An AI Assistant That I Actually Want

Siri and Alexa can't do the thing I actually want most days, order food while I'm driving, switch on the AC at home before I've pulled into the driveway. Alexa sets timers. Siri texts people, if I already know exactly who and what to say. I tweeted about this a month before I sat down to write it out.

One caveat first, this is about being alone in the car. I'm fine talking to a voice assistant by myself. I'm not fine doing it in front of other people. That's a personal bias, not a universal one, but it's exactly why general-purpose voice apps don't work for me. They're built for me to talk to my phone anywhere. I won't.

The core driving problem is mostly solved already. Calls, messages, music, podcasts, navigation, all hands-free. That covers what I need with my foot on the pedal and my eyes on the road.

the idle time nobody's using

What's left is the other half of a drive. Stopped at a light. Stuck in traffic. Waiting in a parking lot. That's idle time, not driving time, and none of the big voice assistants treat it as a separate product.

Idle time is a different use case from the one Alexa and Siri were built for, and neither has noticed. Driving time has a hard constraint, hands on the wheel, eyes on the road. Idle time doesn't. I could type, I could tap, I could scroll, I'm just choosing not to because my hands are free anyway and talking is faster. Order food, order groceries, set the AC at home, check a flight, check a portfolio, none of it is voice-shaped today. It's still open the app, tap through it, screen by screen, the exact thing voice was supposed to replace.

what alexa and siri actually are

Neither can act inside another app, and that's a distribution problem, not a model one. Voice-controlled search with a few party tricks bolted on, weather, timers, music, a text if the contact's unambiguous.

Alexa Skills and Siri Shortcuts both ask third-party apps to build the integration themselves. A food delivery app has no reason to build a deep Alexa skill when the order still has to route through its own app to make any money off the customer relationship. Amazon and Apple don't own that relationship, so the incentive to integrate deeply isn't there on either side. What's left is a search bar dressed up as an assistant, because search is the only job neither company needs anyone else's cooperation to do.

what I actually want

Real access to every app on my phone. Real memory of my patterns. Real authority to transact. Not a narrator describing what it would do if I opened the app myself. Say "order my usual" while I'm driving, and it places the order, instead of reading me a menu. Say "turn on the AC," and it reaches into the home automation app and does it, instead of failing to find the skill and making me pull over to do it myself.

That's not one command and done. Ordering food needs back and forth on what goes in the cart, same as ordering from a person, not a single instruction I have to get exactly right the first time. Then a separate back and forth to pay, consent given, an OTP read out if that's what authorizing the transaction takes. After the order's placed, I still want to ask it things. When's it arriving. Tell the delivery person to leave it at the door. One continuous conversation about one job, not five separate trips into five separate apps.

Everyone's optimizing these assistants for how natural the conversation sounds. That's optimizing the wrong metric. I don't need small talk, I need the job done, start to finish. The real measure of a good assistant isn't how smart the conversation sounds, it's how few steps happen between me saying something out loud and that thing actually getting done. Call it time-to-completion, not conversational quality.

why this doesn't exist yet

This isn't a technology problem anymore. The models are already good enough to understand what I'm asking for. The real problem is liability. Who's responsible when it gets something wrong. If an AI plans a fake trip just to practice, nothing bad happens, no real flight got booked, no real money moved. If it actually books my flight and charges my card and gets it wrong, that's real money and a real mess to clean up. Nobody's built the insurance or warranty layer for that yet, and no one's handing an assistant broad transactional authority until someone has.

There's a second problem, and it's a business model problem, not a technical one. Apple and Google control how much access any assistant gets to the other apps on a phone. Real transactional access to third-party apps means real commerce, and real commerce routed through someone else's assistant is commerce that never touches Apple's or Google's own payment rails or their own take rate. Handing a competitor's assistant that much control over the phone doesn't just make Siri or Google Assistant matter less, it moves the transaction, and the revenue, off their books. Neither company is opening that door for a third party before it's built the equivalent itself.

conclusion

Not someone promising to run my whole life on day one. Someone who picks one bounded, recoverable transaction, one specific reorder, one specific bill, and builds the trust ladder around just that. Propose it, let me approve it a few times, earn full autonomy only once there's a track record. Widen the scope one transaction at a time, at the same pace as the trust.

Trust should scale with what's being asked, not renegotiate from zero on every transaction. That's the model an AI assistant needs for autonomy, not a single all-or-nothing switch.

Model quality won't be the moat here, every serious player will have access to a model good enough for this job. The moat is the trust ladder itself, how fast a company can move a customer from approve-every-step to approve-occasionally to don't-ask-me-again, without one bad transaction resetting the clock. That's slow to build and hard to copy, which is exactly why it's a real business and not just a feature. That's the assistant I'd actually hand my card to. That's the founder I'd actually write a check to.