It is a Tuesday at a PE-owned imaging center outside Atlanta. Two events happen within the same hour.
At 10:07 AM, a patient cancels a 2:00 PM MRI appointment. The scheduling concierge checks the waitlist, identifies a patient whose lumbar spine MRI was scheduled for Thursday, confirms insurance eligibility, verifies the MRI protocol matches the available slot, and fills the cancellation. The center manager receives a notification: “2:00 PM MRI slot filled from waitlist. Patient: confirmed. Auth: verified. Protocol: compatible.” She does not need to approve. The action is routine, within established parameters, and the scheduling concierge has demonstrated 96% accuracy on waitlist fills over the past 90 days. This is Tier 1.
At 10:43 AM, the portfolio intelligence agent surfaces a recommendation. Three imaging centers in the portfolio, including this one, are running MRI utilization below 70% on Wednesday afternoons. A fourth center, 45 miles southeast, is running at 94% utilization and turning away referrals. The recommendation: reallocate the 3:00 PM MRI technologist from this center to the high-volume center on Wednesdays. The projected utilization improvement across the two centers is $18,000 per month in additional revenue.
The center manager reads this recommendation and sees what the portfolio intelligence agent cannot fully weigh. The MRI tech has a 50-minute commute to this center and would have a 90-minute commute to the southeastern center. The tech’s daughter attends a school near this center, and the Wednesday afternoon schedule allows the tech to handle school pickup. The referring physician who sends 30% of this center’s MRI volume has a personal relationship with this tech and has mentioned that the tech’s professionalism is part of why he refers here. None of these factors appear in the utilization data.
The center manager declines the recommendation and notes her reasoning. The portfolio intelligence agent records the decision, the reasoning, and the outcome. It does not override. It does not re-recommend next week. It incorporates the decision into its model of this center’s operational constraints. This is Tier 3.
Tier 1: The Autonomous Layer#
Tier 1 actions run without human intervention. The agent acts; the human is notified after the fact. Four conditions must hold simultaneously.
The action must be within established parameters. The scheduling concierge fills cancellations from the waitlist because waitlist filling is a defined operational action with clear rules: the replacement patient must be eligible, the slot must be protocol-compatible, and the fill must occur before a configurable cutoff time. These parameters are set per entity. One center may allow waitlist fills up to 60 minutes before the appointment. Another may require 120 minutes. The parameters are the envelope within which autonomy operates.
The action must be reversible or low-impact. Filling a cancellation from the waitlist is reversible: if the replacement patient cancels, the slot returns to the waitlist. Submitting a clean claim to a payer is low-impact: the claim either pays or denies, and a denial triggers its own workflow. Reordering standard medical supplies at the par level is routine and self-correcting if the order is slightly off. These are actions where the cost of a mistake is bounded and recoverable.
The action must be high-frequency and well-understood. The system has seen this action type hundreds or thousands of times. The scheduling concierge has filled waitlist slots at this entity for 90 days. The revenue cycle concierge has submitted clean claims for six months. The pattern is established, the outcomes are tracked, and the error rate is below the entity’s configured threshold.
The agent must have demonstrated accuracy above the entity’s threshold. This is the earned autonomy gate. A scheduling concierge that correctly predicts waitlist fill compatibility 96% of the time has earned Tier 1 autonomy for waitlist fills. A revenue cycle concierge with an 88% clean claim acceptance rate has earned Tier 1 autonomy for standard claim submission. The thresholds are configurable. A risk-tolerant entity might set the threshold at 85%. A risk-averse entity might set it at 95%. The architecture does not dictate the threshold; it enforces whatever threshold the entity sets.
The agents that operate most frequently at Tier 1 are the scheduling concierge (waitlist fills, no-show rebooking, routine appointment confirmations), the revenue cycle concierge (clean claim submission, payment posting, standard collections workflows), and the supply chain concierge (par-level reordering, inventory tracking, expiration alerts). These are the high-volume, low-consequence operational actions that consume staff time disproportionate to their complexity.
Tier 2: The Entity Gate#
Tier 2 actions require approval from an entity-level decision-maker before the agent proceeds. The agent recommends; the human decides.
