The “Last Mile” Problem: The Key to Unlocking AI’s Potential in Health Care

Greatness is not a function of circumstance. Greatness, it turns out, is largely a matter of conscious choice, and discipline

-Jim Collins, Good to Great

While many people tend to kick off a new year with a set of predictions for what is to come in healthcare, I am only going to make one: In 2025, organizations using AI in health care will diverge in success - based mostly on how well they address the “last mile problem.” This single prediction may seem narrow, but it actually encompasses the biggest barriers to realizing AI’s transformative potential in clinical settings.

If you imagine AI in healthcare as a pipeline, data science teams develop models, researchers refine algorithms, and leaders champion innovation. But the biggest stumbling block is the final link in the chain: the human component of deploying and using AI tools in real-world workflows. This last step—the one where doctors, nurses, and health system personnel actually need to interact with, trust, and engage with these tools—is often treated as an afterthought or left unaddressed entirely.

In retail or logistics, “the last mile” typically refers to delivering a product from a distribution hub to a customer’s doorstep. For AI in healthcare, the analogy holds: advanced insights must be “delivered” into the hands of healthcare practitioners and patients at the right place, time, and format. You might have an excellent risk stratification model that identifies high-risk patients with unprecedented accuracy. But if it’s buried in a software system that clinicians rarely open, or if it pops up in a way that disrupts workflow, it won’t be used. And if nobody uses it, you might as well not have built it.

Key points to keep in mind:

• The best AI solutions can fail if they’re not embedded into the standard clinical process.

• Clinicians and staff have limited time and are bombarded with alerts, messages, and tasks.

• Effective user interface (UI) and user experience (UX) design is not a luxury— it’s a necessity.

AI in healthcare generates a lot of hype. The potential is indeed huge—predictive models can flag disease earlier, large language models (LLMs) can capture notes automatically, and machine learning can suggest best practices for care. Yet, despite these capabilities, there remains a tendency to view AI primarily as a purely technical endeavor: gather data, build models, deliver insights.

Here are some reasons why the last mile problem is frequently overlooked:

1. Overemphasis on Model Performance
   It’s easy to get swept up in discussions about algorithmic accuracy, F1 scores, or ROC curves. While these metrics matter, they don’t necessarily translate into real-world adoption. You can have a perfect algorithm on paper, but if it requires clinicians to leave their usual workflow, adoption is likely to stall.

2. Misaligned Incentives
   Innovation teams and data scientists might be rewarded for how quickly they can build new AI prototypes, not necessarily for successful adoption or integration. What gets measured often gets prioritized, so bridging the last mile is sometimes relegated to the bottom of a long to-do list.

3. Time and Budget Constraints
   Healthcare systems face myriad financial and operational pressures. Implementing new technology demands resources—not just licenses and servers, but also training hours, revised protocols, and staff buy-in. It’s much easier to secure funding for a cutting-edge pilot than for a lengthy implementation plan.

4. Cultural Resistance to Change
   Healthcare is inherently conservative. The stakes are high, and anything that threatens to disrupt routines can meet resistance. Organizations that don’t understand the depth of this cultural context are likely to face a significant last mile challenges.

One of the most promising AI use cases is ambient AI for clinical documentation. In theory, this is a dream scenario: a “virtual scribe” passively listening to patient-doctor encounters, automatically capturing and structuring notes, and freeing physicians to focus on the patient instead of a keyboard. It sounds like a no-brainer.

Yet many health systems have implemented it and discover that up to half of doctors actually use the feature after it’s been deployed. On the surface, it may look like the technology simply isn’t a fit for those clinicians. But deeper investigation often reveals a classic last mile problem:

• Workflow Disruption: Clinicians must change what they’ve been doing for decades and remember to activate the AI tool or log in to a separate application.

• Privacy Concerns: Patients or providers worry about being recorded.

• Trust in Accuracy: Doctors may feel that if the AI output requires heavy edits, it’s not worth the time saved.

• Lack of Seamless Integration: The final notes might not integrate smoothly with the electronic health record (EHR) system.

All of these represent breakdowns in the last mile of AI. The problem is less about the technology being inadequate and more about insufficient alignment with clinician needs and workflows.

Another scenario that highlights the last mile challenge is the use of predictive analytics to identify high-risk patients—say, those at risk of readmission or those who would benefit most from a particular intervention. Once again, the data science might be robust and the analytics powerful, yet the model’s output can go ignored by clinicians.

Why? A few typical reasons:

• Lack of Context: The risk score is visible in a dashboard clinicians don’t have time to check.

• Insufficient Education: Providers may not understand how the model arrives at the score or how to interpret it in clinical decision-making.

• Inadequate Workflow Triggers: Instead of surfacing alerts within the normal EHR order process, a risk score is merely appended as an extra detail, lost among dozens of other data points.

