.svg)
July 23, 2026
Subscribe on: RSS | SPOTIFY | APPLE PODCAST | GOOGLE | BREAKER | ANCHOR
On this episode of The Digital Patient, Joshua Liu, MD, Co-founder & CEO of SeamlessMD, and colleague, Alan Sardana, chat with Hasan Ahmad, DO, MBA, MSc, Associate Chief Medical Information Officer at Parkview Health, about "Why Deleting Alerts Is the Safest Thing You'll Do This Year, Three Questions Every AI Tool Must Answer Before Go-Live, Why Clinician Resistance Is Usability Feedback and more..." Click the play button to listen or read the show notes below.
Audio:
Guest(s):
- Hasan Ahmad, DO, MBA, MSc, Associate Chief Medical Information Officer at Parkview Health
- Joshua Liu, MD (@joshuapliu), Co-founder & CEO at SeamlessMD
Episode 234 - Show Notes:
[00:00:07] Episode preview
[00:07:26] The data-driven CDS overhaul at Parkview — rather than simply deleting alerts nobody reads, the team analyzed alert performance and restructured them, retiring the ones that weren't changing clinician behavior. The result was roughly a 25% reduction in interruptive alerts and over 100,000 fewer interruptions per month. Alert fatigue, he argues, "isn't an EHR problem, it's more a governance problem."
[00:08:07] Why removing an alert is so hard — it's an asymmetric risk: an over-firing alert gets nobody in trouble, but turning one off makes everyone imagine the worst-case outcome. Alerts overridden 95% of the time protect no one and just train clinicians to click through, making the important alerts invisible. Compounding it are orphan alerts whose original requesters have left, and a process with "a front door" but "no back door."
[00:09:10] Why every alert needs two owners — Hasan argues each alert should have a technical owner checking that it works and is designed correctly, and a clinical or operational owner accountable for the outcome it's meant to deliver. Without a clear owner, there's often no one empowered to say yes to turning an alert off.
[00:10:35] How often alerts should be reviewed — risk-stratify them. Commit to reviewing everything at least once a year, but review the very high-firing alerts quarterly or semi-annually so they never go a full year unexamined.
[00:11:49] Lead with data to take the emotion out of it — everyone has feelings about their alerts, but Epic dashboards carry the performance of every one. Citing a real case, "we have an alert that fires 94,000 times over a six-month period with a 1% action rate, the conversation's already over."
[00:12:21] The redesigned intake process — new alerts now ship with performance criteria: what KPI should this change, how will we know, and what metric defines failure. Alerts are checked at 30 and 60 days, and if something still isn't working by 90 days, it gets sunset rather than wasting clinicians' time.
[00:13:31] One governance body with explicit retire authority — improving decision support requires a single body with clinicians in the room that owns the full lifecycle. Parkview's CDS clinical support committee does exactly that, and safety was monitored throughout the cuts: "nothing bad happened," which itself signals those alerts weren't doing anything.
[00:14:12] Why failure criteria matter as much as success criteria — without a hard failure definition, a tool "just didn't succeed yet" and can linger for years. Defined upfront, it becomes a contract, and retiring something should carry no shame: "We should celebrate the fact that we looked at something, we reviewed the data, and we made a decision."
[00:15:49] Epic GenAI adoption at Parkview scaled dramatically — organizational token use grew from around 60,000 tokens per month pre-launch to millions per month after, signaling adoption well beyond a small pilot. The challenge: getting AI into everyday clinicians' hands without adding tools they have to remember to use.
[00:16:36] The best AI is embedded in a step clinicians already do — problems with clinical AI usually come from process, not technology, especially when a tool is optional and out of the way. Ignoring an extra, out-of-workflow tool "is not somebody being resistant, it's just being rational."
[00:17:27] Adoption runs on relationships and coaching — Parkview invests heavily in personalization and one-on-one, elbow-to-elbow coaching rather than just a tip sheet, letting enthusiasm spread peer-to-peer within a specialty. Usage is treated as a proxy for trust, "because people aren't gonna keep using things that they don't trust," and removing burden first builds credibility for what comes next.
[00:19:47] Which AI features moved the needle inpatient — as a practicing hospitalist, Hasan points to chart summarization, inpatient insights, and hospital-course summarization as the widest-adopted, biggest time-savers, while the AI text assistant is still gaining uptake. The team makes tools unmissable by surfacing them on the first summary page, the note-writing sidebar, the secure chat sidebar (a lifesaver for cross-coverage), and embedded in default note templates.
[00:24:51] Good AI governance plans for turning tools off, not just on — committees often act as a gate for what gets in, but nobody discusses what gets out. Because AI moves fast, models drift, and vendors change things, "what you say yes to today is not necessarily what's gonna be in your system after a year." It's the same discipline pharmacy already applies to its formulary.
[00:26:37] Three questions on one slide before any AI tool goes live — who owns it, how success is defined as a number, and how failure is defined as a number. After go-live, a contract, a champion, and sunk cost make removal far harder. Monitoring must be consistent rather than complaint-driven, and the obligation to provide monitoring data must be written into the vendor contract, with a named owner to review it.
[00:30:13] One AI capability isn't one workflow — chart summarization needs different governance across settings. Inpatient and ED summaries build off a single encounter and are easy to audit for what's missing, but a complex ambulatory oncology chart with hundreds of notes can produce summaries that "look right" while being subtly, dangerously wrong, demanding a much heavier audit before go-live.
[00:35:24] Clinician "resistance" is really usability feedback — clinicians are telling you a tool causes more friction than they'll accept, doesn't match the real workflow, or forces a productivity dip during the learning curve. Many also remember past tools that promised to save clicks and didn't, so they're rationally weighing whether the change is worth it.
[00:37:19] Ambient AI works ambulatory but struggles on inpatient progress notes — it captures what's said aloud well for ambulatory visits and inpatient H&P and consult notes, but for a progress note "70 to 80% of what goes into my note is coming from outside the note." That's a product limitation, not clinician resistance, and until the vendor ingests that outside context, clinicians rationally copy forward instead.
[00:40:54] Measuring and driving real adoption — vendors like to count a one-time "active user," but Hasan tracks the percent of a clinician's total encounters where the tool is actually used, data now surfaced in Epic dashboards at the specialty and individual level. Closing gaps took investment in post-deployment change management: adding smart links to note templates was the rate-limiting step, and many non-adopters simply needed someone to sit with them and set it up.
[00:46:21] Reframe ROI as competing timelines — finance works in quarters, but clinicians take months to build trust, and burnout or retention gains "might show in the second year." Treat these tools as infrastructure rather than traditional software: track early indicators like after-hours EHR time, note quality, and burnout indices, but agree upfront on when outcomes will actually be judged.
Fast 5 Lightning Round:
- What is your favorite book or book you’ve gifted the most?
Atul Gawande's Checklist Manifesto - If you could instantly master any skill, what would it be?
"Instantly speak any language." - Would you rather have Super strength, super speed, or the ability to read people’s minds?
"Definitely not mind reading. I chair some committees and I wouldn't wanna know a lot of the things that people are thinking... So I'll go with super speed for faster efficiency." - What is something in healthcare you believe others might find insane?
"Most clinical decision support is making care less safe. And the single safest thing that many health systems can do this year is delete alerts." - What is the last movie or TV show you saw, and what did you think of it?
FIFA World Cup
The Digital Patient has been recognized as Feedspot's #1 Patient Engagement Podcast of 2025. Thank you to our listeners for making this happen!
.png)
