James Okafor retired from aerospace engineering three years ago. He spent thirty-one years at Pratt and Whitney, the last twelve as a senior propulsion systems analyst. He knows gas turbine thermodynamics the way most people know their commute: every curve, every failure mode, every shortcut that works and every shortcut that kills. He is seventy years old. He lives in East Hartford, Connecticut, on a pension and Social Security. He is not looking for a job. He is looking for a reason to use what he knows.
His daughter sent him a link to BlueMirror’s BGO program six months ago. She thought he might find it interesting. He found it more than interesting. He found it structurally different from anything he had encountered in three years of restless retirement. The system did not ask him to volunteer. It did not ask him to tutor. It asked him a question he had never been asked before: what do you know that other people need?
The Expert Exchange Layer is the architectural answer to that question, for James and for the millions of aging adults whose expertise is not lost when they retire but simply disconnected from the people who could use it. The architecture connects three distinct pools of expertise to the person who needs help, using routing logic that weighs trust, cost, urgency, availability, and what the system has learned about this specific person’s preferences.
The personal circle#
The first pool is the people who already know you. The neighbor who spent twenty years as an electrician and can tell you whether that outlet needs a professional or a trip to the hardware store. The niece who is a registered nurse and fields your medication questions on Sunday afternoons. The friend from church whose husband went through prostate cancer treatment last year and who knows things about insurance appeals that no website publishes.
These experts existed before AI. The system does not create them. The system formalizes routing to them.
The trust model for personal circle experts is lifetime relationship. You trust your neighbor’s electrical advice because you have known him for fifteen years and he has never steered you wrong. The system does not credential-check your neighbor. It does not verify his electrical license (he may not have one; he was a industrial electrician, not a residential one). What it does is track outcomes. When you follow your neighbor’s advice and the outlet works, the system notes a positive outcome. When you follow his advice and the breaker trips, the system notes a negative outcome. Over time, the system learns which personal circle experts produce good outcomes in which domains for you specifically.
The system also tracks negative capabilities: the domains where a personal circle expert should not be trusted even though they seem relevant. Your neighbor is an excellent source for basic wiring questions. He is not a licensed residential electrician, and the system learns, through outcome tracking and through the safety boundaries defined in the architecture, that questions involving electrical panel work, circuit additions, or code compliance should route to a licensed professional regardless of the neighbor’s willingness to help. The neighbor may be willing and even competent. But the liability boundary and the person’s safety require professional credentials for work that could kill someone if done incorrectly.
Payment in the personal circle is reciprocal or nonexistent. Your neighbor does not charge you for electrical advice. You do not charge him for the tax preparation help you provide each February. The system does not insert a payment layer into relationships that have always been free. It tracks the reciprocity so the person can see the exchange: “You’ve helped Tom with three tax questions this year. Tom has helped you with two electrical issues and one plumbing question.” Visibility, not monetization.
The access pattern is availability-based. Your niece the nurse works day shifts on weekdays. The system learns this and routes nursing questions to her on evenings and weekends, or queues them as non-urgent when the question can wait. Your neighbor is usually available Saturday mornings. The system learns this too.
The professional registry#
The second pool is credentialed experts with formal qualifications and fee-based relationships. Dr. Chen the primary care physician. The CPA who handles annual taxes. The elder law attorney who prepared the power of attorney documents. The physical therapist. The home health aide.
The trust model for professional registry experts is credential-based. Dr. Chen is trusted for medical advice because she has a medical license, board certification, and an established patient relationship. The CPA is trusted for tax advice because he holds a CPA license and has filed the person’s returns for seven years. The system verifies credentials through professional registry databases and licensing boards, checks for malpractice actions and disciplinary history, and maintains the verification on a recurring basis. A credential that was valid at onboarding may lapse. A disciplinary action may appear after the relationship began. Recurring verification catches these changes. The system does not interrupt an existing relationship on a flag. It surfaces the information to the person and lets her decide.
Payment follows the existing model: insurance, fee-for-service, retainer. The system does not disrupt existing payment relationships. It does handle administrative friction. Scheduling Dr. Chen for a follow-up, submitting prior authorization paperwork to the insurer, preparing the CPA’s annual document package, delivering the attorney’s requested records: these are tasks that the system performs or assists with, not payment model changes. The scheduling alone is worth attention. A person who needs to see three specialists in a month currently makes three phone calls, navigates three scheduling systems, and manages three sets of appointment logistics. The system handles the scheduling coordination, finds appointment times that do not conflict with each other or with the person’s other commitments, arranges transportation if needed, and prepares the context package each specialist will need.
Context sharing with professional registry experts is governed by professional obligation and applicable law. Dr. Chen gets the health context she needs under HIPAA. The CPA gets the financial information he needs under accountant-client privilege. The attorney gets the legal context she needs under attorney-client privilege. The system packages this context through the membrane described in Series 03, exposing what is relevant and withholding what is not.