• No Clear Next Steps: Even if a clinician knows the patient is at high risk, there might not be a well-defined set of actions to take.

The takeaway: just producing a predictive score doesn’t ensure it will be used to change healthcare. Organizations must deliver not just a prediction but also practical steps for intervention, embedded in a workflow that makes it easier (not harder) for clinicians to act on it.

Having recognized the urgency of the last mile challenge, how do you tackle it? Below is a framework designed to help healthcare leaders and AI champions systematically plan for last mile adoption.

• Observe Current Processes: Spend time shadowing clinicians and staff. Map out exactly how decisions are made, what screens they see, and how they communicate.

• Identify Pain Points: Where does friction exist? Is it in note-taking, patient triage, scheduling, or discharge planning?

• Validate with End Users: Present your workflow maps to actual users and see if they recognize them. Validate assumptions early.

• Co-Design & Feedback: Don’t just gather requirements from clinicians—co-design with them. Involve them in early prototyping, user testing, and pilot evaluations.

• Patient Perspective: For tools that affect the doctor-patient relationship (like ambient AI), don’t forget to include the patient’s comfort and preferences as part of the design.

• Embed in Existing Systems: If doctors live in the EHR, that’s where your AI tool should reside. Minimize extra logins and screens whenever possible.

• Use Visual Cues & Automated Alerts: Intelligent design can highlight critical information in a user-friendly format. But be cautious of “alert fatigue.” Strike a balance between relevance and urgency.

• Design for Speed: If using your AI solution slows clinicians down, they’ll abandon it. Optimize for minimal clicks and near-zero load times.

• Translate Data into Guidance: A risk score is only as good as the next steps it prompts. Provide clear recommendations or protocols for follow-up.

• Contextualized Information: If your model identifies pneumonia risk, embed pneumonia-specific management guidelines right where the score appears.

• Robust Training: Offer training materials that address both the “how” (button clicks) and the “why” (clinical rationale).

• Champions & Super Users: Cultivate a group of early adopters and “super users” who can provide peer-to-peer support.

• Ongoing Support & Iteration: The first rollout is rarely perfect. Iterate based on feedback, and make sure there’s a support line for troubleshooting.

• Metrics & Dashboards: Track usage metrics (e.g., frequency of feature use) and clinical outcomes (e.g., reduced readmissions).

• Regular Check-ins: Survey users to gauge satisfaction, gather improvement ideas, and verify that the tool is meeting its intended goals.

• Reward Success: Recognize clinics and individuals who effectively leverage the AI tool and see improved patient outcomes.

1. Treating the Last Mile as a Single Step: It’s a series of small steps—each one can derail your rollout if overlooked.

2. Underestimating Time for Implementation: Plan for training, pilot testing, revision cycles, and full-scale rollout.

3. Skipping Pilot Programs: Pilots reveal hidden workflow issues and user concerns. They’re invaluable for refining your approach.

4. Ignoring Cultural and Ethical Considerations: If staff feel the AI tool is “watching” them or judge it as another administrative burden, trust breaks down.

5. Failing to Align with Organizational Strategy: AI projects need champions at all levels, from department heads to frontline staff.

Healthcare organizations face growing pressure to deliver better outcomes with fewer resources. AI holds tremendous promise to optimize care, but implementing AI poorly can backfire: wasted time, eroded trust in technology, and a reputation for “tech hype” overshadowing true value. In a time when burnout is at an all-time high, solutions that genuinely make clinicians’ lives easier—and more importantly, improve patient care—are not just nice to have; they’re essential.

Moreover, the AI landscape itself is shifting. Tools are becoming more intuitive and user-friendly, but that doesn’t automatically solve the last mile challenge. In fact, the proliferation of AI solutions can increase complexity, leading clinicians to tune out or reject yet another tool unless it is seamlessly integrated into their workflow.

My one prediction: the organizations that truly move the needle with AI in healthcare this year will be those that master the last mile problem. It might not sound as flashy but it is, without question, where the real work gets done.

Addressing the last mile is not a box to check at the end of a project; it is a design principle that should inform every decision from day one. It requires rethinking workflows, engaging end users at a deep level, and allocating resources to integration and training. It also means measuring success not just by algorithmic accuracy, but by adoption rates, clinician satisfaction, and patient outcomes.

Organizations that invest in these practical details—and refuse to let them become an afterthought—will find themselves standing apart from the crowd. While everyone else is busy talking about the potential of AI, these organizations will actually be using AI to deliver better care. And that’s what real impact in healthcare looks like.

In the end, a shiny new AI model or an exciting technology demonstration is only the beginning. It’s how you ensure it’s adopted and used in daily practice that matters. Solve the last mile problem, and the rest will follow.

This post is not an endorsement or investing advice. It is personal opinions and does not reflect the views of my past, present or future employers, clients, or colleagues.