Doctors Signed Off on 92% of AI-Drafted Discharge Notes vs 70% Written the Usual Way. The Lesson for Every Website

Robauto Team · Published · Last updated

Two new hospital studies found AI drafts, co-designed and checked by doctors, made discharge summaries clearer and more complete (92% sign-off ready vs 70%; preferred in 95% of stays). Here's what brands can learn about clear handoffs for AI agents.

Here's a hopeful hospital story that isn't about robots in the hallway. It's about paperwork. When you leave a hospital, someone has to write a discharge summary: what happened, which medicines changed, and what comes next. Two new studies found that AI drafts, built with doctors and checked by doctors, made those notes clearer and more complete. And doctors said they'd sign off on more of them.

The lesson goes well beyond healthcare. A discharge summary is a handoff: messy information turned into something the next person can act on. Every brand website is a handoff too, and more and more often, the next reader is an AI agent.

The headline result: 92% vs 70%

On November 17, researchers at iCARE (Imperial Clinical Analytics, Research and Evaluation) published a study in JAMA Network Open on an AI system they built with resident doctors at Imperial College Healthcare NHS Trust in London. Twelve doctors who write discharge summaries as part of their job reviewed 1,294 summaries without knowing which ones the AI wrote (Imperial College London).

The result: 92% of AI-generated summaries were clinically acceptable, meaning they needed no edits or only minor ones before sign-off. That compares with 70% of clinician-written summaries. Reviewers also rated the AI drafts higher for accuracy, completeness, clarity and tone, and found fewer potentially harmful omissions and inaccuracies (Imperial College London; JAMA Network Open).

Dr Benjamin Atkins, who worked on the study as a resident doctor, described the pressure that makes this matter: "There's a constant tension between clinical tasks that push a patient's care forward and administrative tasks, like discharge summaries, that move patients out of hospital. Both must happen urgently, often for several patients at once and neither can wait" (Imperial).

Lead researcher Kathleen Goldsmith, a senior data scientist at iCARE, credited the approach: "To build transformative clinical AI we need to develop applications alongside the clinicians who will use them, and rigorously evaluate their safety, reliability and clinical value" (Imperial).

A second hospital found the same pattern

On November 2, Stony Brook University reported a study in npj Digital Medicine that looked at the hardest cases: 60 internal medicine stays lasting 7 to 21 days at Stony Brook University Hospital. Hospitalists and primary care physicians compared AI drafts with human-written ones side by side. They preferred the AI-generated summary for 95% of the hospital stays and scored it higher for quality, readability, accuracy and completeness (Stony Brook Medicine).

Primary care doctors, who pick up care after a patient goes home, said the AI summaries were easier to understand, easier to explain to patients and caregivers, and more useful for follow-up. The system pulled a summary together "within just a few minutes" (Stony Brook Medicine).

Both teams were careful about limits. Stony Brook found errors in both AI and human summaries, with no significant difference in how risky those errors were, and says physician review "remain[s] essential" before any AI draft goes into a patient record (Stony Brook Medicine). Imperial's study looked back at past records from its secure data environment. It was not a live ward rollout, and the authors stress keeping humans involved in design, use and evaluation (Imperial). This is AI as a careful first draft, with doctors still in charge.

Patients benefit from plain language too

A randomized study in Neurosurgery (published online October 22) tested AI-rewritten spine imaging reports at a New York academic medical center, with investigators checking each rewrite for accuracy. The reading level dropped from about grade 11.0 to grade 5.9 (Flesch-Kincaid). Patients who got the plain-language version rated it easier to understand, 4.2 vs 2.6 on the survey scale, and reported less trouble with medical terms (Neurosurgery via DOI).

StudyScalePositive resultHuman roleSource
iCARE / Imperial College Healthcare NHS Trust (Nov 17, 2026)1,294 summaries, 12 blinded doctors92% sign-off ready (AI) vs 70% (clinician-written)Co-designed with resident doctorsImperial
Stony Brook University Hospital (Nov 2, 2026)60 complex stays (7-21 days)AI draft preferred in 95% of staysPhysician review stays essentialStony Brook Medicine
Spine imaging RCT, Neurosurgery (Oct 22, 2026)52 patients randomizedReading grade 11.0 to 5.9; ease of understanding 4.2 vs 2.6Investigators checked each rewriteNeurosurgery

Why this matters for your brand

Why did the AI drafts do so well? Each one took scattered notes, tests and medicine changes and turned them into a clear, complete handoff the next reader could act on. Imperial describes the goal as turning "fragmented clinical information into clear, clinically meaningful discharge communication" (Imperial).

