AI in Medicine: How It's Actually Reshaping Clinical Workflows

The phrase "AI in medicine" gets used to describe everything from a radiology algorithm that flags pulmonary nodules to a chatbot answering patient portal messages. That breadth makes it nearly useless as a search term — and yet it's what many clinicians and administrators actually type when they're trying to understand what's being deployed around them.

This entry focuses specifically on AI in clinical workflows: the tools that sit inside or adjacent to the care encounter itself — not imaging analysis, not drug discovery, not population health dashboards. The three categories that have reached meaningful deployment scale are ambient documentation assistants, EHR-embedded clinical decision support, and AI-assisted medical coding. Each has a distinct evidence base, regulatory posture, and set of practical limitations.

Ambient Documentation: The Highest-Adoption Category

Ambient AI scribes — tools that listen to a clinical encounter and generate a draft note — have seen faster adoption than almost any other category of clinical AI. By early 2026, several large US health systems had rolled out ambient documentation to thousands of clinicians, and the category had attracted significant vendor activity from both established EHR companies and dedicated startups.

The appeal is straightforward: documentation burden is one of the most consistently cited drivers of physician burnout, and ambient tools promise to reduce the time spent typing after patient visits. Published studies have reported reductions in after-hours documentation time ranging from 20 to 40 minutes per clinician per day in some settings — though these figures vary considerably by specialty, visit type, and how "documentation time" is defined in each study.

Regulatory Status

Most ambient AI scribe products are not FDA-cleared. They are generally positioned as administrative tools that assist with documentation rather than as Software as a Medical Device (SaMD) — a classification that would trigger FDA oversight. This positioning is not without controversy: when a tool's output influences clinical decision-making (as a note that gets signed and placed in the record inevitably does), the boundary between "administrative" and "clinical" becomes contested.

What the Evidence Actually Shows

The published evidence on ambient scribes is growing but uneven. Most studies are prospective cohort designs or pre/post comparisons at single institutions. Randomized controlled trials are rare. External validation — testing whether results from one health system replicate at another — is limited.

Reported outcomes cluster around three dimensions: documentation time, clinician satisfaction, and note quality. Time reduction findings are the most consistent. Satisfaction scores tend to be positive in early adoption phases but show more variance at 6–12 months, particularly in specialties where documentation requirements are highly structured (e.g., psychiatry, oncology). Note quality — accuracy, completeness, absence of hallucinated content — is the least well-studied dimension and arguably the most clinically significant.

EHR-Embedded Clinical Decision Support

Clinical decision support (CDS) tools embedded in EHR systems represent the oldest category of AI in clinical workflows — and the one with the most complicated track record. Rule-based CDS alerts have existed for decades; what's changed is the move toward ML-driven predictions that generate alerts or risk scores from patterns in structured EHR data rather than from manually coded rules.

Common Applications

• Sepsis prediction models that score patients on likelihood of deterioration using vital signs, lab values, and nursing assessments
• Readmission risk scoring that flags patients at high risk of 30-day readmission at time of discharge
• Deterioration alerts in inpatient settings that trigger earlier escalation to rapid response teams
• Medication interaction and dosing alerts, including AI-augmented versions that attempt to reduce alert fatigue by better contextualizing risk
• Diagnostic support tools that surface relevant differential diagnoses or order suggestions based on presenting information

The Alert Fatigue Problem

The central operational challenge for EHR-embedded CDS is not accuracy — it's uptake. Studies consistently find that clinicians override the majority of CDS alerts, often exceeding 90% override rates for certain alert types. This isn't always clinician error; many alerts fire in contexts where the clinician has already accounted for the flagged risk. But high override rates mean that even a well-validated model may have minimal real-world impact.

AI-driven CDS has attempted to address this by making alerts more selective — firing fewer of them, better targeted. The evidence on whether this improves outcomes (rather than just reducing alert volume) is still developing. A few prospective studies have shown mortality reductions associated with sepsis prediction tools, but these results haven't replicated uniformly across institutions, and the contribution of the AI component versus accompanying workflow changes is often difficult to isolate.

Regulatory and Certification Landscape

Some EHR-embedded CDS tools fall under FDA oversight as SaMD, particularly those that provide patient-specific recommendations for treatment decisions. Others qualify for the "clinical decision support" exemption under the 21st Century Cures Act, which excludes certain CDS software from FDA device classification when it meets specific criteria around clinical basis and clinician override capability.