Five triggers move an action from potential Tier 1 to mandatory Tier 2. Financial impact above the entity’s configured threshold: a supply order that exceeds par level by more than 20%, a scheduling change that affects projected revenue by more than a configured dollar amount, a coding change that alters the charge structure. An order for $400 in standard supplies is Tier 1. An order for $4,000 in a specialty supply is Tier 2, because the financial exposure warrants human review.
Actions that affect staff assignments cross into Tier 2. Rescheduling a provider’s patient slots, adjusting staffing ratios, modifying shift assignments. These involve human factors that the agents can assess partially but not fully. The staffing concierge can calculate that reallocating Dr. Chen from afternoon clinic to morning procedures increases utilization by 12%. It cannot calculate that Dr. Chen has a childcare arrangement that makes morning shifts impossible without a two-week adjustment period.
Actions that change clinical workflow require Tier 2 approval because clinical workflow changes affect patient safety, provider comfort, and regulatory compliance in ways that compound. The compliance concierge recommending a documentation template change based on audit findings is a reasonable recommendation. Implementing it without physician review is not.
Actions involving external relationships require entity-level approval. A payer contract renegotiation point identified by the payer contract concierge. A referral relationship concern flagged by the referral concierge. A vendor contract approaching renewal. These actions involve organizational relationships that have histories and nuances the agents cannot fully model.
Actions where the agent has not yet earned autonomy default to Tier 2. A newly deployed scheduling concierge starts every action at Tier 2 until its accuracy record establishes the pattern that earns Tier 1 graduation. The default is human oversight. Autonomy is earned, never assumed.
The approval workflow is designed for operational speed, not bureaucratic process. The agent delivers its recommendation through the decision-maker’s preferred channel: a notification on the dashboard, an alert on the mobile app, a structured message in the practice’s communication platform. The recommendation includes the proposed action, the reasoning, the expected outcome, the risks, and the confidence level. The decision-maker approves, rejects, or modifies. The system records the decision and learns from it. A recommendation that is consistently approved moves toward Tier 1 graduation. A recommendation that is consistently rejected or modified adjusts the agent’s recommendation model.
Tier 3: The Portfolio Gate#
Tier 3 decisions cross entity boundaries or affect portfolio strategy. These are the operating partner’s decisions, surfaced by the portfolio intelligence agent with full context but never taken autonomously.
Cross-entity resource changes: reallocating staff between entities, shifting patient volume from one center to another, consolidating administrative functions across practices. These affect multiple entities and require portfolio-level coordination. The portfolio intelligence agent can model the financial and operational impact. It cannot model the human and relationship consequences without entity-level input.
Portfolio-wide payer strategy: responding to a payer policy change that affects the entire portfolio, negotiating a portfolio-level contract, adjusting coding practices across entities based on denial pattern analysis. These decisions require portfolio-level authority because they commit multiple entities to a unified approach.
Capital allocation decisions: investing in equipment at one entity versus another, funding a build-out, approving a technology upgrade. The portfolio intelligence agent can rank investments by projected ROI. The operating partner weighs ROI against strategic positioning, relationship factors, and operational capacity.
M&A intelligence: the portfolio intelligence agent identifies acquisition targets based on operational patterns (BOI-06.03). The agent can assess operational fit. The operating partner evaluates strategic fit, cultural fit, and deal terms.
The critical constraint on Tier 3 is that portfolio decisions affecting specific entities require entity notification and, at higher trust tiers, entity consent. The operating partner cannot reallocate a technologist from Center A without Center A’s manager knowing. At trust Tier 3 and above, the manager’s consent is required, not merely her awareness. The escalation hierarchy respects entity sovereignty even when the portfolio has the authority to override. The architecture makes the override visible and recorded, not silent.
Earning Autonomy#
Agents move actions from Tier 2 to Tier 1 through demonstrated performance. The autonomy scoring model tracks three dimensions.
