Eight days after ChatGPT shipped its new Voice Mode I posted that my work had left my desk. I listed the five things I now get done without touching a screen. Say out loud what I want while I am on a walk. Send it to the computer at home, which starts working. Ask what it has done so far and it tells me. Start three jobs at once instead of one. Change the instructions halfway through just by saying so. The post took 265 reactions and 109 comments, and almost every reply wanted the same thing spelled out. Not the walk. The setup behind it. This article is that setup, written for anyone who runs a go to market team and is tired of being the bottleneck in their own week.
I run Growth Cab, a GTM advisory. We put AI inside outbound, research and reporting for B2B teams selling contracts above fifty thousand dollars. That means I spend most of my week handing work to machines and checking what comes back. So the voice thing is not a gimmick for me. It is the difference between a day where I babysit three tools from a chair and a day where three jobs run while I walk to lunch.
in
“ChatGPT's new Voice Mode is 8 days old. My work left my desk. Here are the 5 things I now do without touching a screen.”
How to Delegate Tasks to AI Agents by Voice
The first thing to get straight is what the voice actually is. When I talk to Voice Mode I am not talking to the thing doing the work. The voice only carries my words to the agent running at home, the one with access to my files, my CRM and my scripts. The voice is a messenger. The agent is the worker. Keep those two apart in your head and the whole method makes sense. Blur them and you will keep asking the messenger to do the job, then wonder why it forgets everything the moment the call ends.
So the rule I follow is simple. Use the voice like an assistant you are briefing, and never like the expert you are hiring. An assistant gets a clear instruction, a deadline and a way to report back. That is exactly what a delegated task to an AI agent needs too. A goal it can check itself against, the context it needs to start, and a place to leave the result so I can read it when I sit back down. The moment I started treating the voice as an interface instead of a brain, the number of things I could hand off in a morning stopped being one.
The Setup That Runs While You Walk
Here is the shape of it. One machine at home stays on, with an agent that can reach the things I would reach at my desk. My documents, my CRM, the prospect research folder, the reporting scripts. The voice layer on my phone does one job, which is to take what I say and pass it over. When I say start the research on the three accounts we talked about Monday, the voice does not do the research. It wakes the agent, hands over the instruction, and the agent starts pulling. I keep walking.
The part people underrate is the report back. Delegation without a report is just hope. So every task I hand off has to answer one question when I ask it later. What have you done so far. If the agent cannot tell me that in a sentence, the task was too vague to hand over in the first place. This is the same discipline you would use with a new hire. If you cannot describe the finished thing clearly enough for someone else to check it, you are not ready to delegate it, and no amount of model quality will save you.
Once that loop works, the count climbs on its own. I used to start one job and wait for it. Now I start three, because starting a second one costs nothing while the first is running without me. Research on one account. A first draft of a reporting deck. A cleanup pass on a messy list. None of them need me in the chair. They need me for ten seconds at the start and two minutes at the end, and the two minutes in the middle are mine.
What Makes a Task Safe to Hand Over
Not everything survives being delegated by voice, and pretending otherwise is how people get burned. The tasks that work share one trait. You can look at the output and know in a minute whether it is right or wrong. Research on a company is checkable. A draft is checkable. A list built to a rule you named out loud is checkable. Those are safe to hand over because the review is fast and honest, and a fast review is what lets you run several jobs without drowning in them.
The tasks that fail are the ones where you cannot tell good from bad without redoing the work yourself. A pricing call. A reply to a client who is already annoyed. The final read on whether a deal is real. I have tried to push those onto an agent and every time I ended up doing them again anyway, so now the agent drafts and I decide. The judgement stays with me. The typing goes to the machine. That line is where the whole thing holds together.
Where Delegating to AI Agents Still Breaks
I want to be straight about the failure modes, because a walk in the sun makes this sound cleaner than it is. The voice mishears. On a loud street it will turn account into a count and the agent will start on the wrong job with total confidence. So I built a small habit of asking it to read the instruction back before it runs anything that matters. Ten seconds, and it has saved me from a full morning of an agent doing the wrong thing very thoroughly.
The second problem is that an agent left alone drifts. Give it a vague goal and it will hand back something that looks finished and is quietly useless. The fix is not more freedom. It is a tighter instruction and a shorter leash on the first run. I let a task run long only after I have watched it get the short version right. And I still read everything before it touches a real prospect. The setup removed the busywork from my week. It did not remove me from the loop, and the day I pretend it did is the day something embarrassing goes out under my name.
So the honest summary is this. Voice Mode did not make me smarter. It moved the boring half of my job off my desk and let me keep the half that needs a person. If you run a GTM team and you are the bottleneck, that trade is worth more than any single model release. Start with one task you can check in a minute, hand it over, and see whether the report back tells you the truth. If it does, you have your first agent. If it does not, you have found a task you did not understand well enough to delegate, which is worth knowing too.
I write up the systems we actually run at Growth Cab in The Revenue AI Brief, one play a day for people putting AI inside their go to market. The original version of this started as a LinkedIn post, and the comments there are worth reading if you want to see how other founders are running the same setup.
