Nadia has spent eight years as a product manager at health technology companies, long enough to have watched three generations of “patient empowerment” products fail in the same way. The first generation gave patients data and called it empowerment. The second generation gave patients recommendations and called it support. The third generation gave patients AI-driven decisions, presented as personalization, without telling them how much the system had decided for them. The person received the output and was expected to feel in control of a process she had never seen.
When Nadia joined the BlueMirror product team, her first question was about the autonomy model. Not “how much autonomy does the system have?” but “how does the person know how much autonomy the system has, and how does she change it?” These are different questions. Most AI products answer only the first.
The answer to both is the Human Agency Scale.
Why a scale, not a switch
Binary autonomy fails because life is not binary. Margaret wants her medication reminders fully automated, her appointment scheduling to require her confirmation, her financial decisions to be advisory only, and her entertainment recommendations to do whatever they want. That is four different autonomy levels within one person’s morning. A single on/off switch cannot express this. A scale with domain modifiers can.
The Human Agency Scale is a 0.0-to-1.0 spectrum that determines how much the system can do without asking. At 0.0, the system does nothing without explicit instruction. At 1.0, the system handles everything and reports back. Nobody operates at either extreme. Real life happens between 0.3 and 0.8, and the right setting varies significantly by domain.
The critical innovation in the HAS is not the scale itself. Scales are easy. The innovation is the domain modifier system that applies different effective autonomy levels to different areas of life based on each domain’s risk profile. A person with an overall setting of 0.7 is not granting 0.7 autonomy everywhere. She is granting 0.7 as a base from which the system computes an effective autonomy per domain, modifying it up or down based on the stakes involved.
Six levels with behavioral signatures
Six points on the spectrum have distinct behavioral signatures, and naming them makes it easier for people to recognize where they sit without translating from decimal notation.
Full Manual (0.0) means the system waits for explicit instructions. It never initiates and never suggests. Every action begins with the person. Every output is the direct consequence of something the person asked for. This is the appropriate setting for someone who wants complete control, for domains where confidence in the system has not yet been established, or for any situation where the person wants to experience the decision-making process directly rather than delegate it.
Informed (0.2) means the system monitors and surfaces information. “Your blood pressure was higher than usual this morning.” No recommendation follows. The system provides awareness and stops. The appropriate setting for domains where the relevant expertise belongs to a professional and the person wants to be kept current without being guided.
Advisory (0.4) means the system monitors, informs, and recommends. “Your blood pressure has been trending up for ten days. You might want to talk to Dr. Patel.” The recommendation is explicit. The decision is entirely the person’s. The system does not schedule the appointment; it proposes that scheduling might be warranted and waits to be asked.
Guided Automation (0.6) means the system acts within pre-approved patterns and notifies afterward. “I refilled your metformin because you were down to three days. The new prescription arrives Thursday.” The person can reverse any action, but the system did not wait for permission. The action fell within the range the person previously authorized and she learns about it after it has been taken.
Trusted Automation (0.8) means the system handles most decisions and surfaces exceptions only. “Your appointments are set for the month. One change from your usual preference: Dr. Patel moved to Thursdays, so I shifted your quarterly check-up. Let me know if Thursday doesn’t work.” Routine decisions happen invisibly. Meaningful deviations from pattern are surfaced. The person’s cognitive load is low because she only sees what genuinely requires her attention.
Full Delegation (1.0) means the system handles everything and the person reviews periodic summaries. This level is available for entertainment and routine scheduling, where any individual decision is low-stakes and reversible. It is never the effective level for healthcare or finance because domain modifiers reduce the effective autonomy in those areas regardless of what the overall setting is.
Domain modifiers: the mechanism that makes the scale real
The domain modifier system is what separates the Human Agency Scale from every “user control” slider that consumer AI has produced. A 0.7 setting without modifiers means the person has granted 0.7 autonomy in healthcare, finance, legal, home maintenance, entertainment, and every other domain equally. This is wrong. The same person who is comfortable with full delegation in entertainment is not comfortable with full delegation in healthcare, because the consequences of a wrong decision are categorically different.
Nine domain modifiers adjust the effective autonomy from the person’s stated base:
Healthcare carries a -0.3 modifier. A person at 0.7 overall has an effective 0.4 in healthcare: Advisory level. Clinical risk, regulatory requirements around informed consent, and the irreversibility of many medical decisions make the advisory level the appropriate floor even when the person wants maximum delegation. At 0.4 in healthcare, the system recommends but does not act on clinical matters. Appointment scheduling and medication reminders can operate at higher effective autonomy within healthcare’s routine coordination tasks; clinical decisions stay at advisory.
Financial carries a -0.2 modifier. Effective 0.5 at a 0.7 overall: advisory and guided for routine bill payment and standard reorders, but nothing irreversible above a defined financial threshold without confirmation. The threshold is person-configurable. The modifier is not.
Legal carries a -0.4 modifier. Effective 0.3 at 0.7 overall. The representation boundary means the system prepares but the person must authorize. The legal agent never crosses the line from assistance into advice, and the heavy modifier reflects that the consequences of stepping over that line are significant and not easily undone.
Home maintenance carries a 0.0 modifier. Effective 0.7. Scheduling an HVAC service or coordinating a plumber is genuinely low-risk automation. The person’s stated preference applies directly.
Nutrition carries a -0.1 modifier. Effective 0.6. Dietary guidance has health implications, so the modifier pulls it toward guided automation with notification rather than trusted automation.
Social and family coordination each carry a -0.1 modifier. Effective 0.6. The relational stakes in social and family decisions warrant a slight pull toward the guided level: the system handles coordination but surfaces decisions that might affect relationships rather than making them invisibly.
