Internet-delivered psychological treatment has moved from self-help books to guided online programs and now to systems that generate their own therapeutic language. Automation is advancing both inside the research field and outside it, where many people already use general-purpose chatbots for support off-label, without trials or oversight.

This keynote asks what can be delegated. In guided internet treatment the therapist's contribution is compressed into a few minutes of written feedback made up of observable, codable behaviors, a description that doubles as a specification for machines and as a way to audit them. Emerging trials make AI-generated and human feedback hard to tell apart, on outcomes and on the relationship alike, but equivalence is easier to claim than to demonstrate. Safety separates cleanly from capability: supervised systems have behaved well, while unsupervised companion bots are another matter.

Using a five-stage taxonomy of language-based systems, I argue that technical capability and clinical effectiveness are separate axes, and that fluency at a higher level does not entail a treatment effect. The real question is therefore not whether AI can replace therapists, but which functions can be delegated, at what level of autonomy, under whose oversight, and disclosed to whom.