AI & work 3 min readDesign human oversight
AI led the task. Who still owned the decision?
A system can do most of the work while people retain decisions that determine its consequences.
The short answer
AI-led work can still depend on human supervision. Name who authorizes consequential actions, resolves exceptions and checks the result before calling the workflow autonomous.
On this page
The plan is finished. The release is not approved.
In a fictional product team, an AI prepares a release plan, drafts messages and arranges test results. A person chooses the release date, checks an unresolved customer issue and authorizes publication.
Saying AI did the task could mean that it prepared most of the material. It could also sound as though it made every consequential decision. Those are different descriptions of the same workflow.
Original conceptual example / Not measured results
A finished draft is not an approval.
About this visual
An independently authored explanatory model. It shows the condition to inspect, not measured performance or a complete implementation. Read the source and limits below.
What the measurement distinguishes
Anthropic’s measurement proposal for frontier-lab work uses an automation scale that distinguishes collaboration, AI-led work under human supervision, and fully autonomous work. Its August 2026 snapshot does not report fully autonomous operation for any measured subset of its AI research and development. The proposal acknowledges difficulties comparing labs and using models to assess model work. These are the lab’s reported measurements, not an independent audit. [1]
The decision this changes
Our interpretation: describe responsibility at decision points. Who chooses the goal, resolves exceptions, approves external action and checks the result? The amount of text the AI produced does not answer those questions.
For the release team, a finished draft is evidence of preparation. Approval is a separate event. If the customer issue remains unresolved, record what the reviewer decided and why. Do not make the word autonomous carry details the workflow has not specified.
For builders Methods, implementation details and source evidence
Make intervention observable
In a proposed oversight record, store the action awaiting review, evidence shown, reviewer decision and time of intervention. Distinguish a check before an irreversible action from a review after it. Both can be useful, but they control different risks.
Count exceptions and work returned for correction, not only approvals. If a person cannot inspect the relevant evidence or stop the action, a checkbox called human review may overstate the control.
What we don’t know
We have not inspected Anthropic’s internal systems or verified its measurements. A lab’s automation definitions do not establish how another company’s work is organized. Our release example is fictional; it supports no claim about job replacement or staffing reductions.
Try this Open the full exercise and its limits
Try this
Draw a small map of a non-sensitive workflow your team already understands. Mark where AI prepares work, where someone decides, and where an external action occurs. Add one case where required evidence is missing.
Ask who can pause that case and what they would inspect. The expected result is an explicit owner and unresolved condition, not a blanket automation percentage. This is a proposed responsibility review, not a tested control system.
What to watch next
When a report says AI leads a task, look for the supervision definition, exception handling and action boundary. Which human decision would disappear if you summarized the workflow in one percentage?
One AI decision, made clearer.
Get a short email with one visual explanation and one practical check. Follow the full article when you want the evidence.
Send me the briefingOpens our email signup page. Confirm by email to receive the briefing. Free; unsubscribe anytime.
Preview a short email →Follow the evidence
Sources & further reading
- Anthropic: Measurements for understanding the pace of AI development inside frontier labs ↗
2026 measurement proposal; reports an August snapshot and methodological limitations. Checked September 27, 2026.
Back to cited passage ↑
A correction or a different perspective? Write to us.