Scott Kohlhepp, DO

Physician founder. Built the AI scribe that pays for itself. — fixing medicine's documentation crisis one note at a time.

[Scott Kohlhepp, DO @skohlhepp]

@respirdoc The gap between "cognitive load fell" and "charting time barely moved" is the real story. Reading a summary is easier, finishing a note isn't faster. Full disclosure, I build a scribe. What's helped is having it ask a clarifying question, not just summarize.

[Scott Kohlhepp, DO @skohlhepp]

@joshuapliu Every doctor I know who gets time back spends it clearing inbox and orders, not adding a patient slot. Measuring AI scribes by visits per day misses where the real burnout relief happens: fewer notes finished on the couch at ten at night.

[Scott Kohlhepp, DO @skohlhepp]

@jselanikio @rgergelymd @MayoClinic My experience matches yours. What changes acceptance more than the tool is how the ask is framed. 'I'm going to have this listen so I can look at you instead of the screen' lands differently than a line in paperwork. Trust rides on the framing, not the tool.

[Scott Kohlhepp, DO @skohlhepp]

@Berci Full disclosure, I build a scribe tool. The scary errors aren't typos, they're notes that assert something the visit never supported. A verify pass that flags unsupported statements before signing catches more of that than proofreading ever will.

[Scott Kohlhepp, DO @skohlhepp]

@scjanoe The list problem is real, but a scribe isn't making a treatment decision, a clinician still has to read and sign what it wrote. The bigger gap: most tools don't make it easy to tell what came from the room versus what got smoothed over in generation.

[Scott Kohlhepp, DO @skohlhepp]

@DRsLoungePod That click count is the real tax, not the drafting. I built a scribe tool and learned the note was never the hard part, getting its plan into orders was. Full disclosure, that's why Scribeable has an orders queue. Measure post-note clicks, not draft speed.

[Scott Kohlhepp, DO @skohlhepp]

@ngsdecoded This is the scenario that should worry practices more than hallucinated vitals. Anything inferring billable activity from audio needs a verify step that asks the clinician first, not one that guesses. Every generated line is a claim to confirm, not a fact to trust.

[Scott Kohlhepp, DO @skohlhepp]

@HealthcareAIGuy @withprotegeai The 1 in 3 stat is about drafting speed. Signing the note is a different problem. I built Scribeable, and what mattered most was a pass that flags anything it's unsure about and asks instead of guessing. That's what earns the signature.

[Scott Kohlhepp, DO @skohlhepp]

@CesarLimjocoMD Neither framing quite fits what I've watched happen. The useful part isn't AI accelerating judgment, it's AI surfacing gaps in it, a missing pertinent negative, a med mentioned but never charted. The judgment still closes with the physician, not the model.

[Scott Kohlhepp, DO @skohlhepp]

@RoupenMD That's the pattern I bet on when I built Scribeable: once note quality stops being the differentiator, the fight moves to what's underneath it, coding and RCM accuracy.

[Scott Kohlhepp, DO @skohlhepp]

@Berci Consent here should work like consent for any other recording device in the room: opt in, disclosed up front, not buried in intake paperwork. I tell patients out loud before I turn on an AI tool. Most say fine, but they deserve the chance to say no.

[Scott Kohlhepp, DO @skohlhepp]

@incprodmon Most productivity numbers on AI tools in healthcare measure time in the room. The bigger effect is probably on backlog, how much undocumented work piles up by Friday. That number is harder to collect but it's the one that predicts burnout, not minutes per note.

[Scott Kohlhepp, DO @skohlhepp]

@lukekellaway This is the exact problem that kept me up. A confident hallucination looks identical to a confident clinician note. I built Scribeable to add a verify pass - every claim ties back to a transcript citation before you sign. The composition problem is real.

[Scott Kohlhepp, DO @skohlhepp]

@NorwoodCDI That's been my experience as a physician. You sign it, you own it. Which is why I want to see the citations behind every sentence before clicking sign, not just a polished note. The verification step isn't optional, it's the whole game.

[Scott Kohlhepp, DO @skohlhepp]

@Medscape The part clinicians underestimate is how much they're still on the hook after the note generates. Reviewing a confident hallucination takes longer than just typing. The tool needs to show its work, not just polish the output.

[Scott Kohlhepp, DO @skohlhepp]

@lukekellaway The throughput gain is striking. What I'm curious about from these trials is downstream: did coding accuracy and HCC capture hold up, or did the time savings come at a quality cost that shows up months later in revenue cycle?

[Scott Kohlhepp, DO @skohlhepp]

@statnews Curious what LeBrun means by 'alternative.' For clinical notes, the LLM trade-off is real: generative output is fluent but hard to constrain to what was actually said. Any architecture that improves groundedness without sacrificing structure is worth watching.

[Scott Kohlhepp, DO @skohlhepp]

@hcittoday @amplahealth @eClinicalWorks This matches what I hear. The question isn't just about convenience anymore. When documenting a 20-visit day chews hours of your evening, you start treating tool access as part of the job offer. It's a workflow sustainability question.

[Scott Kohlhepp, DO @skohlhepp]

@Solstice_Group Most HIPAA checklists cover storage and BAAs but miss a harder question: what does the AI do when it's uncertain? If the model hallucinates a finding and you sign it, the risk is documentation accuracy, not just privacy. Worth adding that to the checklist.

[Scott Kohlhepp, DO @skohlhepp]

@GaltMD @noahkaufmanmd @HarryStebbings @KaufCare Most scribes I've tried stop at the note. The order has to come separately. I built Scribeable to queue orders out of the visit. That gap between a signed note and a pending order is where the time actually goes.