Earning carries a -0.2 modifier. Effective 0.5. Financial and cognitive dimensions of earning activities require more conservative handling than purely domestic tasks.
Entertainment carries a +0.2 modifier. Effective 0.9 at a 0.7 base. Low stakes and high automation benefit combine to make near-full delegation sensible in entertainment. The worst consequence of a wrong entertainment recommendation is a movie the person does not enjoy.
Home environment carries a 0.0 modifier. Effective 0.7. Ambient management of lighting, temperature, and household systems is routine automation that the person has configured for her space.
Modifiers are defaults, not mandates. If Margaret wants an effective 0.8 in healthcare after two years of experience with the system and a high degree of confidence in its clinical coordination, she can set it. The system records the override, notes the deviation from default, and adjusts. The override is her deliberate choice, visible in the interface, and changeable at any time. The default exists because most people have not thought through the implications of 0.8 in healthcare. The override exists because some people have.
Setting the scale at onboarding
The HAS is configured at onboarding through concrete questions rather than a ten-domain configuration spreadsheet. “How much do you want the AI to handle on its own?” followed by plain-language options that describe behavior rather than numbers. “Handle routine tasks and tell me afterward.” “Recommend and wait for my go-ahead.” “Keep me informed but let me decide everything.” The system translates the responses into HAS values and applies domain modifiers automatically.
The first thirty days operate conservatively regardless of what the person states as her preference. The system asks more than strictly necessary. It learns what the person actually wants through behavioral signals: which recommendations does she accept without modification, which does she override, which does she ignore entirely. The pattern across thirty days of real interactions is more informative than a stated preference given without experience of the system.
After the initial period, the system proposes autonomy adjustments based on demonstrated competence. The proposals are explicit, not silent. “I have managed your medication refills without error for six months. Would you like me to handle these automatically from here?” The person agrees or declines. The system does not quietly drift toward higher autonomy. It asks for each promotion.
When the system makes a mistake, it proposes reducing autonomy in the affected domain until confidence is rebuilt. “I made an error with your last grocery substitution. I would like to check in with you on substitutions for the next few weeks while I recalibrate.” The mistake is not hidden. The adjustment proposal is honest about its cause. A system that hides its errors to protect its autonomy level is not a system the person should trust.
The bidirectional principle
Autonomy is not a one-way ratchet. The system earns the right to do more. The person also earns the right to do more.
If Margaret’s cognitive assessment improves after a medication change and she starts managing her own appointments again, the system does not maintain the higher automation level it used during the period when it was handling everything on her behalf. It notices the pattern and asks: “You have been managing your appointments yourself this week. Want me to stop handling those?” The question is genuine. Either answer is equally acceptable. The system does not treat the reclamation of a task as a problem to be managed or a regression to be corrected.
The person is never locked into dependency. A system that gradually takes over more and more, making itself indispensable, is not serving the person. It is creating a fragility: the person whose capabilities atrophy because the system handles everything is less able to evaluate the system’s decisions, less able to catch errors, and more vulnerable if the system fails. The bidirectional principle means the system has a structural obligation to surface opportunities for the person to re-engage with domains it manages, not just to accept further delegation.
This is the feature that most AI products do not build because it actively works against engagement metrics. A person who re-engages with tasks the system was handling uses the system less. That is, from an engagement perspective, a failure. From a service perspective, it is a success: the person is more capable, more aware, and better positioned to be a meaningful participant in decisions about her own life.
The cognitive dimension
The Human Agency Scale has a third modifier that operates in real time rather than as a domain-fixed setting: the cognitive state modifier. The Cognitive State Estimator produces a continuous 0.0-to-1.0 estimate of the person’s current cognitive function from behavioral signals. When that estimate drops, the effective HAS level adjusts downward automatically. When it recovers, the effective level returns to the person’s stated preference.
A person with a stated preference of 0.65 in healthcare and a cognitive state estimate of 0.72 (slightly reduced) has an effective level of approximately 0.55: the system acts more conservatively, surfaces more decisions for her attention, and uses simpler language. The adjustment is not labeled. She does not receive a notification that her autonomy has been reduced. She experiences a system that seems to need a little more from her today, which is an accurate description of what the system is doing and why.
The cognitive modifier is the subject of a deeper treatment in BMT-04.05. The important principle here is that the HAS is not a static document. It is a live, responsive system that serves the person where she is rather than where she was when she set her preferences.
Nadia’s pair of questions, how does the person know how much autonomy the system has, and how does she change it, have complete answers. She knows because the HAS, the domain modifiers, and the current effective level per domain are all visible in the interface at any time. She changes it by telling the system, by her behavior over time, or by accepting or declining the system’s proposals. The scale responds to her, continuously, and it does not change without her involvement except where the domain ceilings and cognitive modifier apply, which are themselves visible, documented, and in the case of domain modifiers, overridable by her deliberate choice.
Cross-References#
Earned Autonomy (BMT-04.02). How the HAS evolves through demonstrated competence rather than remaining static at the configured level.
Cognitive Capacity and Consent (BMT-04.05). How the scale responds when cognitive capacity changes, and what the cognitive modifier does at the deeper architectural level.
When Agents Disagree (BMT-02.06). How HAS settings affect conflict resolution when concierge agents have competing priorities for the same resource or time.
The Cognitive Concierge (BMT-01.07). The concierge agent most directly affected by HAS adjustments as capacity changes, and whose output drives the cognitive modifier.
Technical Appendix BMT-04.01-A is available to partners and investors at partners.bluemirror.tech.