AI Jobs Transition (2): The work that still needs a person
A job can be highly exposed to AI and still require a person for accountability, trust, or physical execution. That protects the role from full automation, but not from redesign.
Jonathan
Founder
AI Jobs Transition · 2/6
When people estimate whether AI will replace a job, they usually count the tasks a model can perform. The more useful question is what happens after those tasks are removed.
If the remaining work includes signing a legal decision, earning a patient’s trust, teaching a child, repairing a pipe, or taking responsibility when something goes wrong, the person may still be the part of the system that cannot be removed.
OpenAI’s AI Jobs Transition Framework calls this human necessity. It explains why high AI exposure can produce two very different outcomes: full substitution in one occupation and extensive redesign in another.
Accountability cannot be delegated with the draft
The first form is regulatory and accountability necessity. Some services require a licensed person to approve a result, represent a client, or remain answerable for the consequences.
AI can research cases and draft a filing. It does not thereby become the attorney of record. It can produce a transcript, but some jurisdictions still require a court reporter to certify it. It can summarize clinical information without becoming the professional accountable for patient care.
This boundary is not merely a capability gap. Better models do not automatically receive a license, fiduciary duty, or legal liability. Organizations need an explicit accountable actor even when AI performs most of the preparation.
For an AI-native company, that changes product design. The system should not hide the human checkpoint as an embarrassing remnant of incomplete automation. It should make the checkpoint precise: what decision is reserved for a person, what evidence they receive, what they approve, and what audit trail remains.
In relational work, the human is part of the product
The second form is relational necessity. Teaching, nursing, counseling, negotiation, and persuasion deliver more than information. Trust, care, motivation, and human connection are part of the service.
A teacher may use AI for lesson plans, exercises, and feedback. The classroom still needs someone to notice confusion, maintain attention, resolve conflict, and take responsibility for students. A nurse may automate documentation and information synthesis while bedside care remains irreducibly physical and relational.
Preferences can change, and some customers already accept AI-only interactions. But teams should not confuse “the model can produce an acceptable answer” with “the customer values the service in the same way without a person.”
The right product question is not whether AI can imitate the conversation. It is whether removing the person changes trust, compliance, willingness to disclose information, or the perceived quality of the outcome.
Physical work creates a bottleneck outside the model
The third form is physical necessity. Plumbers, physical therapists, field inspectors, installers, and care workers act in the real world. Language models can diagnose, plan, document, and instruct, but the work still requires hands, presence, and adaptation to an uncontrolled environment.
This does not make physical occupations immune to AI. Administrative work can disappear, less experienced workers can handle more complex cases with guidance, and scheduling or diagnosis can improve. The bottleneck simply moves to physical execution.
Robotics may move that boundary over time. The framework is deliberately near-term: it describes what keeps people central under current technical and institutional conditions, not what will remain uniquely human forever.
Human necessity protects the role, not the headcount
This is the distinction that matters most. If a worker remains necessary for one critical task, the occupation may survive while employment falls.
Imagine a legal team where AI handles research, first drafts, document review, and routine correspondence. Lawyers remain necessary for representation and accountability, but each lawyer can supervise far more work. If lower prices do not generate enough new demand, the firm needs fewer lawyers for the same caseload.
The same pattern can appear in teaching, medicine, accounting, or any regulated profession. Human necessity blocks full substitution. Productivity still changes the staffing ratio.
That means “human in the loop” is not a labor strategy. It is only a description of system architecture. To understand employment, we must also ask how large the loop is, how often human judgment is required, and whether demand grows enough to absorb the additional capacity.
Design the human role before automating the workflow
Teams introducing agents into professional work should define the human boundary before choosing what to automate:
- Which decisions require a licensed or accountable person?
- Where does the service depend on trust, care, or persuasion?
- Which actions require physical presence?
- What evidence must the system collect before asking for approval?
- Can one person review the output without becoming a rubber stamp?
- Who owns the outcome when the model, reviewer, and workflow disagree?
The goal is not to keep a person somewhere in the process so the system can be called safe. The goal is to place human judgment where it changes the outcome, then design the context, interface, and audit trail that let that judgment work.
AI will remove many tasks without removing the occupation. The organizations that manage the transition well will make the remaining human work more explicit, more valuable, and more accountable—not merely smaller.
Sources
- Alex Martin Richmond, The AI Jobs Transition Framework, OpenAI Economic Research, April 2026.