Accuracy measures whether the agent’s autonomous actions produce correct outcomes. The scheduling concierge’s waitlist fill accuracy is the percentage of fills that result in completed appointments without scheduling conflicts, insurance denials, or patient complaints. Accuracy is measured against outcomes, not against predictions. A fill that the agent predicted would succeed but that the patient later cancels is not an accuracy failure; a fill that created a scheduling conflict is.
Consistency measures whether the agent’s performance is stable over time. An agent that performs at 95% accuracy for three months and then drops to 80% in month four has a consistency problem. The autonomy model weights recent performance more heavily than historical performance, because recent performance reflects current system conditions. A billing system update that changes denial patterns should trigger a temporary Tier 2 reversion for the revenue cycle concierge until the agent demonstrates that it has adapted to the new patterns.
Consequence severity weights the impact of errors. A scheduling error that wastes a 30-minute slot has low consequence severity. A coding error that triggers a payer audit has high consequence severity. Actions with high consequence severity require higher accuracy and consistency thresholds for Tier 1 graduation. The thresholds are asymmetric by design: low-consequence actions graduate at 85% to 90% accuracy, while high-consequence actions may require 95% or higher.
The thresholds are configurable per entity because risk tolerance varies. An aggressive growth-oriented practice may set lower thresholds to maximize operational speed. A conservative practice with a recent compliance issue may set higher thresholds. A practice in a highly regulated state may set thresholds that reflect the regulatory environment’s expectations. The architecture provides the scoring model; the entity provides the policy.
Demotion is also possible. An agent that has earned Tier 1 autonomy for a specific action type can be demoted back to Tier 2 if its accuracy drops below the threshold, if the underlying system changes in ways that invalidate the historical performance data, or if the entity’s governance policy changes to require more oversight. Demotion is automatic when triggered by performance metrics and manual when triggered by governance changes. Either way, it is immediate. The system does not continue operating at a trust level it has not earned.
Transparency as Architecture#
Every escalation is logged with full reasoning. Every Tier 1 action is logged with the parameters that authorized it. Every Tier 2 recommendation includes the reasoning chain that produced it. Every Tier 3 recommendation includes the data, the analysis, the projected outcomes, and the confidence intervals.
This is not optional transparency added for compliance. It is an architectural requirement for the earned autonomy model to function. The entity cannot evaluate whether an agent has earned Tier 1 autonomy without a complete record of its decisions and outcomes. The operating partner cannot evaluate a Tier 3 recommendation without understanding the analysis that produced it. The auditor examining the system’s behavior six months later cannot assess compliance without the decision log.
The audit trail connects to the audit architecture described in BOI-05.02. Every escalation decision, every tier assessment, every autonomy graduation and demotion is a permanent, immutable record. The trail serves three audiences simultaneously: the entity manager reviewing current operations, the operating partner evaluating portfolio governance, and the external auditor verifying compliance. Same data, different views, same underlying record.
The transparency also serves the agents themselves. The learning loop described in BOI-02.01 uses escalation outcomes as training signals. When an entity manager consistently overrides a Tier 2 recommendation in a specific way, the override pattern becomes a learning input. The agent adjusts its recommendations. The escalation trail is not just a record of what happened; it is the mechanism by which the system improves.
Cross-References
BMT-04.01 The Human Agency Scale: the consumer autonomy framework that establishes the principle of scored human agency, adapted here for operational decision authority across three tiers.
BMT-04.02 Earned Autonomy: the earned autonomy model describing how consumer-facing agents gain trust through demonstrated accuracy, applied here to operational agents and entity-level governance.
BMT-04.04 The Escalation Hierarchy: the consumer escalation framework that BOI extends from individual person decisions to entity-level and portfolio-level operational decisions.
BOI-02.01 The Operational Brain: the DAG execution engine that operates within the escalation tiers, pausing workflows at tier boundaries when human decisions are required.
BOI-05.01 Trust Tiers: the portfolio trust tier framework that intersects with escalation authority, governing what cross-entity actions are possible at each trust level.
BOI-05.02 The Audit Trail: the audit architecture that records every escalation decision, tier assessment, and autonomy change as permanent, immutable evidence.
Technical Appendix BOI-02.04-A is available to partners and investors at partners.bluemirror.tech.