Capability / delegation / consequence
When does an AI agent
become an economic actor?
Not when it answers. Not necessarily when it completes a task. The threshold appears when delegated intelligence can represent intent, commit resources, and leave consequences behind.
Test the threshold →The question beneath the label
Intelligence can advise.
Authority can alter the world.
A system that finds a cheaper flight is a tool. A system authorized to buy it, negotiate a refund, rearrange the trip, and accept a penalty has crossed into a different category.
The singularity may first appear not as a single superintelligence, but as millions of narrow delegations: software that can transact, negotiate, hire, allocate, and coordinate at machine speed. The decisive variable is not IQ alone. It is the combination of capability and permission.
The delegation threshold
Build an agent.
Watch its status change.
Move the five controls. The classification is an explanatory model—not a legal test or scientific scale. Its purpose is to expose which ingredients people quietly bundle into the phrase “economic actor.”
The system completes multi-step work, yet a human still owns the consequential decision and market commitment.
Authority and consequence are gates: high general capability alone cannot produce the final classification.Four commitment tests
A transaction is more
than a tool call.
It can represent a principal
The agent must infer and preserve someone else's goals well enough to act when preferences are incomplete or conflicting.
It can cross a commitment boundary
Advice becomes agency when the system can spend, exchange, promise, schedule, hire, or otherwise bind resources within delegated limits.
Its actions survive the session
Economic action unfolds through time: counterparties respond, prices change, obligations remain, and the agent must recover when its plan collides with reality.
Responsibility remains legible
An actor without an audit trail is not autonomous in a useful institutional sense; it is an unassignable risk.
The disagreement
One transition.
Two starting points.
“Economic agency begins when useful work can be completed.”
- Agents already sustain longer workflows, use tools, and cross occupational task boundaries.
- A principal can set a budget and rules, then let the system search and execute inside them.
- Markets have always accommodated agents—employees, brokers, and software—without requiring perfect autonomy.
“No authority, accountability, or standing—no actor.”
- Fluent negotiation can conceal a poor model of what the human actually wants.
- Legal responsibility still lands on the person or institution that deploys the system.
- Real markets contain fraud, recourse, identity, taxes, and obligations that benchmark tasks omit.
Observatory assessment
Agency is an envelope,
not an on-off switch.
Today's strongest systems can act as operators and bounded proxies. They become economic actors only inside a specific delegation envelope: who they represent, what they may commit, how long they may act, what can be reversed, and who answers when the outcome is wrong.
The practical threshold will arrive unevenly—first in digital markets with machine-readable rules, low reversibility costs, and strong audit trails; later where preferences are tacit and errors cannot be undone.
Evidence record
Read the claims
at their source.
Company research is useful evidence about observed systems, not a neutral verdict on economy-wide impact. The Observatory distinguishes reported findings from its own synthesis.
Anthropic · September 2026Project Swap: Can AI agents negotiate for people?In a controlled book-barter market with 201 participants, agents reproduced 61% of participants' pairwise preferences after a short interview; incomplete preference information was the largest source of failure.
Anthropic · September 2026What work can robots do?Technical exposure and economic viability diverge sharply: the study estimates current robots are cost-competitive for only a small fraction of job tasks.
OpenAI · June 2026How agents are transforming workOpenAI reports that more than 70% of May Codex users asked it to perform tasks estimated to take a human longer than one hour.
OpenAI · July 2026How AI is expanding what people do at workAn analysis of 800,000 messages found AI use crossing traditional occupational boundaries and creating new combinations of tasks.
Anthropic · January 2026Economic Index: The economic primitives of AI useA task-level framework separates complexity, skill, purpose, autonomy, and success—useful evidence that economic effect cannot be reduced to raw capability.
Anthropic · June 2026Economic Index: CadencesAnthropic's usage data tracks a growing share of longer-running, agentic tasks while emphasizing that observed use is not the same as economy-wide impact.
The dossier changes when the evidence does
Follow the
delegation frontier.
Receive new dossiers, forecast revisions, and the evidence that changed our mind.