What you should never delegate to an AI assistant

Most 'AI can't do X' lists age badly because they're really 'AI can't do X yet' lists. The boundaries worth writing down are the ones that don't move as models improve — where delegation fails not because the assistant is insufficiently clever, but because the task's value depends on a human doing it, or because being wrong is unrecoverable.

Some decisions carry personal or legal accountability that no delegation moves: signing contracts and financial commitments; regulatory filings; hiring and firing decisions; medical, legal, and major financial choices. An assistant can prepare all of these — assemble the comparison, draft the contract summary, screen against a rubric. The decision itself needs an accountable human, because 'the AI decided' is not a defence you want to need, anywhere, and least of all to MOM, IRAS, or a court.

Condolences, apologies, difficult feedback, negotiations, and the customer conversation after your mistake: the recipient's experience depends on a human having shown up. Delegating these doesn't fail because the words are wrong — modern models write eerily adequate sympathy — it fails when detected, and it eventually is, converting a relationship moment into an insult. Use the assistant to prepare (context, history, talking points); deliver in person.

The structural weak spot of language models is confident output whose wrongness only an expert can see. Novel legal positions, tax edge cases, structural engineering, medication interactions: reviewing output here requires the same expertise as producing it, so delegation saves nothing and adds risk. Delegate the retrieval and summarisation of settled knowledge; keep novel judgment with professionals.

Related, and easy to miss: never delegate the checking of the assistant's own work to the assistant. Verification needs independence — a human, a second system, or reality.

Before delegating anything, ask: if this goes wrong in the worst plausible way, is the outcome a correction or a catastrophe? Corrections — a rescheduled meeting, a clumsy draft, a wrong research lead — are the assistant's learning budget. Catastrophes — a lost client, a legal exposure, a breached confidence — stay human, whatever the demo showed. Our concept pages encode this by listing, for every assistant we design, exactly what it may do alone and what always needs approval; a vendor who can't produce that list for their product hasn't thought about it.