A patient comes to you with a story you’ve heard in countless variations. She’s seen a nurse practitioner at an urgent care, a rheumatologist two states away when she was between jobs, and, more recently, an AI symptom checker that told her it was probably stress. By the time she’s in the exam room, you have maybe fifteen minutes, and a chart that reflects exactly one of those encounters.
During several decades in practice, I’ve watched the conversation around AI in health care shift from “Will this replace me?” to something more grounded: Patients are increasingly consulting an AI model before they call the doctor’s office. A physician podcast recently captured it well: Doctors are becoming the second opinion, with AI serving as the first. That framing is accurate, but it misses the part that actually matters.
During those same decades, documentation of clinical records has evolved significantly, from barely legible handwritten scribbles on paper charts, kept by individual providers and rarely shared, rife with opportunities for errors, to a current state where information is digitally collected and maintained within walled, proprietary ecosystems. Through all those years, until this very moment, one thing has not changed; it doesn’t matter whether the first opinion comes from a chatbot or the second opinion comes from a board-certified physician. Both are only as good as the record they’re reasoning over. And in 2026, that record is still, overwhelmingly, incomplete.
The chart you have is not the chart that exists
Every clinician knows this intuitively. The “complete” record in front of you is simply whatever happened to make it into your EHR. It’s not the walk-in clinic across town. It’s not the specialist visit from the insurance plan the patient had two jobs ago. It’s not the medication list from the pharmacy they switched away from.
Real physician voices, twice a week
Free, and one click to unsubscribe.
Most systems that claim to unify records still rely on the patient logging into a separate portal for every prior provider, remembering a password they set once, and manually authorizing each release. Most patients don’t do that, and frankly, we shouldn’t expect them to.
The result is a chart that looks complete because it’s formatted like one, but functions like a fraction of the story. Modern EHR interfaces reinforce the illusion: tidy medication lists, clean problem summaries, and neatly organized timelines that imply completeness even when half the patient’s medical life is missing.
AI systems are now being asked to reason over that same partial picture.
A composite case we’ve all seen
Here’s a generalized example, not one patient, but a pattern many clinicians have encountered. A patient on a stable anticoagulant regimen arrives in an ED with new confusion. The ED has no record of the anticoagulant. It was prescribed by a physician outside their system, at a visit that predates the patient’s current insurance. A well-meaning resident, working from what’s visible in the chart, orders a workup that doesn’t account for the medication at all.
Nothing catastrophic happens this time. But the near-miss is real, and it wasn’t caused by poor clinical judgment. It was caused by an incomplete record that looked, on screen, indistinguishable from a complete one.
That’s the critical point: An incomplete chart doesn’t announce itself. There’s no banner that says, “Forty percent of this patient’s history is missing.” It just looks like a chart. And a physician (or increasingly, an AI system doing first-pass triage) has no way to distinguish “This problem doesn’t exist” from “This problem exists but isn’t visible to me.”
AI adds another layer of risk: When it fills gaps with statistical guesses, those guesses can appear authoritative. A human clinician can sense when something is missing. An AI model cannot.
This is becoming a liability question, not just a quality one
We don’t discuss this enough, perhaps because it’s uncomfortable. But a decision made on an incomplete record, where a fuller record was reasonably obtainable, is a different kind of exposure than a decision made on genuinely unavailable information.
As more of the “first pass” shifts to AI (symptom checkers, triage bots, documentation assistants), the standard for what counts as a reasonably complete record before a clinical judgment gets made will be scrutinized. Quality committees will examine it first. Plaintiffs’ attorneys will examine it next.
There’s also an equity dimension: Patients who move frequently, change insurance often, or rely on urgent care instead of primary care are disproportionately likely to have fragmented records. Incomplete data doesn’t just create risk; it creates unequal risk.
This is not alarmism. It’s a systems problem we should address deliberately now, rather than relearn case by case.
What actually needs to change
The solution is not asking patients to be better record-keepers, nor asking physicians to spend more of their fifteen minutes chasing down old charts by fax. Patients have never been good at being their own medical records custodians, and it’s not a fair expectation to place on someone who is sick. The fix must happen at the infrastructure level.
We need retrieval systems that:
- Do not depend on a patient logging into every portal they’ve ever touched
- Work through the provider relationship rather than around it
- Surface gaps instead of quietly presenting a partial record as a whole one
We already have the technical building blocks to do this. What’s missing is the institutional will to treat “Does this AI or physician have the full picture?” as a patient-safety question worth solving, not a background inconvenience we’ve learned to tolerate. Until record completeness is treated as a core requirement, AI will continue to reason over partial stories, and clinicians will continue to inherit the consequences.
The question that actually matters
The next time an AI tool hands a patient a first opinion, or a physician forms a second one, the question worth asking isn’t simply, “Was the reasoning sound?” It’s “What fraction of this person’s actual medical life was that reasoning based on?”
Until we can answer that with confidence, every opinion in the chain, human or otherwise, is doing its best with less than the full story.
Miguel Venereo is a board-certified obstetrician-gynecologist and senior health care executive serving as chief medical officer at Medikarma, where he leads clinical strategy and data-integrity initiatives aimed at improving the reliability, completeness, and clinical usability of medical records across diverse care settings. His career spans hospital systems, Medicaid plans, and Medicare Advantage organizations.
Venereo has held leadership roles within Community Care Plan, Centene, and UnitedHealthcare, where he directed programs in utilization management, care coordination, predictive modeling, and quality-measure performance. His work in both Medicaid and Medicare Advantage environments has centered on strengthening documentation accuracy, reducing avoidable variation, and improving outcomes through better information flow and record completeness.
Across his career, including major administrative roles at Memorial Healthcare System, Venereo has led multidisciplinary initiatives that modernized peer review, enhanced patient-safety practices, and advanced system-level performance improvement. His ongoing work focuses on leveraging reliable clinical data and interoperable records to support equitable care delivery, reduce readmissions, and improve decision-making for clinicians, health plans, and patients. Medikarma’s About and For Organizations pages describe this work in more detail.
