The person of record

The professions most likely to stay human-anchored are not necessarily the ones with the hardest tasks. They are the ones where some productive part of the work cannot be detached from a particular person.

Capability forecasts usually rank work by whether a model can perform the visible task. That misses three independent constraints:

1. Substrate accessibility: can the actor obtain the context needed to act competently?
2. Execution verifiability: can its actions and consequences be evaluated, attributed, and reconciled?
3. Authority portability: can the right to commit, attest, decide, or accept liability be transferred to that actor?

An agent may be capable, informed, and verifiable while still lacking standing to perform the final act. It may also have authority in principle while being unable to perceive the local substrate on which competent action depends. “Can the model do the task?” compresses these different limits into the least interesting one.

Substrate	Execution	Authority	Likely operating model
Machine-addressable	Verifiable	Transferable	Predominantly agent-operated
Machine-addressable	Verifiable	Person-bound	Agent execution with a human of record
Person-bound	Verifiable	Person-bound	Human substrate anchor with extensive agent support
Locally addressable	Verifiable	Transferable	Embedded or on-premises agent operation
Inaccessible	Poorly verifiable	Person-bound	Persistently human-heavy work

The local row matters. A substrate need not be widely distributable to become machine-operated. Patient data, classified systems, a factory control network, or a particular physical site can remain inside its boundary while models move to it. Centralized SaaS access is one deployment model, not a condition of automation.

Four ways work stays bound

Some substrate is representationally bound. A person’s perception, motor skill, local intuition, or exception knowledge has not yet been captured in a form machines can use. Multimodal models, sensors, simulation, and robotics continually erode this category. Much tacit knowledge is currently unrepresented rather than permanently unrepresentable. These occupations may be delayed without being protected.

Some substrate is locationally bound. It can be represented but must remain inside a hospital, factory, deal room, jurisdiction, or secure environment. This resists distribution more than automation. It favors local inference, private models, secure access cells, and capability gateways.

Some substrate is relationally bound. The productive asset is not information about a relationship but the counterparty’s willingness to extend trust to this person. A dossier can describe a rainmaker’s clients, a physician’s patient history, or a diplomat’s contacts. It cannot transfer their standing. Agents can operate much of the surrounding workflow while the person remains the relationship-bearing endpoint.

Some authority is institutionally bound. A named engineer, physician, auditor, director, fiduciary, judge, or responsible officer must answer for an outcome. This is not a failure to encode the knowledge. Person-binding is the institution’s chosen accountability mechanism.

The last two bindings are more durable than the first two, but not natural laws. Trust can migrate from individuals to institutions, brands, protocols, or artificial actors. Law can move liability from a professional to an employer, vendor, or certified system. These constraints change politically and culturally rather than automatically with benchmark performance.

Human-anchored does not mean human-operated

A profession can retain a human core while losing most of its present task composition.

An engineer of record may not personally produce most calculations, drawings, simulations, comparisons, or evidence. The engineer can instead define the assurance envelope, resolve exceptions, challenge assumptions, approve material claims, and accept residual risk. A senior analyst can own semantics, acceptance criteria, disputed adjustments, and attestations while agents perform much of the transformation and reconciliation. A trusted adviser can remain present at the relationship-bearing decision while agents prepare and operate everything around it.

This is more than supervising an assistant. The person becomes the anchor of an institutional capability.

The corresponding measure of organizational capacity is no longer merely headcount:

How much agent-operated scope can one accountable person safely anchor?

Contract quality, verification strength, consequence severity, exception rate, substrate volatility, and relational load all constrain that ratio. Better agents can widen it. They cannot make it infinite.

The obvious failure is nominal accountability without effective control. A person may sign for hundreds of agent-operated processes while lacking the time, evidence, or comprehension to challenge any of them. The human of record then becomes a liability sink: legally present, operationally absent.

An assurance substrate has to protect that person as well as the institution. It needs materiality-based escalation, comprehensible evidence, visible unresolved assumptions, limits on delegated scope, overload detection, refusal paths, and a record of what was never presented for judgment. A signature is not governance if the system made meaningful review impossible.

Management separates along the same line. Transporting status, assigning routine work, chasing dependencies, and translating updates are increasingly substrate operations. Resolving incompatible preferences, maintaining trust, allocating legitimate authority, and deciding which ambiguity matters remain human-anchored longer. Management does not simply disappear. Administrative management contracts while judgment-bearing management becomes more leveraged and more accountable.

There is a symmetry with software assurance. AI divides outputs into derived realizations and artifacts of record. It divides participation into replaceable operators and persons of record. In both cases, the exact identity becomes load-bearing where fungibility is deliberately refused.

The future organization will have to make both bindings explicit:

Which output must remain bound to this exact artifact?
Which decision must remain bound to this exact person?

Everything around those bindings may change radically. The bindings themselves are where the institution says consequence becomes accountability.