AI scribes: The promise, the uptake pattern–and some pitfalls in the mix

AI scribes for clinicians are clearly a compelling idea—but to what extent are adoption and results matching up? It can be hard to tell, given persistent gaps in ROI evidence and limited adoption benchmarks, In this post, we take a look at what's what as of today in this space. I will briefly recap the vision (burnout relief for physicians!), unpack how this tech works, explain where its limits lie (today), and take a look at what's known about patterns of adoption in the healthcare industry. Finally, I will take a look at how the ROI story is evolving—in many ways the most interesting part of the story.

Recapping the vision: AI scribes as burnout relief for physicians

It's easy to understand the appeal of AI scribe tech for physicians. You can imagine a before-and-after pair of dramatically different scenes.

• In scene 1, it’s 7:45 PM. Somewhere in the U.S., a physician is toggling between tabs, reconstructing notes from a full clinic day. Her memory of the 11:10 AM visit is fading, the chart’s still blank, and her inbox is overflowing. Multiply that scenario system-wide and you get the modern documentation crisis: burnout by a thousand clicks.

  One study  found that for every hour of patient care, physicians spend nearly two hours in the EHR. The fallout: Fatigue, late notes, missed details, and clinician churn. Nothing good here for patients, providers, or the organizations that employ the providers.

• In scene 2, enter the AI scribe—not with fanfare, but with a quiet notification: “Note ready for review.” At organizations using ambient AI, physicians often close notes minutes after the visit ends.

  There are plenty of glowing statistics about what this means for the workforce. A June 2025 AMA study showed the average U.S. physician using an AI scribe saves 6.3 minutes per patient encounter—totaling over 15,000 hours across 2.5 million patient visits. Other AI scribes research shows that clinicians report better patient interactions (84%), greater patient focus (79%), and reduced burnout (~40% in early pilots).

  Advocates say an overall more efficient clinician workforce could translate into improved access and quality of care for patients. The tech is also positioned as benefitting the business side of provider organizations: Faster, more accurate notes would create major efficiencies in billing and reimbursement—including, potentially, a way to cut down on denials—a major pain point for providers today. And so on.

Operational detail: What AI scribes do today

AI scribes are not clinical decision support software (CDSS) or other types of autonomous agents. Instead, they are solely meant to be workflow allies for clinicians. AI scribe tools passively listen during clinician-patient encounters, using ambient natural language processing (NLP) and large language models (LLMs) to capture, summarize, and structure the conversation into usable clinical documentation.

Here’s what today’s AI scribes do well:

• Structured clinical note generation. AI scribes convert unstructured conversation into SOAP (Subjective, Objective, Assessment, Plan) notes, referral letters, discharge instructions, and visit summaries, often directly into the EHR.

• Ambient encounter capture. Unlike traditional voice recognition software, most modern AI scribes operate in the background, requiring no voice commands or manual transcription.

• Process-aware documentation. Some tools go beyond the conversation to capture screens, orders, or mouse clicks to contextualize the interaction for more complete documentation.

• Customizable workflows. Many scribes allow for user customization, ranging from preferred formatting and note length to specialty-specific content inclusion (e.g., problem lists, assessment logic).

• Multimodal inputs and outputs. In addition to ambient audio, some solutions incorporate keystrokes or cursor behavior. Outputs may be structured notes, clinical summaries, or even templated EHR entries that can be reviewed, edited, and signed by the clinician.

What AI scribes don’t do (and why that matters)

Despite their rapid progress, AI scribes operate within real limits, defined by regulation, liability, and clinical norms. Under today’s rules, many of the tasks below would qualify as medical device functions or carry risks that health systems aren’t yet ready to take on. Still, examples from adjacent digital health tools suggest these boundaries may not hold forever.

• Do NOT: Diagnose or recommend care. Scribes don’t generate differentials or suggest treatments. That would trigger FDA scrutiny as software-as-a-medical-device. But tools like Penda Health’s AI Consult already surface potential diagnostic misses in real time—pointing toward future integration.

