in conversation with ai
What happens when you work with AI by voice?
By Sonia Nash.
I have to admit, talking to AI did not come naturally to me despite the fact that I spend a large portion of my day interacting with AI tools. I believe that subconsciously, I was compartmentalising my interactions: keyboard for software/AI, voice for people. I was aware that most AI assistants let you speak directly to them, but I didn’t see the point. Until I caught myself thinking out loud, then afterwards typing what I was thinking. And so I wondered what would happen if I was thinking out loud directly to my AI assistants. Would I get better results or just longer, messier ones in line with my messy thinking? Is there a productivity increase related to voice use? And what is the impact on wellbeing and loneliness, particularly for solo entrepreneurs / remote workers who may not have a colleague to talk to. And so, I started speaking to my tools.
The difference between dictation and voice mode
These 2 voice elements are not the same. Both let you speak freely with your Ai assistant/agent, but dictation only allows YOU to speak, which means that the response from Claude or ChatGPT will be written. In voice mode, it listens and talks back so this feels like a real conversation. And it is pretty good despite the questionable voice names. In Claude, you can select between 5 voices in your settings: Buttery, Airy, Mellow, Glassy and Rounded. I’ve asked Claude to explain why the voices had such bizarre names and it couldn’t explain either. I settled for Mellow and it is pretty good.
So, verdict
The first observation is that it is indeed faster. There aren’t enough studies and evidence out there to state a concrete number, but I reckon based on my personal experience that it is 2 to 3 times faster to vocalise my idea/command to AI than it is to type it, and I include the thinking time here where I use to start typing, pause to think about what next, then type it. And I’m not even including my personal OCD with spelling; even if AI tools are incredibly efficient at understanding what you mean despite typos, I would always correct my typos before pressing enter. Inefficient? Yes. But as someone who speaks 4 languages and who is deeply in love with the written word, I refuse to compromise on grammar. Similarly, my text messages are full words only. Call me old fashioned (or just old), but this is something I do not want to lose. So if we took this into account as well, perhaps my productivity gain by using voice is even higher than stated.
And this has opened new ways of working for me. As someone who has to spend an inordinate amount of time in the car between client sites, I can now offload meeting notes straight after the meeting, impressions, things to remember, I can draft emails and follow-up as I go instead of waiting to get home. I also do that on the occasional walk to bring together fresh air and work.
Do we say more when we speak?
I found myself providing more information and more context behind my commands because it was so easy and effortless. But the question isn’t whether we say more or less, but whether what we say makes more sense or provides more useful context for the task. As someone with a tendency to waffle, writing is the perfect platform for me as I have the time to organise my thoughts into a concise block with some sort of structure. My verbal prompts were initially all over the shop, a lot longer with multiple asks buried within several layers of context bits. I resisted the urge to “rework” my own prompt as I wanted to see how well the tools performed in this scenario. And results vary; sometimes the tool didn’t understand the ask, and this is a prime example of “user error”. If your prompt is a giant mess, you will not get the best output; sometimes the tool will emphasise a part that you didn’t think was that important, but your waffling made it look like it was important.
Where voice vs text made sense for me
Building an asset: iterative feedback on a page or design is ideal for voice. You react the way you would with a colleague or an intern, they mention or build something, you react. It’s conversational ping-pong like you would do with a human.
Summarising documents: giving instructions as to what you want doing on that document. It’s usually quite specific and waffle-free, ideal for voice.
Ideation: this is one I have debated for a while. If you have a beginning of an idea but you’re not too sure where you’re going to go with it, voice could be a good way to start. You could offload your half-formed thoughts, the context and let AI do the structuring for you. I found that writing is best suited to the next stage following on from that, once you are narrowed down and you’re looking at comparing options. But again, I feel this is very subjective.
Where voice falls short
Precision work: obvious but worth stating. Have you tried to dictate or listen to a line of code? Code, formulas, exact figures, exact file paths, a misheard number could screw up your entire analysis or the piece of work you are delivering to a client. Not worth the risk.
Public environment: again an obvious one but you don’t want to be blurting out commands and important information in an open-plan office or worse, in public places. GDPR and all that.
Listening back: hearing a long AI answer read aloud is slower than reading it, and while the voice is fairly decent, it is still discernibly non-human. It does sound like droning on at times and I feel I switch off easier versus reading an answer.
The interesting aspect: breaking the silent workday
There’s an entire strata of workers who can spend most of the day without speaking to a soul: remote-workers, lone workers, field engineers between jobs. A July 2025 study by the University of York found that people in jobs that moved to remote work after 2020 reported more loneliness, especially women, and worse mental health. And this is something I can relate to; building my own business has meant many days working in isolation and entire days without uttering a word. So does talking to AI help? I will first answer based on my own observations; I feel it has made a positive difference to days where I have no client/partner meetings. It made solitary working genuinely less solitary. However, it is no substitute for the real thing aka talking to human.
There are several studies out there for MIT, OpenAI and Harvard Business School on both the positive and negative impacts of voice on loneliness. Research suggests that at low to moderate use, voice did better than text at reducing loneliness and emotional dependance. However, high use went hand in hand with more loneliness, more dependence and less socialising with real people. Worth noting that these studies looked at chatbots and companion apps, nobody has yet studied lone workers using voice AI for everyday work tasks.
How to get good at using voice AI
Start with one habit. Pick one recurring task, such as meeting debriefs, and do it by voice for two weeks.
Speak like a briefing, not a chat. The research suggests a structured way of speaking works better than a ramble. Use a simple shape: context, goal, constraints, ask.
Finish with a clear ask. "So what I need from you is…" is the most useful sentence in voice prompting. It stops your request getting lost in the context.
Ask for a read-back. "Before you start, summarise what you think I'm asking for." If you have quite a lengthy prompt, this is quite useful at catching misunderstandings in seconds. I’ve only just started doing this and I wish I had known sooner.
Don't over-compress. If a dictation tool tidies your words into a short summary, check it hasn't stripped out what mattered.
Spell out the tricky bits. Names, acronyms and numbers: say them clearly and check the transcript.
Mix modes. Speak to brief and give feedback, read to review, type to fine-tune. Don’t be afraid to mix your interaction modes, if writing feels right for that specific task, use it even if you started with voice and vice-versa.
Check your privacy settings. Most AI tools delete your audio once it has been turned into text and they state that your voice isn’t used to train models. Worth checking that this toggle is indeed off. It won’t change anything to the input or output, but you’ll have more peace of mind.
Used well, it’s one of the simplest ways to get more from the AI tools you already have. So try it out this week. Then use the time you have saved to call someone.