Every roadmap deck in 2026 has an 'agents' slide. Most of them are wrong about which work is ready — not because the models can't do it, but because the surrounding system can't yet be trusted to let them.
The workflows that are genuinely ready share three traits: the input is bounded, the success criteria are checkable, and a wrong answer is cheap to catch. Support triage, document extraction, and first-draft generation all qualify. Anything that takes an irreversible action on a system of record does not — until the permissions and audit story is built first.
The honest test we run with clients: could a new hire do this task in an afternoon with a checklist, and would you catch their mistakes before they mattered? If yes, an agent can do it now. If the task needs judgment you can't articulate, the agent will fail in exactly the places you can't see.