build around each approach. Whichever mix you choose, the return depends on the people you enable to use it.

The case for building

The economics of building AI tools in-house have changed dramatically. With access to foundation models through APIs, a capable IT or data science team can develop solutions at a fraction of what vendors charge: sentiment analysis on patient calls, intelligent ticket routing, predictive scoring for readmission risk, automated chart abstraction, prior authorization support, patient message triage, supply chain forecasting, and scheduling optimization based on census predictions. A strong team can prototype these tools in weeks at pennies per transaction.

The cost argument is only part of it. Building in-house fosters innovation and gives your team a deeper understanding of the technologies themselves. A case report from Penn Medicine leaders published in *Frontiers of Health Services Management* highlights this: in-house development gives health systems the agility to adapt quickly to emergent needs and build solutions tailored to specific institutional goals. When your team builds, they develop institutional knowledge about AI capabilities and limitations that makes every subsequent decision better informed.

The talent challenge is real, though. McKinsey research on generative AI talent found that 51% of employees who identify as heavy AI users or creators plan to leave their jobs within three to six months. McKinsey’s guide to digital and AI transformation for healthcare payers reinforces this: organizations that have successfully attracted AI talent have done so by emphasizing healthcare’s mission and pairing domain experts with technical specialists. The talent plan needs to be as intentional as the technology plan.

The hidden weight of building

When your best developer leaves, you inherit an application that nobody else fully understands. That risk alone should factor into every build decision. Beyond personnel, that low-cost tool still needs ongoing maintenance: monitoring model drift, updating prompts, testing outputs, and troubleshooting when something breaks at 2 AM on a holiday weekend.

Custom tools also become black boxes to the people using them. If a clinician can’t follow the logic behind a recommendation, adoption stalls. If the organization can’t explain how a tool works to regulators, that’s a compliance risk. Sometimes the vendor subscription premium is buying you security, accountability, and support infrastructure, and that can be worth every dollar.

The case for buying

Vendor solutions come with advantages that are easy to undervalue: dedicated support, regular updates, security certifications, and a contractual relationship that includes accountability. Your technical team stays focused on their core responsibilities instead of becoming a permanent support desk for homegrown tools.

The buy decision makes the most sense when the tool addresses a well-defined, broadly shared problem where vendors have already invested in validation, compliance, and integration. AI-powered chart summarization is a good example. It sounds straightforward: feed clinical notes into a model and get a summary back. In practice, the clinical validation required is substantial. A 2026 study published in npj Health Systems evaluating AI-generated chart summaries found that while physicians were generally positive, they identified omissions in nearly a third of summaries, along with confusing content and hallucinations. Getting chart summarization to a level where clinicians trust it in their workflow requires extensive testing across specialties, patient populations, and documentation styles. Vendors who have invested in that validation across hundreds of health systems offer something most internal teams would struggle to replicate on their own.

Revenue cycle management is another area where buying often wins. AI-powered coding, denial management, and prior authorization tools from established vendors come with trained models, integration with clearinghouses, and regulatory update cycles that reflect the complexity of the payer landscape. The cost of a subscription looks different when you weigh it against the engineering hours, compliance risk, and maintenance burden of building and operating that toolset internally.

The key question is whether the vendor’s solution solves a problem your organization shares with hundreds of others, or whether your needs are specific enough that a custom approach would deliver meaningfully better results.

When your EHR becomes the AI platform

Here’s where the lines between “build” and “buy” start to blur.

At HIMSS 2026, Epic announced Agent Factory, a no-code visual builder that lets organizations design, customize, deploy, and monitor AI agents across clinical, operational, and patient-facing processes. Agent actions are traceable, and organizations can equip agents with local policies and knowledge bases. Epic is positioning Agent Factory as a sandbox for innovative health systems who want to invent and reimagine workflows on their own timeline. For healthcare AI startups, this development is worth watching closely: when your potential customers can build AI agents inside their own EHR, the competitive landscape shifts.

Agent Factory is designed for IT teams, informatics leaders, and data science groups. McKinsey’s analysis of agentic AI in healthcare recommends that organizations be “focused transformers,” picking a few high-impact domains to start rather than scattering efforts across dozens of agents. They also flag the build vs. buy question directly: organizations will need to evaluate whether to build their own agentic systems or partner with startups, and enterprise architects will play an increasingly important role in making those calls.

Clinicians are gaining more control within AI tools that already exist. Epic’s AI Charting, launched in February 2026, lets clinicians personalize how their notes are structured during visits. That kind of configuration is valuable and signals where things are headed.

The harder governance question is what happens outside the EHR: clinicians and staff using general-purpose AI tools like ChatGPT, Claude, or Gemini for clinical tasks and documentation. That activity is already happening at most organizations, often without IT’s knowledge or any formal oversight. IT’s role shifts from building tools to testing, validating, and monitoring what’s being created and used across the organization. The team becomes the quality assurance layer.