• Do NOT: Assign billing codes or submit claims. They may support downstream documentation improvement (e.g., via clinical specificity), but coding requires manual validation or additional tools like computer assisted coding (CAC) or autonomous medical coding platforms. But as vendors move toward tighter integrations, the boundary is starting to blur. Unified platforms that handle both ambient documentation and autonomous medical coding are emerging, and a few have already demonstrated end-to-end functionality under pilot conditions.

• Do NOT: Interface directly with payer systems. True bidirectional payer connectivity remains rare—even for mature RCM platforms. But as a directional proof point, VBC enablement platforms like Aledade and Lumeris are already exchanging quality and performance data with payers on behalf of their risk-bearing provider clients. While not explicitly AI scribe vendors themselves, they suggest a future where AI-powered documentation tools could eventually be embedded more deeply into the payment feedback loop.

• Do NOT: Flag care variation. Scribes don’t yet alert providers to deviations from clinical guidelines—a risky move without regulatory clearance. But this function could emerge as digital quality measures and pathways get baked into EHRs.

• Do NOT: Replace human review. Final notes still require clinician sign-off for legal and clinical accuracy. As Dr. Sara Murray, Chief AI Officer at UCSF Health noted, ”These tools draft notes and patient instructions for doctors … But they still need to be reviewed and edited by physicians to ensure everything’s accurate.”

In short: AI scribes don’t make clinical decisions or run revenue cycle operations. But they can reduce the cognitive and clerical workload that contributes to physician burnout—particularly in high-volume, documentation-heavy settings. They’re also increasingly being deployed in tandem with complementary tools like clinical documentation improvement (CDI) software, computer-assisted coding (CAC )systems, and autonomous medical coding platforms—creating a more integrated, intelligent clinical documentation tech stack.

Untangling the new RCM stack: What sits beside AI scribes?

Setting aside the common vision of AI scribes as burnout reducers for physicians, let's turn our attention to the more behind-the-scenes question of how they fit into revenue cycle management (RCM)

If AI scribes increasingly shoulder the burden of real-time documentation, how will these tools fit into the broader constellation of AI-powered revenue cycle management (RCM) solutions?

The short answer: They don’t replace other tools—they work alongside them.

Case in point: clinical documentation improvement (CDI) tools have traditionally operated downstream—flagging gaps in specificity after notes were written. But that’s changing. As AI scribes capture richer, real-time clinical narratives, a new generation of CDI tools is stepping in upstream. Rather than “Hey doctor, please revise this,” we’re now seeing “Hey doctor, say this” nudges embedded directly into the documentation process. AI scribe vendors like Nuance and 3M are already piloting ambient CDI guidance that suggests phrasing during or immediately after the encounter—helping clinicians document with greater specificity from the start. This shift from reactive to proactive CDI has big implications: more accurate coding, fewer retrospective queries, and a smoother workflow for physicians

Computer-assisted coding (CAC) tools come next, scanning finalized notes to suggest relevant billing codes. Unlike scribes, which focus on front-end documentation, CAC operates closer to the billing line—augmenting coders with algorithmic suggestions. Some systems still require a coder’s final sign-off, while others operate with greater autonomy.

That brings us to autonomous medical coding: the logical evolution of CAC. These tools aim to fully automate the coding process, translating documentation into billable codes in real time, often with little to no human intervention. While not yet universal, adoption is accelerating—especially among larger systems grappling with coder shortages and scaling complexity. But there’s more on the horizon. Advanced ML coding engines are beginning to actively “learn” what’s at risk, based on payer-specific denial trends, and modify output dynamically.

For example: Experian Health’s AI Advantage, integrated into its ClaimSource system, uses ML to analyze payer data in nearly real-time and proactively pause or modify high-risk claims before they’re sent. It’s not just about generating correct medical codes anymore; it’s about adapting them intelligently, on the fly.

Beyond documentation and coding, revenue cycle automation solutions cast a wider net. These are workflow engines built to eliminate manual handoffs in tasks like prior authorization, eligibility verification, or claims follow-up. They’re complementary to scribes and coders but serve very different functions.

Clinical workflow automation platforms tie it all together. They route documentation to the appropriate stakeholders (e.g., CDI, coding, billing), assign ownership, and flag delays, ensuring handoffs happen smoothly and nothing gets lost in the shuffle.

