What’s Working in Healthcare AI: Lessons From the Field
By James Forsythe, Virtual Technology Executive Lead
Almost every healthcare leader we work with has heard some version of the same question from their board: “What’s our AI plan?” It’s a fair question, and the pressure behind it is real. According to the AMA’s 2026 physician survey, 81% of physicians now use AI professionally, more than double the share in 2023. AI has made its way into day-to-day clinical and administrative work, whether an organization has a formal strategy in place or not.
Over the past year, our Virtual Technology Executives (VTEs) have sat in planning sessions, operating meetings, and annual reviews with community health centers, specialty groups, and multi-site practices working through AI adoption.
Across these conversations, the organizations seeing real traction shared five habits:
- Start with a measurable operational problem, not a product.
- Prove value with a small, low-cost pilot before committing to broad licensing.
- Put governance, policy, and permissions in place before AI tools reach staff.
- Steer staff toward approved tools instead of one-off AI add-ins.
- Treat data hygiene as the foundation that lets a successful pilot scale safely.
1. Start With the Business Case, Not the Product
The organizations making the most progress aren’t approaching AI as something they need to buy simply because it’s available. They’re evaluating it like any other technology investment: What problem will it solve, and is the investment worth it? Instead of debating vendors, leadership teams map where AI could actually move the needle: ambient clinical documentation, call and scheduling automation, prior authorization, denials and claims work, quality and UDS reporting, etc.
Another good starting point is evaluating your IT service data to find where staff time is actually going. When Abacus evaluates IT service data, we look beyond ticket volume to identify patterns, recurring requests, and opportunities to improve operational efficiency. That insight can help organizations identify where automation and AI may reduce administrative burden, streamline workflows, and free staff to focus on higher-value work. By grounding recommendations in real operational data, we can help healthcare organizations prioritize AI initiatives with measurable impact and build a roadmap around their specific needs.
The lesson is consistent. Teams that start with an operational problem they can measure end up with an AI roadmap. Teams that start with a product end up with just a subscription.
2. Prove Value With a Small, Low-Cost First Step
Once an organization identifies where AI could help, the next step doesn’t have to be a large investment. In many cases, a smaller pilot is a better way to understand what’s useful before committing to additional licensing. A common example is a Microsoft Teams-based assistant that gives staff quick answers from approved internal documentation, such as the employee handbook. Because staff can use it without an individual Copilot license, a pilot like this doesn’t require a large licensing commitment to get started.
We’ve found that starting with a smaller group, such as department directors, gives organizations a chance to see how people actually use the tool and where it falls short. A handful of staff who already hold the right licenses can also test more capable AI agents. Organizations should also be intentional about which document libraries those agents can access, ensuring responses are based on approved internal information.
The advantage of this approach is that leadership can evaluate actual usage, feedback, and potential gaps before deciding whether broader licensing makes sense.
3. Put Governance Ahead of Licensing
Before introducing AI tools across an organization, leadership needs to understand how those tools will be used, what information they can access and what safeguards need to be in place. We’ve seen healthcare leaders take a step back when new AI licensing requests come forward, recognizing that buying licenses before establishing policies and controls can create more problems than it solves.
For example, rather than immediately assigning licenses to staff, organizations can first evaluate the AI capabilities already included in their existing platforms, and determine where additional capabilities are actually needed. This gives leadership time to review and update existing IT and AI policies if needed before expanding access, ensuring the right governance and safeguards are in place.
An important part of that process is also understanding existing file permissions. Enterprise AI assistants inherit the permissions of the person using them. If file permissions are too broad, an AI tool can surface sensitive documents to employees who should never have access to them. Without reviewing those permissions first, an organization could unintentionally expose information simply by introducing a new tool.
This is an area where many healthcare organizations still have work to do. In a January 2026 MGMA poll, only 20% of medical group practices reported having AI governance or a formal AI-use policy in place, and more than half had neither a policy nor one in development.
4. Hold the Line on One-Off AI Add-Ins
Having an AI policy is important, but it also needs to carry over into everyday decisions. Staff regularly request third-party AI add-ins for routine tasks, like building a spreadsheet. The organizations handling this well set a clear rule: no individual exceptions for outside AI tools. Instead, they point users to the approved AI features already built into their existing tools.
This may seem like a small decision, but it helps establish consistency around how company and patient information is handled. Every one-off add-in is another place patient or business data can leave the organization’s governed environment, and the risk is no longer theoretical. IBM’s 2026 Cost of a Data Breach Report found that 43% of breached organizations reported a security incident involving unapproved AI tools, more than double the prior year. Healthcare remained the costliest industry for breaches for the thirteenth consecutive year, at an average of $6.64 million per incident.
Keeping staff within approved tools makes it easier to maintain existing security, retention and compliance controls. It also helps prevent the organization from accumulating multiple AI subscriptions without a clear understanding of who is using them or what data they’re accessing.
