Why automation vs augmentation is the wrong frame.

The binary asks about the worker and the task. The unit AI substitutes is a piece of a bundle - not a worker, not a single task. When AI handles part of the bundle, the bundle reconfigures: some pieces survive in the original role, some get rebundled into a different role, some get absorbed by an algorithmic platform. 'Automated' and 'augmented' are both inadequate descriptors of what happened.

Augmentation studies measure worker-level productivity gains. The gains rarely accrue to the worker. Value migrates upstream to the provider and, within the firm, to whoever sits above the algorithm. 'The worker is augmented' is a useful sentence only if the worker captures the gain - which is the question the binary doesn't ask.

Both poles of the binary presuppose a clean partition between an AI task and a human task. Agentic AI executes across the chain - calling tools, taking actions, reading replies, deciding the next call. There is no single task to assign. The binary's grammar cannot describe what an agent does.

Across roles where AI is most capable, the surviving human role centers on signature, attestation, sign-off, risk-bearing. AI produces the read, the filing, the diagnosis. What survives is the liability function, not the capability function. The binary calls this 'augmentation' because the human is still in the picture - but the role bundle has reconfigured around the constraint AI cannot dissolve, not around AI making the human better.

The binary lives entirely inside the firm. It has no name for what happens when unbundled fragments leave the firm and get reorganized through algorithmic platforms - dispatched, ranked, priced, monitored by the algorithm. That third locus is neither automation nor augmentation; it is a structural reorganization of where work happens.

Both poles assume a steady state. Centaur explicitly assumes a stable human and AI complementarity. Asymmetric capability gain means yesterday's augmentation becomes today's automation, and the answer keeps moving. A binary that returns one of two answers cannot describe a phenomenon that updates its answer every six to twelve months.

Augmentation assumes AI affects one worker at one task. Cross-functional foundation models substitute across sales, engineering, support, and HR with the same model. The 'augmentation' of a sales rep is also the reorganization of the sales function and the displacement of the surrounding org structure. The binary thinks one role at a time; the model does not.

The binary assumes the capability to do the work lives inside the firm, so automating or augmenting is a choice the firm makes with assets it owns. Increasingly the model, the data, and the orchestration sit upstream with providers and platforms - the firm rents the capability rather than holding it. That reframes the decision: the firm no longer fully controls whether the work gets automated, and the surplus flows to whoever owns the layer the work depends on.