Practical Lessons on Paperwork from the ASGE AI Workshop

Artificial Intelligence and the Jevons Paradox in Gastroenterology

A fax sitting in a queue for two weeks. A prior authorization that takes fifteen days. A thirty-minute phone call just to clear a routine colonoscopy. None of that is why any of us went into medicine, and fixing it doesn’t need a breakthrough. That was the thread running through the ASGE AI Workshop, Separating AI Hype from Clinical Reality, held September 18, 2026, at the Gaylord National Resort in Maryland, as part of the AI Weekend, where we heard significant results came from practices that automated the paperwork nobody wanted to handle.

What people reported

Three examples stuck with me.

• One large practice swapped a 30 to 40-minute phone screen for direct-access colonoscopy for a 10 to 15-minute online questionnaire that patients can fill out at any hour, in several languages. The AI sorts risk and hands a nurse a short summary. Labor costs per case dropped, even with extra recall volume.
• Another practice took medication prior authorization from 15 days to 3 and cut the coordinator’s time on it from 8 hours a day to 2. Its fax triage went from more than two weeks to first contact to under a day, and referral conversion rose from under 20% to over 40%.
• That same practice dropped a $50,000-a-year website contract for about $1,500 a year in tools, and an AI-assisted invoice audit turned up roughly $37,000 in billing discrepancies.

These are small, repetitive jobs with clear rules. None was worth custom software before, and now they are. Staff time moved to more useful work, usually without anyone losing a job.

What’s still unproven

I’d keep a healthy skepticism here, and so did most of the speakers.

Nearly all those numbers are self-reported by the practice or the vendor. They’re promising leads that still need independent confirmation.

Demonstrations can look better than reality. One vendor described a scheduling agent that reported a successful booking when the system behind it never saved the appointment. Another speaker said a weekend prototype and a production system look the same in a sales meeting.

How long a project takes has more to do with what it touches than how good it is. At a large academic medical center, a documentation co-pilot built on an already-approved platform, with no new vendor and no EHR write-back, was in daily use within days. A more rigorously validated tool from the same team, which was published, independently benchmarked, and tested on real patient cases, required live system integration and spent three years working through governance review. The lesson is that integration depth, not tool quality, may drive how long approval takes.

Beyond the workflow wins, the GI evidence is thin. The best controlled data we have so far comes from a broader trial across specialties. In a randomized trial of 238 UCLA physicians, one AI scribe cut time in the note by about 9.5% and the other showed no significant change. Clinicians reported occasional clinically significant inaccuracies, and the time measure leaves out editing time, so the savings may be overstated [1].

Finally, a legal briefing showed how far regulation lags technology: consent rules vary by state, liability is largely untested in court, and vendor contracts need a close read. The advice in short: treat AI as a capable but fallible assistant, always reviewed and never the final word.

A practical playbook

Speakers kept landing on the same rule: let AI handle the friction (retrieval, drafting, routing, tracking) and keep a human accountable for every clinical judgment and every exception. Four steps came up repeatedly.

1. Start with a baseline. Measure time per task, delays and denials before you buy anything. Without those numbers, you can’t tell whether the tool helped.
2. Pick one painful workflow. Scheduling, referral triage and prior authorization paid off first. Give it a single owner who answers for the result.
3. Buy the workflow, not the demo. Ask how deeply it integrates with your EHR, where a human reviews, and how you’d get out of the contract.
4. Pilot small and keep watching. Set your stop and scale criteria before you start, check every output at first, and keep auditing after launch. Confirm your malpractice policy has no AI exclusion.

The bigger picture

In the 19th century, the economist William Stanley Jevons noticed that more fuel-efficient steam engines didn’t reduce coal use. They made engines so worthwhile that people burned more coal overall. In a recent NEJM Perspective, Dhruv Khullar suggests AI in medicine could follow the same pattern, with easier care driving more demand [2]. For GI, that could mean a prior authorization that clears in three days, or an endoscopy suite that no longer sits idle after a prep-failure cancellation, becomes more referrals, more screening and more patients who finally get in. That’s a good problem if we plan for it. And as systems improve, small errors cost more, so the human steps that remain matter more.

The goal is to put the time we recover back toward patients, starting with the screening backlog.

Before you automate anything

One question from the workshop has stayed with me: why are we doing this at all, and why is a person doing it? Most of us have spent years getting faster at tasks that probably shouldn’t exist, like re-keying a referral or chasing a fax. If we only automate a broken process, we get a faster version of the same problem.

Another speaker offered a simple test. Would you still buy this if the vendor stopped calling it AI? If you can’t say what it replaces, you may be buying one more layer.

The harder question comes after the friction is gone. When the prior authorization clears in three days, what do we do with the hours we get back? More screening, more time with patients, more teaching? That choice belongs to us. If your unit is piloting something, I’d like to hear what you’re measuring and what you plan to do with the time it frees.

References

1. Lukac PJ, Turner W, Vangala S, Chin AT, Khalili J, Shih YT, Sarkisian C, Cheng EM, Mafi JN. Ambient AI Scribes in Clinical Practice: A Randomized Trial. NEJM AI. 2025 Dec;2(12):10.1056/aioa2501000.
2. Khullar D. Artificial Intelligence and the Future of the Clinical Workforce. N Engl J Med. 2026 Sep 17;395(11):1041-1043.

This article is based on discussions and presentations at the ASGE AI Workshop, Separating AI Hype from Clinical Reality, held September 18, 2026.

Explore AIGI’s programs and resources at AIforGI.org.