silicon valley keeps trying to sell the universal assistant.
the pitch changes costumes. siri will handle your phone. alexa will run the house. google assistant will know what you mean before you do. chatgpt, claude, and gemini now arrive with better language and the same old promise: one polite box that can do everything.
i don't think that is where the useful thing is.
uber was not great because it gave everyone a private driver. it was great because dispatch got better. spotify was not great because it replaced record-store clerks with "music intelligence" in the abstract. it was great because it routed you through a messy little graph of songs, people, skips, repeats, moods.
the luxury was expertise plus routing.
the old universal assistants failed partly because the models were weak. but the shape was wrong too. ask siri to set a timer and it is fine. ask it to negotiate across calendars, preferences, prices, airports, children's bedtimes, and one person who refuses early flights, and the whole fantasy starts sweating.
the older AI people already knew this, in a cruder form. expert systems were built for bounded domains. MYCIN tried to help with infectious disease diagnosis. XCON configured DEC VAX computer orders. these systems were narrow, brittle, and sometimes genuinely useful.
they also ran into boring, fatal engineering problems. knowledge acquisition was expensive. rules conflicted. maintenance got ugly. closed-world assumptions broke as soon as reality arrived in the wrong shape. a human expert could say "hmm, not that kind of exception." the system needed someone to encode the exception, then the exception to the exception.
so expert systems became a cautionary tale.
fair enough. but the caution was not "specialization is fake." it was "hand-maintained rule boxes do not scale well."
modern agents are different things wearing similar clothes. an expert system is a rule base. a tool-using agent is a controller that can call software. a mixture-of-experts layer is a neural routing trick inside some models. human delegation is an organizational pattern with politics and trust and blame.
those should not be blurred together. but they rhyme.
the useful AI interface probably looks less like jarvis and more like a dispatcher. a task comes in. the router decides whether this is scheduling, legal, research, customer support, code, finance, or "please ask a human before touching anything." bounded agents do bounded work. handoffs are visible. uncertainty is visible too.
that last part matters. the fantasy butler hides delegation behind a smooth voice. a good dispatcher shows its calls. it tells you which specialist answered, what assumptions they used, and where they stopped.
nature does build generalists sometimes, if the niche rewards it. humans are annoyingly good proof. but even we outsource constantly. teeth to dentists. taxes to accountants. dinner to whoever can be bothered. civilization is mostly specialization with receipts.
AI may not replace that pattern. it may finally make the routing cheap enough to put everywhere.