When Reconstruction Breaks Down

A risk score resulted in an early discharge. The model achieved the self reported accuracy of the data that trained it. It passed every review. Every documented step of the workflow that produced the score was followed. The physician reviewed the documented process, scored the risk as consistent with everything in front of them, and signed off. After discharge, the outcome was incorrect.

Each group investigation of the event has clear explanations for the part they owned. The data team outlines sound inputs. Vendor model review outlines operational efficacy. Workflow ran as intended for the operations team. The clinician exercised sound judgment to the extent of their knowledge at the time. Every explanation is defensible. The final outcome contradicts the sum of them. Investigation reaches an endpoint before discerning the true source. That source was outside the evidence. It existed in the gap where no single explanation conflicts with another and where no one was tasked to search.

This already occurs in health systems where AI is shaping discharge decision making, resource distribution, risk stratification, and clinical documentation at scale. The output enters a workflow that multiple teams touch. Each team interprets it through a different frame and acts on that interpretation without having ever coordinated on what the output was supposed to mean once it left the hands of the builders. Every team made a defensible decision in their role. The model did not make the decision, but it designed the framework in which the decision was made, and the conditions of that framework were spread across roles that never had a convergent dialogue about what they were producing in unison or how the output was going to land once it entered a clinical context.

Review of adherence determines the scope and confirms what is obligated. Clinical quality review determines the care decision and confirms it was within standard. IT confirms the system operated as designed. Analytics confirms the logic was sound. All four confirmations are correct, and all four are incomplete in the same way. None of them accounts for what occurred in the space where the four functions intersected to produce the singular output that the physician acted upon. That space has no owner, and the investigation framework has no place for it. What gets recorded in the post incident report is a claim of what each function did adequately. Interrelation between them goes undocumented, because the report was structured to capture individual responsibility, not collective causation, and no one has been asked to create a report that accomplishes both.

The meetings where this plays out do not resolve. One stakeholder recounts their piece, then the next recounts theirs, both narratives are correct, and the gap between them is where the outcome was produced. That meeting was never designed to address the gap because no one put it on the agenda, because no one owns it. It ends with a shared understanding that nothing failed, and every condition that produced the outcome remains in place for the next decision.

Health systems are structured by defining roles where quality is separate from IT, analytics, infrastructure, and logic, each function acting as the single point of responsibility for what it owns. This model works when decisions can be traced back to a particular function that produced them. AI produces situations that shape decisions across several functions at once, and the operating model that health systems run on is doing what it was designed to do. It is simply not designed for a system where the point of consequence is the interaction between functions rather than the output of any single one of them.

The cost of this gap is not what leaders expect, because it does not arrive as a single large event that triggers a response. Time is the cost. Weeks lost in review cycles that do not resolve. Meetings where each stakeholder leaves with a clear understanding of their own deliverable and no understanding of how the deliverables combine into a decision. Decisions that remain unmade because reporting lines run vertical and were never designed to share authority over the same output. The organization does not stop, but it slows, and the slow is invisible in the very areas where AI was expected to bring speed and clarity. When boards start to notice, the question they ask is whether AI is being used in the organization ethically, to which every function answers yes, and every function is telling the truth. No one asks whether there is oversight over the collaboration between functions, because there is no report that surfaces the issue and no role that exists to bring it to the table.

Vendor contracts reflect the same pattern. The vendor secures the model touch points. Health systems secure the deployment touch points. In the space where the output moves between them and takes on unassigned meaning, both contracts leave influence to assumption. That space is where the next incident is already being built.

The question leaders keep asking is who is answerable for this output. That question leads back to the same clean lines from every function and the same gap that the question does not design the investigation to name. A different question would change the reality of the output. Who has named ownership over the design of the output. Ownership of that kind would be constituted at the intersection of data quality, model behavior, workflow design, and clinical decision, and it would exist before the outcome, not after it.

Such a role is not present in health systems today. There is no standing audit of the interplay between functions, and no one is assigned ownership of what lives at the point where those functions converge. What resides in that absence is not a gap waiting to be filled. It is a gap that already defines what is produced, ahead of the awareness of the organization, shaping decisions and outcomes in systems that are operating exactly as they were designed.

The article draws on recurring patterns observed in incident reviews across health systems deploying AI into discharge, risk stratification, and clinical documentation workflows. It extends themes from earlier editions on the oversight debt curve, the audit trail illusion, and why oversight without ownership fails. The pattern this edition names, the absence of a named owner for cross functional output design, is the structural condition that produces the outcomes those earlier editions describe.

Christopher Hutchins
Founder & CEO, Hutchins Data Strategy Consultants

Related from Hutchins Data Strategy Consultants: Healthcare Data Analytics Consulting. On The Signal Room podcast: Data Quality and AI Strategy.