Enabling your workforce

Build and buy decisions put tools in place. Enabling decides what your organization gets out of them.

Enabling your workforce starts with training on the AI tools people already have access to. From there it gets more interesting: finding the employees who take to these tools and giving them the authority to build, then reskilling teams so they can manage what gets deployed, implement the next tool well, and spot opportunities nobody upstream had scoped. Employees are asking for this. In McKinsey’s Superagency in the Workplace survey, employees ranked formal training from their organization as the initiative most likely to increase their AI use, ahead of better tool access.

Consider what most health systems already own. Organizations running on Microsoft have Copilot Studio and Power Automate sitting in the same ecosystem as their email, and trained teams can use them to extend the AI products the organization has already purchased. Premera Blue Cross has built 900 agents in Copilot Studio, many by employees without engineering backgrounds; one contract-review agent cut a task that took 30 to 45 minutes down to about three. Brown University Health has built more than two dozen agents covering emergency department guidance, translation, scheduling, and operations. That capability was sitting in the license the whole time.

The same shift shows up in content work. Education and training teams that learn to generate pages, images, and video with AI can produce course materials and communications in a fraction of the usual time, which means the platforms you bought start paying off in places the business case never listed.

Clinicians are asking for this investment too. In the AMA’s 2026 physician survey, more than 80% of physicians reported using AI professionally and 92% said they want more education and training on it. The gap is measurable: KLAS Arch Collaborative research found that fewer than a quarter of clinicians using AI tools received adequate training, and ambient documentation users who know how to get the most from their tool report EHR experience scores 43 points higher than colleagues who use the same tool without that grounding.

You can build a strong portfolio and buy from the best vendors in the market. BCG’s research on AI value finds only about 5% of companies are generating value from AI at scale, and many respond to that gap by adding more tools while underinvesting in the people who use them. Enablement belongs in the AI budget next to build and buy, with real dollars behind it and a named owner.

The governance thread

Every path requires governance, just different kinds. Building internally requires code review, model monitoring, and succession planning. Buying requires rigorous evaluation and contract management. Enabling org-level AI creation through platforms like Agent Factory requires clear policies on what can be built, how it gets tested, and who maintains it. And the growing use of general-purpose AI tools by staff requires discovery, education, and guardrails.

The hardest part is the overlap. A single department might be using a vendor AI tool, a custom-built model from the data science team, an Agent Factory workflow from informatics, and a clinician running patient prep through ChatGPT. Governing each category in isolation misses the point. The real challenge is coordinating across all of them with a consistent approach to oversight, accountability, and risk.

Finding the right balance

The right approach starts with honest answers to a few questions: Do you have the talent to build and sustain custom tools? Are your vendors delivering real value? Is your organization ready to take advantage of no-code AI platforms inside your EHR? And are your clinicians and staff already using AI tools that nobody in IT knows about?

That last question might be the most important one to ask right now.

Frequently asked questions about building, buying, and enabling healthcare AI

How should a health system decide whether to build or buy an AI solution?

Start with the workflow, risk level, internal talent, integration needs, and long-term ownership cost. Building can fit a differentiated workflow when the organization can maintain and monitor the solution. Buying can fit a common problem where a vendor has already invested in validation and support. Enabling applies to every path: training the workforce, developing internal builders, and reskilling teams determine how much value the build or the purchase returns.

When does building healthcare AI in-house make sense?

Building makes sense when the use case is specific to the organization, the internal team has the technical and clinical expertise to support it, and leaders are prepared to fund maintenance after launch. The decision should account for validation, security, model and prompt updates, monitoring, documentation, and succession planning in addition to initial development cost.

What should a health system evaluate when buying an AI product?

Evaluate how the product performs on local data and workflows, what information it accesses, where that information is processed, how the vendor handles model changes, and who is accountable when performance declines. The contract should also address security, data use, support, monitoring, incident response, and what happens to organizational data when the relationship ends.

Does using AI built into Epic or another enterprise platform reduce the need for governance?

Embedded AI may simplify contracting, integration, identity management, and support. The health system still owns local validation, appropriate-use decisions, workforce education, monitoring, and accountability. A feature that is available inside an established platform can affect different patient populations and workflows differently, so availability should trigger a defined review rather than automatic adoption.

Can a health system build and buy at the same time?

Most health systems will. One organization may buy an ambient documentation tool, build an internal analytics model, and configure an agent through its EHR or service-management platform. A shared intake process, AI inventory, risk tiers, validation standards, and named owners allow those approaches to operate under one governance system, and a single enablement program can train the workforce across all of them.