The ONC's information blocking rules and interoperability requirements also intersect with CDS deployment — particularly for tools that rely on data from multiple sources or that operate across EHR platforms. Health IT staff deploying these tools need to track both regulatory tracks simultaneously.

AI-Assisted Medical Coding and Revenue Cycle

Medical coding — translating clinical documentation into ICD-10, CPT, and HCC codes for billing — is one of the clearest fits for NLP-based AI in healthcare. The task is well-defined, the training data is abundant (decades of coded records), and the cost of errors is measurable in denied claims and compliance exposure.

AI coding tools range from computer-assisted coding (CAC) that suggests codes for human review, to fully automated coding workflows for lower-complexity encounter types. The more mature products in this space have been in deployment for several years and have accumulated real-world performance data, though most of that data is vendor-disclosed rather than independently published.

Comparing the Three Categories

What Clinicians and Administrators Should Actually Verify

The gap between a vendor's deployment claims and what a clinician or procurement team can actually verify is wide in this space. Before deploying any of these tool categories, there are specific questions worth pressing on.

1. Is there peer-reviewed evidence, or only vendor-disclosed data? Vendor white papers are not equivalent to published studies. Ask whether the performance claims have been independently replicated.
2. Was the evidence generated at an institution similar to yours? A model trained and validated at an academic medical center may perform differently at a community hospital with a different patient mix, EHR configuration, and documentation culture.
3. What is the FDA or ONC regulatory status, and does the vendor's classification hold up? Some tools are positioned as non-regulated when their actual function could support a different classification. Your legal and compliance team should review this independently.
4. How are errors surfaced and corrected? For ambient documentation tools especially, what is the workflow for catching and correcting AI-generated errors before they enter the permanent record?
5. Are there known performance disparities by patient population? Algorithmic bias in clinical AI is documented across multiple application areas. Ask vendors whether their tools have been evaluated for differential performance across race, language, age, or other demographic dimensions.

The Generative AI Layer

Large language models now underpin most ambient documentation products and are increasingly embedded in CDS interfaces for tasks like summarization, prior authorization drafting, and patient communication. This introduces capabilities that weren't possible with earlier rule-based or classical ML approaches — and risks that are qualitatively different.

The hallucination problem — where a generative model produces plausible-sounding but factually incorrect output — is particularly consequential in clinical documentation. An AI-generated note that includes a medication the patient wasn't prescribed, or omits an allergy that was discussed, creates patient safety exposure that doesn't exist with a blank template.

Equity Considerations in Workflow AI

Clinical workflow AI tools are often discussed as though their benefits are uniformly distributed — less documentation burden for all physicians, better decision support for all patients. The reality is more complicated.

Ambient documentation tools trained predominantly on English-language encounters may perform worse for clinicians conducting visits in other languages or for patients with limited English proficiency where interpreters are involved. Speech recognition accuracy varies by accent and dialect. CDS models trained on EHR data from majority-white patient populations may generate systematically different risk scores for patients from underrepresented groups — a pattern documented in sepsis prediction and readmission risk tools.

These aren't hypothetical concerns. They're documented in published literature. Procurement teams and clinical informatics staff should ask vendors for subgroup performance data, and should treat "we haven't found any disparities" differently from "we've specifically evaluated for disparities and here's the data."

Where the Evidence Gaps Are

The field is moving faster than the evidence. Several gaps are worth flagging explicitly for anyone trying to make deployment decisions based on the current literature.

• Long-term outcomes data is almost entirely absent. Most studies follow cohorts for weeks or months; almost none track clinical outcomes over years.
• Multi-site external validation remains rare. A tool that performs well at its development institution may behave differently elsewhere, and few published studies test this directly.
• Note quality and error rates for ambient AI tools are understudied compared to efficiency metrics. Time savings are easier to measure than documentation accuracy.
• Interaction effects between multiple deployed AI tools in the same workflow are essentially unstudied. Most health systems are now running several AI tools simultaneously; how they interact is unknown.
• Patient perspectives on AI involvement in their care — particularly awareness and consent — are rarely reported in deployment studies.