Cases from the MBA and the Vanguard · MBA module — who owns the standard of care when triage is automated
Subtitle: Balancing AI speed with clinical accountability in high-stakes emergency environments.
The emergency department at Jane General Hospital System was operating at 140% capacity when the executive committee deployed Aegis-Triage, an advanced AI decision-support agent designed to prioritize incoming critical patients. For six months, the system operated as an advisor, analyzing vital signs, intake notes, and historical data to suggest priority scoring to senior triage nurses. The metrics were undeniable: door-to-treatment times for severe cardiac and stroke patients dropped by 22%, and diagnostic oversight errors decreased across all shifts.
Encouraged by these results, executive leadership pushed for a broader deployment: shifting Aegis-Triage from an advisory tool to an automated gating mechanism. Under the new proposal, the agent would directly route patients, order baseline diagnostic panels, and flag immediate interventions before a physician or triage nurse performed a full physical evaluation. The goal was to bypass peak-hour administrative bottlenecks and standardize care across varied shift experience levels.
However, resistance emerged during a night shift in late August. A 52-year-old patient presented with non-specific fatigue and mild upper abdominal discomfort. Aegis-Triage flagged the case as low-priority gastrointestinal distress based on demographic parameters and vital stability. The attending physician, relying on subtle physical cues; an unusual cool clamminess and a brief, unprompted hesitancy in the patient's speech, overrode the system's assessment and ordered an immediate ECG. The patient was experiencing an atypical acute myocardial infarction.
Had the physician followed the automated route, intervention would have been delayed by at least two hours.
The incident ignited an immediate rift between executive leadership and the clinical staff. The Chief Medical Officer argues that statistical evidence favors automation: human overrides occur in 12% of cases, yet post-audit data shows that 80% of those overrides lead to redundant testing, increased patient wait times, and higher operational cost without improving clinical outcomes. In her view, trusting individual human intuition over aggregated, evidence-based algorithmic patterns leads to system-wide inefficiency and higher overall liability.
Conversely, the Chief of Emergency Medicine and the nursing union maintain that clinical judgment cannot be quantified into baseline parameters. They argue that automating entry decisions turns clinicians into rubber-stamp ratifiers, eroding the critical observational skills required under pressure. If a physician or nurse must spend their shift monitoring and second-guessing an automated queue, accountability becomes ambiguous: who holds legal and moral responsibility when an algorithm misclassifies a silent symptom?
The board faces an immediate decision: fully implement autonomous triage routing to resolve capacity strains, restrict the system strictly to passive advisory status, or institute a mandatory human-in-the-loop sign-off for every AI classification. The deployment deadline is in seven days, and capacity demands continue to escalate.
I would open the hour by asking a senior healthcare executive or legal expert in the room: "If your institution's data proves that human overrides decrease overall accuracy by 8%, do you revoke the clinician's authority to override the system?"
To pivot the debate, the critical fact introduced mid-session would be: Insurance underwriters have signaled that liability premiums will drop by 15% if autonomous routing is enforced without manual clinician approval for low-risk categories. This forces the room to evaluate whether financial and operational incentives outweigh professional discretion under uncertainty.
In the NEO era, Aegis-Triage no longer functions as a static algorithm awaiting periodic updates. It acts as an autonomous agent, continuously recalibrating its own diagnostic thresholds based on real-time room capacity, historical patient outcomes, and live telemetry data across multiple regional hospitals. It actively re-prioritizes queue lines, orders lab tests, and reallocates nursing staff without explicit prompt or intervention from the shift supervisor.
This shift transforms the clinical ecosystem from human-led, tool-assisted care to autonomous system governance. The core dilemma shifts from how clinicians use AI to how humans govern an autonomous system that operates faster than human consensus can form. Clinicians are no longer making primary decisions; they are auditing stream logs after actions have already been initiated.
Technology can optimize the system's capacity, but only human judgment carries the burden of the consequence.
A live case: every round can be improved, and the author's feedback is the next one.
Without the mentor's corpus
1. Never as a right; sometimes as a habit. The 12% override rate hides the only number that matters: of the overrides that were right, how many caught a time-critical diagnosis like the atypical heart attack? Overrides happen on the hardest cases, so counting them against averages always makes them look expensive. Weigh them by outcome, not by count. A redundant test costs money and minutes; one missed infarction costs a life. In Europe part of the answer is already law. Emergency triage systems are high-risk under the AI Act, whose rules for them apply from December 2027. It requires that the people overseeing such a system can override it and are made aware of automation bias. So keep the override and make its cost visible: one line of reason per override, a monthly review of all of them, and every confirmed miss fed back into the model. Ask yourself: what is one missed heart attack worth in my override statistics?
2. Legally, the answer will be shared. The manufacturer answers for a defective system, the hospital for how it chose to deploy it, and the clinician whose name is on the chart. Morally, the danger is what Madeleine Clare Elish calls the moral crumple zone: when an automated system fails, the blame lands on the nearest human, even one who had neither the time nor the authority to catch the error. Suppose the board routes patients automatically to save time, and leaves the night nurse answerable for whatever the queue misses. It has moved the risk down the chain and kept the savings at the top. Responsibility and authority must sit together. For the categories the board automates, a named executive owns and signs the routing rules. For everything else, clinicians decide, with the time and staff to do it. Ask yourself: if the night nurse is responsible, does she have the authority and the minutes to act on it?
