James Okafor trusts his own judgment on propulsion systems without hesitation. He spent three decades making decisions where the wrong answer meant a catastrophic failure. He does not trust his own judgment on tax strategy with the same confidence. He is competent. He files correctly. But the tax code changes every year, and the optimization opportunities for retirees drawing from multiple income sources are not intuitive. He wants help with taxes. He wants control over medical decisions. He wants the system to handle his grocery ordering without asking.
Three domains. Three different relationships to delegation. The same person.
The Expert Exchange Layer handles this through a five-level agency spectrum that maps the person’s preferred relationship to AI assistance across every domain and every expert type. The spectrum runs from full human control, where the AI presents options and the person always decides, to trusted delegation, where the AI handles the task and the person reviews the outcome when she chooses. The person sets the level. The system learns the right level through interaction. The levels are not permanent. They move.
Five levels#
The first level is FULL_HUMAN. The AI presents information and options. The person always makes the decision. Nothing proceeds without her explicit approval. This is the default for legal decisions, major medical decisions, and any action the person has flagged as requiring her direct involvement. James has his legal matters at FULL_HUMAN. When his elder law attorney recommends updating his healthcare proxy, the system presents the recommendation, the attorney’s reasoning, the relevant documents, and the scheduling options. It does not proceed until James says proceed.
The second level is ADVISED. The AI recommends a specific course of action. The person confirms or rejects the recommendation. The recommendation includes the reasoning and any relevant alternatives. James has his healthcare at ADVISED. When his health concierge identifies that his blood pressure has trended upward and recommends mentioning it to Dr. Patel at his upcoming visit, the system presents the recommendation with the supporting data. James confirms. The system notes the confirmation and includes the data in the pre-visit package sent to Dr. Patel’s office through the FHIR integration. If James rejects the recommendation, the system records the rejection and does not include the data. The rejection itself is informative: repeated rejections of health alerts may indicate that the alert threshold is too sensitive for this person, and the P-RLHF preference model adjusts accordingly.
The third level is BOUNDED_DELEGATION. The AI acts within explicit rules the person has set. Below the boundary, the AI proceeds autonomously. Above the boundary, the AI escalates. James has his bill payment at BOUNDED_DELEGATION with a $200 threshold. Utility payments, insurance premiums, and subscription renewals below $200 proceed automatically. The system pays them on time, from the correct account, without asking. An unexpected charge above $200, such as a home repair estimate from a contractor, triggers an escalation. The system presents the charge, the context, and the payment options. James decides.
The boundary is not just a dollar amount. It can be a condition set: “pay recurring bills automatically, but escalate any new vendor” or “proceed with pharmacy charges covered by insurance, but escalate out-of-pocket charges above $50.” The bounds can be as simple or as specific as the person wants. A person who sets a single dollar threshold gets a simple rule. A person who sets domain-specific conditions gets a more nuanced one. The system enforces whatever bounds the person defines.
The fourth level is LEARNED_DELEGATION. The AI observes the person’s patterns over time and acts according to what it has learned. The person does not set explicit rules. The system infers them from behavior. James has his grocery ordering at LEARNED_DELEGATION. Over six months, the system has observed that James buys the same twelve staples every week, substitutes store brands for four of them without complaint, insists on a specific brand of coffee and a specific brand of hot sauce, and adds seasonal items that correlate with weather and meal patterns. The system now composes the weekly order, applies the learned substitution rules, and submits it for delivery. James reviews the order if he wants to, but he usually does not. When he does review and overrides a substitution, the system updates its model.
The fifth level is TRUSTED_DELEGATION. The AI handles the domain and the person audits the results when she chooses. James has his entertainment at TRUSTED_DELEGATION. The system manages his streaming queue, his podcast subscriptions, his library hold requests, and his newspaper delivery. He glances at the results occasionally. He has never overridden an entertainment decision. The system interprets this as sustained trust and operates with minimal confirmation.
Domain defaults#
The five levels map to domains through a default table that reflects the stakes, the reversibility, and the regulatory context of each domain.
Healthcare defaults to ADVISED because health decisions carry irreversible consequences and professional clinical judgment is involved. The system recommends. The person confirms. The exception is routine monitoring (daily vital signs trending, medication adherence tracking), which defaults to BOUNDED_DELEGATION because the monitoring itself is not a decision; it is a data collection process the person has already consented to.
Finance defaults to BOUNDED_DELEGATION because most financial actions are routine and reversible within limits. The threshold that defines the boundary is set by the person during onboarding and adjusted over time. The system does not assume a universal threshold. A person on a fixed income of $1,847 per month has different thresholds than a person with a defined benefit pension of $4,200 per month. The system asks during setup and adapts.
Legal defaults to FULL_HUMAN because legal decisions are frequently irreversible and carry binding consequences. No legal document is signed, no legal process is initiated, and no legal communication is sent without the person’s explicit approval. The system prepares, organizes, and recommends. It does not act.
