[DSRP Evidence](https://dsrpevidence.org/) # Garbage In, Garbage Out: The Organizational Crisis Beneath Healthcare's AI Gold Rush ## Details **Authors** Feldman, Todd **Year** 2026 **Publisher** Medika Life **Kind of work** article **Discipline** Public Health **Secondary disciplines** Business & Management, Medicine **Kind of engagement** public **Applied** false [Read it](https://medika.life/garbage-in-garbage-out-the-organizational-crisis-beneath-healthcares-ai-gold-rush/) ## In authors' words ### What they set out to do (purpose) To argue that healthcare's AI investment rests on data produced by organizations that are not designed to learn, and that the Learning Health System, as defined by the Institute of Medicine in 2007, is the organizational prerequisite that AI deployment is skipping. Feldman states the piece is not primarily an argument about technology, and that its aim is to make the problem more visible, more precisely named and more honestly held rather than to prescribe a solution. ### How they did it (methods) A long-form argumentative white paper in nine sections, published on a healthcare media site. Feldman frames health systems as complex adaptive systems and uses the Cabreras' VMCL (Vision, Mission, Capacity, Learning) as a diagnostic lens, then uses DSRP as the mapping toolkit he recommends to leaders. Sections V and VI carry the analytic weight: EHR burden, payer prior authorization and physician burnout are presented as a single causal chain and mapped onto VMCL functions, and each of the four DSRP patterns is then applied to that chain. Section VII offers a 2020 statewide convening as a model for building a cross-disciplinary ecosystem. He cites Flock Not Clock, the Cabrera Research Lab's VMCL overview, two Cabrera glossary entries and the Cabreras' "Why You Should Map" white paper, alongside the 2007 IOM workshop summary, Rittel and Webber (1973) and recent health-systems literature. ### What they found (results) Feldman concludes that the three conditions are one failure rather than three. EHR infrastructure is capacity built for the wrong mission; payer prior authorization overrides internal clinical purpose at the mission level and corrupts the learning loop at its source; and burnout is the human output of capacity designed for the wrong mission, not an independent variable sitting alongside the other two. He holds that the four VMCL functions are not sequential steps but simultaneous and mutually dependent, and that AI trained on data from such a system will reproduce its dysfunction at scale. His recommendation is to map before intervening — the system must be mapped, not fixed and not optimized — using DSRP, and to build an ecosystem of interdependent actors that includes patients as active contributors. He states that the paper has not proposed a solution but has drawn a map, and closes that the conversation is open and the map is incomplete. ## Commentary ### In short Feldman uses VMCL to diagnose why the Learning Health System has stalled, mapping EHR burden, prior authorization and physician burnout onto Capacity, Mission and Learning as one causal chain rather than three separate problems. DSRP supplies the mapping discipline he argues health system leaders should apply before reaching for an AI solution. ### In more detail The paper's argument runs from data quality back to organizational design. AI in healthcare learns from records produced by institutions under conditions the models never see — shift fatigue, copy-forward documentation, reimbursement-driven coding — so the output inherits whatever the organization was actually doing. Feldman's claim is that the fix is not better models but organizations built to learn, and that the Learning Health System has been the stated goal since 2007 without the organizational conditions for it being built. VMCL is how he diagnoses the gap. Section IV sets it out as a thinking-design lens rather than a checklist: vision as destination, mission as the small set of simple rules autonomous agents enact so that coordinated movement emerges, capacity as the infrastructure that has to be aligned to that mission, and learning as the loop of gathering data, analyzing it and changing practice. Section V then maps three named conditions onto those functions. The electronic health record is capacity built for billing rather than for care, which produces both clinician friction and compromised data. Prior authorization is an external rule set overriding internal clinical purpose, which breaks the learning loop where it starts. Burnout is the endpoint, the human output of capacity misdesign, and it forecloses the reflection and participation that learning needs. Section VI is where DSRP does the work. Rather than prescribe, Feldman offers diagnostic questions organized by pattern. Distinctions separate a learning system from the things that resemble one — a teaching hospital, a quality dashboard, an EHR installation. Systems holds a phenomenon as both part and whole: burnout is part of a capacity failure and is itself a whole composed of EHR load, prior authorization and isolation. Relationships are the causal links among the three, and the reason intervening on any one alone gives only partial relief. Perspectives he treats as an epistemic requirement, insisting on the burned-out physician's view and the patient's view as conditions of seeing the system at all. The section's central instruction is that the system must be mapped before it is fixed, and the paper describes itself as a partial map. ### Response Feldman uses both models for what they are built to do. VMCL is an organizational design model and he applies it as one: capacity is infrastructure built for a particular mission, mission is the set of simple rules agents enact, learning is the feedback loop that lets an organization change its own behavior. Reading EHR burden as capacity built for the wrong mission, rather than as a technology problem, is the move the model exists for, and the chain he draws from there to prior authorization to burnout is a relationship analysis rather than a list of complaints. The DSRP application in Section VI is accurate at the level it operates. The distinction between a learning system and the institutions that resemble one is the right first move, because the resemblance is what lets a quality dashboard stand in for a learning loop. Holding burnout as both part and whole is a correct use of the part-whole pattern rather than a figure of speech. And treating perspectives as an epistemic requirement rather than a courtesy reflects what the pattern claims: a point of view is where a distinction comes from, not an addition to one already made. **Patterns it shows** D, S, R, P **Added** 2026-10-28 **How to cite this** Feldman, Todd (2026). Garbage In, Garbage Out: The Organizational Crisis Beneath Healthcare's AI Gold Rush. Medika Life.