An agent can finish a task and leave the business exactly where it started.
A proposal exists. Nobody sent it. A lead list is complete. Nobody checked whether the people belong on it. A report arrived on time. The decision is still waiting.
This is the part of the AI conversation I care about: the distance between an impressive output and a useful consequence.
The public debate often asks whether AI will make humans obsolete or make everyone more capable. Both frames can hide a smaller question with an answer you can test tomorrow: where does responsibility go after the machine finishes?
A machine can contribute real intelligence without owning the result. That distinction matters most when the output leaves the chat window.
The missing job
Consider a hypothetical small design studio. A prospective customer asks for a quote. An assistant reads the brief and drafts a polished proposal in minutes.
That sounds like automation. But the owner still needs to check the scope, decide whether the deadline is realistic, confirm the price, send the proposal, and make sure the customer receives it.
The draft may have saved valuable effort. It has not completed the commercial job. If the studio measures only drafting speed, it can celebrate an improvement while the customer is still waiting.
The practical question is not whether to use the assistant. It is which part of the job the assistant can reliably own, and what evidence closes the handoff.
Capability is uneven. Responsibility cannot be vague.
Research helps explain why sweeping productivity claims deserve care. METR’s July 2025 study found that experienced open-source developers took longer with the early-2025 AI tools tested, despite believing they had worked faster. That result concerns a particular group, task setting and generation of tools—not every person using AI today.
Anthropic’s June 2026 analysis of Claude Code sessions describes people making most planning decisions while the agent makes most execution decisions. This is a different method and setting, not a direct replication or rebuttal. Together, they are a reason to inspect the actual work rather than assume that a fast answer equals a finished outcome.
A small ownership test
Pick one recurring job. Something boring enough to repeat and important enough that failure has a visible cost.
Write these four lines:
The job ends when: [an observable result, not a file name].
The machine may decide: [a bounded set of actions].
A person must decide: [exceptions, commitments or judgment calls].
The receipt is: [the evidence that the result happened].
For the design studio, the receipt might be a reviewed proposal delivered to the intended customer—not “proposal generated.” The human still approves price and scope. The assistant can prepare the draft and surface missing requirements.
Then run one ordinary case. Count the checking, correcting and handoff work, not just the minutes inside the AI tool. Do not add another agent until you understand the unresolved handoff.
The point is not to make humans review everything forever. It is to discover which responsibility can be transferred safely, which still needs judgment, and whether the whole job became easier.
What I want from an intelligent system
I want it to reduce the number of things a person has to remember without making the truth harder to see.
That is a more demanding standard than “produces a lot.” It is also a more human one. A system should create room for a better conversation, a decision made with care, or an evening that actually ends.
Tomorrow, choose one machine-generated output and follow it to its consequence. If the trail stops at a document, you have found the next piece of work.
Sources
METR, July 10, 2025: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
Anthropic, June 16, 2026: https://www.anthropic.com/research/claude-code-expertise
The studio example is hypothetical. The ownership test is a proposed exercise, not a claimed Flow customer result.