Finally, revenue optimization tools operate at the strategic level. These aren’t transactional tools but intelligence platforms that help leaders spot patterns, uncover revenue leakage, and evaluate whether their RCM investments are delivering measurable ROI.

It's important to remember that today, each of these tools typically tackles a different part of the documentation-to-revenue chain, differentiating which tools do what, where they overlap, and how to deploy them together without investing in duplicative capabilities (not easy to do when every point solution tends to want to expand its purview!)..

AI scribe adoption is accelerating–with variation by segment

Not only is investment in AI scribes way up— it more than doubled from $390M in 2023 to $800M in 2024—systems of all types—IDNs, ambulatory groups, ACOs—are deploying the tools. Some health systems now report more than 90% of PCPs using AI scribes daily. One Mass General Brigham exec described ambient scribe tech as having “commoditized” faster than any other IT tool in the system’s history.  

But the drivers and dynamics of adoption vary by healthcare stakeholder or organization type:

• Health systems (e.g., IDNs and AMCs): Enterprise health systems are embracing AI scribes as part of broader workforce and technology modernization efforts. Adoption is often framed as a retention and burnout-mitigation tool, especially in primary care and frontline specialties. Some health systems are also tying AI scribe use to broader ambient documentation strategies.

• Independent hospitals: For smaller, standalone hospitals, AI scribes are seen as less part of a sweeping digital transformation and more a tactical fix—helping offset gaps from medical scribe staffing shortages and offering an affordable on-ramp to workflow automation.

• Ambulatory groups: High-volume outpatient groups are leaning on less expensive, less robust AI scribe offerings to preserve margins and clinician throughput. Adoption tends to be geared toward access and growth: the tool is a way to increase capacity and productivity without overburdening clinicians.

• Risk-bearing providers / ACOs: Healthcare organizations that embrace risk-based contracting efforts and are financially at risk for attributed patient populations are more likely to also view AI scribes through a compliance and quality lens. The promise here isn’t just time savings, but also better, more structured documentation that supports accurate coding, risk adjustment, quality measurement and performance, care coordination... and the downstream financial performance that hinges on all of these factors.

The ROI is real(ish), but pitfalls remain

So far, so good—but the ROI challenge is a real one, and building a strong ROI case, from a provider perspective, is not easy to do. A 2025 assessment conducted by the Peterson Health Technology Institute (PHTI) found evidence that AI scribes decrease the cognitive load and burnout for clinicians, but determined that finding a tangible financial ROI remains elusive.

Key barriers that cloud the AI scribe ROI case:

1. AI scribe tech is expensive. As Stanford Healthcare’s Chief Data Scientist and Stanford HAI faculty member, Dr. Nigam Shah warned, “Everybody thinks that AI will help us with our access and capacity… but if it increases the cost of care by 20 percent, is that viable?”

2. No standardized ROI metrics or benchmarks. Most systems still rely on anecdotal or qualitative feedback, not hard financial impact metrics.

3. Gains in clinician time don’t always translate to more patients seen. Many physicians want to repurpose saved time toward reclaiming work-life balance—not productivity.

4. “Soft ROI” cases are difficult to quantify. Benefits like reduced burnout, improved retention, and better documentation accuracy are real—but hard to translate into line-item budget justification.

5. RAF-related returns are tricky to message. Improved specificity can improve coding accuracy, which can also boost RAF scores—but few organizations are comfortable tying financial ROI explicitly to risk adjustment increases.

6. Deployment missteps can further erode gains. Inadequate training, unclear documentation protocols, or reliance on outdated models can undercut adoption and ROI.

In summary...

AI scribes are a meaningful quality-of-life investment—and an increasingly relevant part of the RCM tech stack. But they’re not a plug-and-play productivity fix. ROI is possible, but systems must know what kind of return they’re aiming for. We will continue to watch this part of the healthcare tech landscape with interest and we think you all should too.

Comments? Questions? Feel free to reach out to me at chris.loumeau@unionhealthcareinsight.com. And don't forget to subscribe to our blog (scroll down!).