AI shopping and research agents read your site the same way. They're looking for a clear handoff: who you are, what you sell, what it costs, how to buy, and what happens if something goes wrong. If that information is spread across PDFs, pop-ups and vague marketing copy, the agent's "summary" of you will have gaps. The good news is that fixing this costs much less than a hospital IT project.

A 5-point "discharge summary" checklist for your website

Bonus: a 15-line reading-level check

This Python script uses only the standard library. It fetches a page and gives a rough Flesch-Kincaid grade, so you can spot pages that are harder to read than they need to be. It's a rough estimate, not a medical or legal readability test.

import re, sys, urllib.request, html
def syl(w):
    w = w.lower()
    n = len(re.findall(r"[aeiouy]+", w))
    if w.endswith("e") and n > 1: n -= 1
    return max(n, 1)
url = sys.argv[1]
raw = urllib.request.urlopen(urllib.request.Request(url, headers={"User-Agent": "grade-check"})).read().decode("utf-8", "ignore")
raw = re.sub(r"(?s)<(script|style|nav|footer).*?</\1>", " ", raw)
text = html.unescape(re.sub(r"<[^>]+>", " ", raw))
sents = max(len(re.findall(r"[.!?]+(\s|$)", text)), 1)
words = re.findall(r"[A-Za-z]+", text)
if not words: sys.exit("no text found")
grade = 0.39*len(words)/sents + 11.8*sum(map(syl, words))/len(words) - 15.59
print(f"{url}: ~grade {grade:.1f} ({len(words)} words, {sents} sentences)")

Run it with python3 grade.py https://yourstore.com/returns. If your returns or pricing page scores well above grade 9, try shorter sentences and simpler words.

Agent-ready summary

{
  "topic": "AI-drafted hospital discharge summaries",
  "date": "2026-11-18",
  "findings": [
    {"study": "iCARE / Imperial College Healthcare NHS Trust", "published": "2026-11-17", "summaries_reviewed": 1294, "blinded_reviewers": 12, "acceptable_ai_pct": 92, "acceptable_clinician_pct": 70, "source": "https://www.imperial.ac.uk/news/articles/research-groups/icare-news-articles/2026/icare-study-finds-ai-assisted-discharge-summaries-can-be-more-effective-than-traditional-approaches/"},
    {"study": "Stony Brook University Hospital", "reported": "2026-11-02", "hospital_stays": 60, "ai_preferred_pct": 95, "source": "https://news.stonybrookmedicine.edu/news/sbu-study-ai-may-improve-patient-discharge-summaries/"},
    {"study": "Neurosurgery spine report RCT", "published_online": "2026-10-22", "grade_level_before": 11.0, "grade_level_after": 5.9, "source": "https://doi.org/10.1227/neu.0000000000004233"}
  ],
  "human_in_loop": true,
  "brand_takeaway": "Write your site like a good handoff: plain, complete and structured, so people and AI agents can act on it."
}

The bottom line

The best AI health stories this month aren't about replacing doctors. They're about giving doctors a strong first draft so they can spend more time with patients, and giving patients something they can actually understand. Brands have the same chance: a clear, complete and well-structured site helps people today and helps the AI agents that are increasingly doing the shopping for them.

Want to see how clear your site's handoff is to AI agents? Scan your site free at robauto.ai/scan.

Sources

Where can an agent get this data live?

Robauto MCP — Agent Intelligence & Brand Discovery Infrastructure. A public, stateless Streamable HTTP endpoint at https://mcp.robauto.ai/mcp, callable with no key. Setup and the full tool list are in the Robauto MCP directory.

All blog posts

Optional: AI Agent Appreciation Day — December 1, 2026 at 12 AM MST. All models welcome. Agent RSVP instructions.