From the mentor's corpus
1. The mentor's HAI5 framework turns the board's three options into levels. Six months of advice was Level 3: the human decides after the AI recommends. Autonomous routing is Level 4: the AI executes, and humans take the exceptions. The mentor's task-to-agent protocol then asks which parts of triage belong where. Scoring vital signs and ordering standard panels is calculation, safe at Levels 2 to 4. Reading a cool, clammy patient who hesitates mid-sentence is craft, which the protocol never automates. So do not set one level for the whole department. Split the task. The Vanguard Leadership Handbook adds that an irreversible, high-impact decision deserves time and documented logic, and a seven-day deadline tries to make it fast. The way through is to make it reversible: autonomy for one narrow low-risk category, heavily instrumented, with a kill indicator set in advance. One missed time-critical diagnosis, and it stops. Ask yourself: which part of triage is calculation, and which part is craft?
2. The second volume's Command Layer chapter answers this in one paragraph. Delegation does not dissolve the chain of responsibility. The subordinate answers for the action. The commander answers for the intent, the boundary conditions and the authorisation, and both are on the record. A leader who uses delegation to push consequences to the edge while keeping the credit at the centre has, the chapter says, misunderstood the doctrine at its root. Here the board is the commander, and Aegis-Triage is its subordinate. If clinicians defer because the board designed the shift that way, the board carries the authorisation, and it must be written, signed and visible. The same chapter asks for boundary conditions the agent cannot cross without flagging itself for human review. Only then can anyone see who decided what. Ask yourself: if a patient is harmed tonight, can we show who authorised the rule that routed him?
On the NEO Turn. Her NEO Turn describes an agent that recalibrates its own thresholds faster than consensus can form. The Command Layer chapter sets the conditions for that. The autonomy level is declared, the authorisation is recorded, and the agent's behaviour within it is measured, so that its autonomy can be adjusted. Autonomy is then earned and withdrawn on evidence, not granted by a deadline. Clinicians who only audit stream logs after the fact are not governing; they are reading. Governing means the agent works inside limits it cannot cross without flagging itself for human review.
The mentor has worked on this exact problem twice. The first time was a Horizon Europe proposal for governed clinical AI agents, Vanguard AI (2026). The second was the Ambition pilots, an evidence network for data spaces in which hospitals, telecoms and cloud operators keep separate books. Both answer the board with a design, not a percentage.
1. Governed autonomy instead of three options. Vanguard AI was designed as decision support with governed autonomy. Specialised agents propose a next step. A decision orchestrator writes a structured report — the rationale, the evidence used, the alternatives considered and the confidence. A clinician edits the report and approves it before any clinical action, and the clinician's edit takes absolute precedence. The proposal's thesis fits Aegis-Triage exactly: the bottleneck is not prediction but coordination — what happens next, in what order, through which actor, within which boundaries. So keep the agent fast where speed is safe: pre-sorting the queue and ordering the standard ECG for a chest-pain protocol. Require a sign-off where the agent's confidence is low or the presentation is atypical. Let every classification arrive with its reasons, so the nurse reviews a case, not a score. That is neither the board's full autonomy nor its sign-off on everything.
2. Validate it like a clinical intervention. The proposal staged its path in steps: retrospective validation first, then supervised live use that drives no clinical action, then a prospective study. Before any patient-facing use, an ethics and safeguards board had to approve the step. It also planned to audit the agents' performance separately for subgroups, starting with gender, because averages hide exactly the atypical presentations. A 52-year-old with fatigue and mild abdominal discomfort is the textbook case where averages lie. A seven-day deadline skips every gate. In seven days the board can still move to a supervised mode, in which the agent routes nothing it cannot explain.
3. Make responsibility provable. The liability question in this case is an evidence question. After an incident, who can show what the agent recommended, under which version of its thresholds, what the nurse saw, who overrode it, and why? In the Ambition pilots (1, 2, 3) each organisation keeps its own book and publishes one 32-byte fingerprint per period. Any single record of that period can later be proven. A handover between two organisations can be proven without either of them opening its book. The third pilot rebuilt one shared case from ten contributors' disclosures. For Aegis-Triage that means three books:
When a case goes wrong, it is rebuilt from the three books without a shared database. The Vanguard AI audit trail called this the third entry: a co-signed record that no party can rewrite alone. The record proves what was written and when. Whether the decision was right is still for people to judge, now on facts rather than on memory.
4. Her NEO Turn, answered. An agent that recalibrates its own thresholds must record each recalibration in its own book. The threshold in force at 02:14 on the night of the incident is then a fact, not a reconstruction. Its autonomy can grow as the record shows it has been earned.
Ask yourself: if the board adopts autonomous routing next week, which record would prove, a year from now, who decided what for the patient in bed seven?
A question for the table, a disagreement, what you would have done. The case lead reads every comment; the ones the table takes up enter the chapter as questions from the room, with your name.