Skip to main content
September 2026
← All 28 pairs

Pair 3×4 · structured conjecture

AI autonomy × Deployable task coverage

Accountability tracks the old locus of work long after control has migrated to the system — the human signature survives capability's departure and just changes what it's covering for.

The full 2×2. Click to enlarge.

The four scenarios

The tempting reading is a single maturity ladder: capability grows, so we hand over more, and full delegation is just the future arriving. That's wrong — coverage and autonomy decouple. What causes delegation isn't capability, it's a support ecology: insurance, traceability, liability allocation, incident history. A broadly capable system can stay advisory for a decade because the liability won't move; a narrow one can get delegated tomorrow because speed beats hesitation and failures get buried in a contract.

What do these axes mean? ▸

Each axis is a spectrum. The card takes the two poles of each and reads off what the corner where they meet produces downstream.

Axis 3 · AI autonomy

Low
Advisory AI drafts and recommends
High
Delegated AI acts and binds others

Whether AI advises a human who decides and executes, or acts directly on the world with little mediation.

Axis 4 · Deployable task coverage

Low
Narrow / brittle patchy or pilot-scale
High
Broad / dependable whole workflows, production-grade

How much real work the system can do reliably — a few narrow pockets, or most cognitive-professional tasks end-to-end.

Advisory · Broad/Dependable

The Human Still Signs

The signature stops meaning what everyone assumes: the human picked among options the system framed, and could detect few of the subtle defects. Legal responsibility and actual causal contribution quietly diverge while the law keeps pretending they coincide — a human name on every decision, almost none of the control that name implies. If sign-off were treated as a real control surface, with teeth, this corner could run broad capability with a genuine veto intact. Mostly, it isn't.

Signs it's emerging

  • A regulator or professional body accepts named human sign-off while permitting AI-generated work at a volume no named reviewer could inspect

    The legal form of human control would have survived after the practical capacity to exercise it disappeared. — Review cadence: yearly.

  • An organisation reports a negligible human override rate while continuing to describe the workflow as human-decided

    A signature without exercised disagreement is evidence that advice has become delegation by deference. — Review cadence: yearly.

Signs it isn't

  • Reviewers get protected time, access to the underlying evidence, and real pause-and-override authority — and measured override rates show that they use it.
  • Liability follows causal control to the vendor, operator, or system deployer instead of defaulting to the person whose name appears last.
Read the essay: The Human Still Signs

Delegated · Broad/Dependable

Exceptions Only

Ordinary operation becomes invisible; supervisory judgment gets trained entirely on the unrepresentative tail — failures, escalations, edge cases — so the sense of when to intervene degrades precisely because supervisors only ever meet the pathological. Institutional tempo outruns governance, and the human fallback erodes before anyone decides to give it up. The honest version of this corner — real pause-and-reverse authority, not a signature — at least drops the pretence of control that a mere signature keeps up.

Advisory · Narrow/Brittle

The Mortgaged Apprentice

The real cost isn't the tool: juniors skip the tedious substeps that used to install judgment, so the people who'd normally be ready to supervise more autonomous systems later simply never build the taste to do it. The bill arrives a cohort late, exactly when supervisors are needed most — a slowly hollowing institution that still looks safe. It's also the best corner for deliberate skill-formation, if anyone designs for it instead of just shipping the copilot.

Signs it's emerging

  • A professional body or regulator mandates periodic unassisted-practice hours in an AI-assisted specialty

    Aviation’s answer (mandated hand-flying) arriving in a second domain would mean institutions have started pricing supervisory decay instead of discovering it at the emergency. — Review cadence: yearly.

  • Replication — or refutation — of the 2025 Lancet endoscopist-deskilling result (unassisted detection 28.4%→22.4% after AI exposure) in another clinical domain

    The first empirical anchor for skill-consumption either generalises across professions or stays a colonoscopy curiosity; the scenario’s weight moves with it. — Review cadence: yearly.

Delegated · Narrow/Brittle

Plumbing With a Throttle

Local speed beats social fear, so narrow, brittle systems get handed real authority inside obscure operational pockets — trading, cyber, claims, logistics — and each pocket reads as ordinary automation, not an AI transition. The most consequential delegation in the economy accumulates below the threshold anyone's measuring, because adoption surveys ask whether you use AI, not whether a brittle system can move money or alter security posture unsupervised. It's the sharpest attacker-advantage corner on the whole grid, invisible until a cross-domain failure cascades — but also genuinely boundable: a hard kill-switch and mandatory incident reporting would make this the safest place to practice delegation, if anyone bothered.

How this whole reading could be wrong

What would undo the pair's thesis — not any single corner.

This pair assumes delegation is gated by a support ecology (insurance, traceability, liability pricing) rather than by capability directly — if that turns out backwards, and organisations delegate the moment a system is capable regardless of whether the surrounding assurance exists, the whole decoupling claim collapses and this becomes a simple capability ladder after all. The Q4 stealth claim is also narrower than first drafted: cold cross-family review found narrow delegated systems can be high-incident and backlash-prone in customer-facing settings, not universally invisible — the stealth argument holds mainly for proprietary, high-leverage pockets, not as a blanket description of the corner.

Related pairs

Other cards that share one of these variables.

More pairs with AI autonomy

More pairs with Deployable task coverage