The AI agent marketplace#
The third pool is commercial AI agents that provide instant, specialized expertise. LegalAI for document review, scanning a lease for unfavorable clauses in seconds rather than the hours a human attorney would take. TaxBot for quarterly tax estimates, computing estimated payments from the person’s current income and deduction data. CardioAI for medication interaction checks, cross-referencing the person’s full medication list against a drug interaction database with sub-second response time.
The trust model for AI agents is vendor credentials and performance history. An AI agent enters the marketplace with a certification from the platform (described in BMT-08.05) that validates its accuracy, safety, and bias profile for its declared capability. The certification is not permanent. It includes a recertification schedule (typically annual, quarterly for high-stakes domains like healthcare and finance) and ongoing performance monitoring. An agent whose medication interaction checks miss a known interaction loses certification immediately, not at the next review cycle.
Payment is per-query or subscription, depending on the agent vendor’s pricing model. The system presents costs to the person before routing a query to a paid agent. A person who sets a per-query spending limit of $2 will not be routed to an AI agent that charges $5 per query unless the system escalates the cost decision to her first. The cost governance integrates with the buying agent’s spending controls described in BMT-01.03. A person who has set a monthly AI agent budget of $30 will see the system shift toward free or lower-cost alternatives as the budget ceiling approaches, with transparent notification about why routing changed.
Context sharing with AI agents is governed by the membrane and the agent’s exploration bounds. An AI agent does not see the full MoC. It sees the minimum context required for its declared function. LegalAI reviewing a lease gets the lease text and the person’s state of residence for jurisdiction analysis. It does not get the person’s health data, financial situation, or family relationships. The context package is assembled by the system, not requested by the agent. The agent receives what the system decides it needs, not what the agent asks for. This distinction is critical: a well-designed AI agent might request more data than it needs, reasoning that more context produces better results. The membrane does not honor overbroad requests. It delivers the minimum viable context defined by the agent’s capability schema, filtered by the person’s consent.
Hybrid teams#
The fourth arrangement is not a pool but a composition: AI triage followed by human review. The AI agent runs the initial analysis. The human expert reviews the result and decides.
The composition is the most cost-effective arrangement for queries that are routine 80% of the time and complex 20% of the time. A medication interaction check is routine when the interaction database returns a clear answer. It is complex when the interaction is conditional (depends on dosage, depends on renal function, depends on other medications that modify metabolism). The AI agent handles the routine check in sub-second time. When the check returns a conditional result, the system routes to a human pharmacist or physician with the AI’s analysis attached as context. The human expert does not start from scratch. She starts with the AI’s work, reviews it, and applies the clinical judgment that the conditional case requires.
The hybrid composition also applies in reverse: human first, AI second. A physician who makes a treatment recommendation can have the AI agent verify the recommendation against the person’s full medication list, allergy history, and insurance formulary before the recommendation reaches the person. The physician provides the judgment. The AI provides the verification. The combination produces a recommendation that is both clinically informed and practically validated.
The person does not manage this composition. She asks a question. The system decides whether AI alone can answer it, whether a human is needed, or whether the hybrid path (AI first, human second) is appropriate. The routing logic learns her preferences: some people always want a human for medication questions, regardless of complexity. Others prefer the AI for initial triage and accept human involvement only when the AI flags uncertainty. The system adapts.
How the system routes#
The routing decision weighs five factors. Urgency: a chest pain question routes to the highest-confidence available source immediately, bypassing cost and preference considerations. Cost: for non-urgent queries, the system considers whether a free personal circle expert, a paid professional, or a commercial AI agent is the most cost-effective match. Trust: the person’s accumulated trust in specific experts and expert types influences routing. Availability: an expert who is unavailable cannot be routed to regardless of other factors. Learned preference: P-RLHF, the preference learning system described in BMT-02.05, tracks which routing decisions the person accepts, overrides, or regrets, and adjusts future routing accordingly.
The routing is not a simple priority ranking. It is a weighted decision that produces different outcomes for different people in the same situation. Margaret, who has a strong relationship with her local pharmacist and has overridden AI routing to pharmacy questions four times, gets routed to her pharmacist first. Dorothy, who has accepted AI triage for all routine medication questions and has never overridden it, gets routed to the AI agent first. Both are correct. The system serves the person’s demonstrated preferences, not a population-average optimization.
James Okafor appears in this architecture twice. He is a person being served, receiving expertise from all three pools. And he is an expert, sitting in the personal circle of the colleagues and former students who know his propulsion knowledge, and increasingly in the professional registry through the BGO program that is converting his expertise into structured, deployable knowledge. Series 08.04 describes how that conversion works. For now, the point is this: the Expert Exchange Layer does not divide the world into experts and recipients. It recognizes that most people are both, at different times, in different domains.
Cross-References#
BMT-03.02 Trust Tiers. The trust tier architecture that governs how AI agents in the marketplace are credentialed and how their context access is bounded.
BMT-02.05 Preference Learning. The P-RLHF system that learns routing preferences and adapts the decision matrix to the individual.
BMT-04.02 Earned Autonomy. The delegation framework that governs how much routing authority the system earns through demonstrated competence.
Technical Appendix BMT-08.01-A is available to partners and investors at partners.bluemirror.tech.