5. Build on a Clean Data Foundation
Many of the concerns around AI adoption come back to issues organizations may already have: overly broad access permissions, poorly managed data, and inconsistent file-sharing practices. AI doesn’t create these problems, but it can make them much more apparent. The organizations making steady progress pair AI planning with foundational work: cleaning up access and permissions, classifying and protecting PHI with data loss prevention tools, and tightening how sensitive files are shared externally. Before introducing any new AI tools, organizations need to understand where their data lives, who has access to it, and how it’s being protected. Without that foundation, expanding AI use can introduce unnecessary security and compliance risks.
What Hasn’t Worked As Well
Not every early move has paid off, and the missteps are as instructive as the wins. A few common challenges have come up repeatedly in our conversations with healthcare leaders.
Licensing ahead of a defined use case. At current per-user pricing, putting an AI assistant in front of every clinician and medical assistant is hard to justify without evidence of where it saves time, and several leaders we work with have said exactly that. Starting with targeted pilots gives organizations a better understanding of where premium licenses are worth the investment.
Letting adoption happen informally. When staff begin using AI tools without organizational guidance, it can create security, compliance, and data privacy risks, particularly when sensitive information is involved. Establishing a clear AI policy that defines approved tools, acceptable use, and a process for evaluating new solutions helps organizations adopt AI responsibly while giving staff the guidance they need to use it effectively.
Treating the AI plan as a board deliverable. Plans built to satisfy a board tend to list tools. A useful AI strategy should identify the problems the organization wants to solve, who is responsible for addressing them and how success will be measured. That gives leadership something they can actually put into practice and revisit over time.
Questions worth asking now
When the board asks, “What’s our AI plan?” instead of just providing a list of tools, healthcare leaders should be able to explain where the organization plans to use AI, what it expects to improve, and how it will manage the risks that come with adoption. That means identifying the operational challenges worth addressing, determining which AI capabilities make sense for those needs, establishing the right security and compliance measures, and setting clear expectations for measuring results.
An effective AI strategy should also give leadership a way to prioritize investments, assign ownership and evaluate progress as the technology and regulatory landscape continue to change. It doesn’t have to answer every question upfront, but it should give the organization a clear starting point and a process for making informed decisions as adoption grows.
If your organization is shaping its own approach, these are a few questions worth discussing with your leadership and IT teams:
- Which operational problem would we most like to solve, and how will we measure success?
- Do we know who can access what across our file shares, SharePoint and Teams today?
- Do we have a written AI policy that staff can actually follow?
- Are we getting full value from the AI capabilities already included in our current licensing?
- What is the smallest pilot that would give leadership real evidence?
Frequently asked questions
Do we need to buy Copilot licenses for everyone to get started?
No. Many organizations can start by evaluating the AI capabilities already available through their existing platforms, such as Copilot Chat in Microsoft 365. From there, a targeted pilot can help identify which employees and roles would benefit from the more advanced features included with paid Copilot licenses. The goal isn’t to avoid investing in licenses; it’s to make sure the right people have the right capabilities based on how they work and where AI can deliver value. This allows organizations to expand access strategically rather than committing to an organization-wide rollout before they understand the need.
What is the biggest AI risk for healthcare organizations?
Oversharing and unintended access to sensitive information. AI assistants like Copilot can only access what the user has permission to see, but in many Microsoft 365 environments, those permissions are broader than they should be. Without permission cleanup and data classification, an assistant can surface PHI to staff who have no business need to access it.
Organizations also need to consider how employees access AI across devices. For example, an employee may start a conversation with an AI assistant on their work computer and continue that same conversation on a personal phone if they’re signed into the same account. That means organizations need clear guidelines around approved devices, account access and how sensitive information is handled outside the workplace. AI security isn’t just about which tools employees use; it’s also about where and how they access them.
Does a smaller practice really need a formal AI policy?
Yes. The policy doesn’t need to be lengthy or overly complicated, but it should establish clear expectations for how AI can be used, which tools are approved and what information employees can share. For healthcare organizations, having guidance is especially important. AI policies need to account for HIPAA and other applicable compliance requirements, not just general technology use. Working with an IT and compliance partner who understands the healthcare environment can help smaller practices put the right safeguards in place without having to figure it all out on their own.
Where should we start if we haven’t done anything yet?
Start with a clear understanding of where AI could make a meaningful difference in your organization. Identify one operational challenge you can measure, review who has access to sensitive information across your systems, and establish basic AI guidelines for employees. An AI Risk & Readiness Assessment can help identify gaps, prioritize next steps and build a practical roadmap for adopting AI securely and responsibly.
Ready to take the next step?
If your organization is evaluating how to adopt AI securely and responsibly, learn more about Abacus’ Virtual Technology Executives service, which provides strategic guidance, and our Managed AI Services. From understanding your current environment and identifying potential risks to developing a strategy for implementation, our team can help you determine what makes sense for your organization and how to move forward with confidence.