Social defaults to LEARNED_DELEGATION because social preferences are highly individual, change gradually, and are best inferred from behavior. The social connection concierge learns that James calls his daughter on Sunday mornings, meets his former colleagues for coffee on Wednesdays, and prefers text messages to phone calls for casual communication. It schedules reminders, suggests activities based on the person’s social graph, and arranges connections without imposing.
Entertainment defaults to TRUSTED_DELEGATION because the stakes are low, the consequences are reversible, and the person’s preferences are readily inferred from consumption patterns.
These defaults are overridable. Every person can adjust any domain to any level at any time. A person who wants FULL_HUMAN control over every domain, including entertainment, can have it. The system will present every streaming recommendation for approval. It will be tedious. It will be respected. The architecture does not second-guess the person’s autonomy preferences.
How levels evolve#
The agency levels are not static. They evolve through the earned autonomy mechanism described in Series 04. The system starts conservative: new domains and new expert relationships begin at ADVISED or FULL_HUMAN. As the system demonstrates competence through accurate recommendations, successful autonomous actions, and positive outcomes, it earns the right to higher delegation levels.
The earning is asymmetric. Earning trust takes time: consistent positive outcomes over weeks and months. Losing trust is fast: a single bad outcome in a high-stakes domain can drop the delegation level immediately. James’s grocery ordering reached LEARNED_DELEGATION after four months of accurate ordering with no complaints. If the system ordered a food to which James is allergic (information it should have from his health context), the delegation level for grocery would drop to ADVISED immediately and would need to earn its way back.
The regression path has its own rules. A domain that drops due to a safety event (allergy, medication error, financial overdraft) drops two levels and requires twice the demonstration period to climb back. A domain that drops due to repeated overrides (the person keeps rejecting recommendations) drops one level and requires the standard demonstration period. A domain that drops because the person manually adjusts her settings drops to whatever level she specified, with no mandatory demonstration period for climbing back. The distinction matters because the cause of the drop signals different things. A safety event means the system failed. Repeated overrides mean the system misjudged the person’s preferences. A manual adjustment means the person changed her mind. Each deserves a different recovery trajectory.
The system does not promote itself. It does not ask the person to grant higher delegation levels. It operates at the current level and, when the person demonstrates comfort (consistently approving recommendations without modification, declining to review autonomous actions, expressing satisfaction in feedback), the system adjusts internally. The next time the person looks at her delegation settings, she sees the current level reflected. She can adjust it up or down.
The override#
At any delegation level, the person can say “I decide this time.” The override is instant. It applies to this specific decision only. It does not change the default delegation level for the domain.
The override exists because no delegation model can anticipate every situation. James has grocery at LEARNED_DELEGATION, and the system handles his weekly order competently. But when his daughter and grandchildren visit for Thanksgiving, the grocery context changes. The system does not know the grandchildren’s food preferences (they are not BlueMirror users). James overrides the learned delegation, reviews the order himself, adds the items his grandchildren like, removes the items they will not eat, and submits. Next week, the system returns to its learned model.
The override is logged in the audit trail as a human decision event. The P-RLHF preference learning system notes the override context (Thanksgiving week, family visit) but does not interpret it as a general loss of trust in grocery delegation. Single-event overrides in identifiable contexts are distinguished from repeated overrides that suggest the delegation level is too high. If James overrides grocery decisions three times in a month outside any identifiable special context, the system considers whether to reduce the delegation level to ADVISED.
What mixed agency does not solve#
Mixed agency does not solve the problem of a person whose cognitive capacity changes over time. A person who competently set FULL_HUMAN for healthcare when she was cognitively sharp may no longer be able to exercise that control two years later. The system’s escalation hierarchy (Series 04) handles this through caregiver involvement and cognitive capacity assessment, but the agency levels themselves do not automatically adjust based on cognitive decline. The adjustment requires the caregiver’s input and the ethical framework’s safeguards, not a unilateral system decision.
Mixed agency also does not solve the problem of a person who sets inappropriate delegation levels. A person who sets TRUSTED_DELEGATION for healthcare because she does not want to be bothered with medical decisions is making a choice the system will respect but will not encourage. The system will operate at TRUSTED_DELEGATION for healthcare if instructed. It will also note the setting in a way that is visible to the person’s designated caregiver or healthcare proxy, because the ethical framework requires transparency about delegation settings in high-stakes domains.
James does not think about these edge cases. He thinks about the fact that his taxes get done correctly, his groceries arrive on time, his health information is tracked without nagging, and his legal matters remain firmly in his own hands. The architecture gives him exactly the control he wants in every domain. The complexity of achieving that per-domain calibration is invisible to him. As it should be.
Cross-References#
BMT-04.01 The Human Agency Scale. The foundational framework for measuring and calibrating the person’s preferred relationship to AI assistance.
BMT-04.02 Earned Autonomy. The mechanism by which the system earns higher delegation levels through demonstrated competence.
BMT-04.04 The Escalation Hierarchy. The system that ensures decisions beyond the current delegation level are surfaced to the person or her caregiver.
Technical Appendix BMT-08.02-A is available to partners and investors at partners.bluemirror.tech.