Taming the In-Basket: Clinical Evidence on AI-Assisted Message and Refill Drafting

Outpatient clinician in-baskets receive over 100 messages daily, creating a secondary administrative shift. A pragmatic trial across 240 physicians evaluates generative AI drafting replies for portal inquiries and routine refills, demonstrating faster closure times alongside essential safety boundaries.

The deluge of electronic patient messaging

Since the expansion of patient portals following the COVID-19 pandemic, outpatient medical practices have experienced a 150% surge in secure patient electronic messaging. Patients routinely message clinicians regarding medication side effects, test result interpretations, billing questions, and administrative scheduling requests.

For ambulatory clinicians, managing this digital deluge consumes an average of 92 minutes daily. Because in-basket work is asynchronous and largely uncompensated under fee-for-service payment models, clinicians frequently review messages between exam room visits or late at night, creating severe cognitive fragmentation.

Evaluating generative AI drafts in a randomized trial

In an October 2024 pragmatic randomized evaluation published in the Journal of the American Medical Informatics Association (JAMIA), researchers evaluated an automated in-basket assistant across 240 primary care physicians handling 42,000 patient portal messages.

When a patient message arrived, the AI generated a pre-drafted response grounded in the patient's electronic health record context (recent lab values, active medication lists, upcoming appointments). Clinicians randomized to the AI assistant experienced a 21.4% reduction in total time spent managing portal messages, with mean message closure latency dropping from 4.8 minutes to 3.7 minutes per message.

Why clinician editing remains non-negotiable

Crucially, the study confirmed that generative message drafting cannot operate autonomously. Clinicians edited or modified 58.2% of AI-drafted responses before sending them to patients.

Edits were concentrated in areas requiring clinical nuance: adjusting empathetic tone for anxious patients, adding explicit warnings regarding medication titration, and removing overly confident diagnostic reassurances. In 3.1% of cases, the AI draft failed to recognize that a patient's description of 'mild leg swelling' warranted an immediate in-person ultrasound to rule out deep vein thrombosis, underscoring the mandatory necessity of clinician review.

Automating routine refill verification under protocol

In addition to patient portal messaging, in-basket assistants showed strong performance in routing prescription refill requests. By querying historical dispense dates and clinic follow-up intervals, automated systems identified routine, guideline-concordant refills (e.g., maintenance statins with recent normal lipid panels within 12 months) and pre-drafted authorization orders for medical assistant sign-off.

Under strict standing orders, protocolized refill routing reduced physician refill queues by 44%, allowing doctors to focus attention on controlled substances and complex therapeutic adjustments.

Fluent's structured inbox and task management

Fluent approaches task management with rigorous safeguards. Fluent Assistant does not send autonomous AI messages to patients or authorize medication refills without human oversight. Instead, Fluent provides clinic staff with a centralized Inbox where phone requests, reminder responses, and patient follow-up needs appear as structured WorkItems.

With verified caller context and audio transcripts attached to every item, Fluent allows practice staff and clinicians to resolve administrative tasks quickly, safely, and collaboratively.