How to Set Up a Voice-First AI Workflow Before OpenAI's Hardware Devices Ship

OpenAI’s president just confirmed the thing we’ve all been guessing at: real hardware, a whole family of it, designed for talking to AI instead of typing at it.

Greg Brockman sat down with Joanna Stern and said OpenAI is “building a family of devices” for its chatbots. No release date, no specs, no confirmation of the smart speaker rumors — just “you can expect them soon.” I wrote about OpenAI’s earlier hardware moves when the Codex Micro macro pad was announced, and that post holds up. But this interview is a bigger deal, because Brockman said something specific about how we’ll use these devices.

His prediction: we’ll talk to our computers rather than type for “the vast majority of what we do.”

That’s not a side comment. That’s OpenAI telling you what their hardware is for before it exists. And here’s the part almost everyone missed — you don’t have to wait for the hardware to benefit. Voice-first workflows already work today, on tools you already own. I spent an afternoon rebuilding my daily AI tasks around voice input to find out what actually works. Some of it was great. Some of it fell apart in funny ways.

What I actually did

I picked my five most-repeated AI tasks and forced myself to do every one of them by voice for a full workday. No typing prompts. I used ChatGPT’s voice mode on my phone, voice dictation in my browser, and a custom GPT I’d built months ago and mostly ignored.

The tasks: drafting client emails, summarizing long articles, generating social captions, brainstorming blog ideas, and checking my schedule against a to-do list.

The idea came straight from Brockman’s framing. If the “vast majority” of AI interaction is heading toward speech, then the people who already think in voice commands will have a head start. I wanted to know if that head start is worth anything right now, or if it’s just early-adopter theater.

Short answer: it’s worth it for about half my tasks. The other half reminded me why keyboards survived this long.

What Brockman actually said

The interview covered more than devices, and a few details matter for anyone planning around this.

On the Apple lawsuit — the one filed over allegedly poached hardware engineers, which I covered when it was filing news — Brockman kept it diplomatic: “We are focused on our own development and technology… We have no interest in other companies’ trade secrets. We are plenty innovative.”

On timing: nothing confirmed. He didn’t confirm the rumored smart speaker for 2027, didn’t confirm a wearable, and gave no date beyond “soon.” If you want the full interview, it’s on Joanna Stern’s YouTube channel, and The Verge’s writeup is the cleanest summary I’ve found.

On privacy, which might be the sleeper story here: Brockman acknowledged that people don’t fully trust features like ChatGPT’s temporary chat mode. His answer was that OpenAI is working on “verifiable, auditable guarantees” — technical proof that only the AI can review your private data. No timeline on that either, but it tells you where their heads are.

And one detail that made me laugh: soundproof “beak” masks are starting to appear in offices for people who talk to their AI all day. Brockman doesn’t use one, but he said the masks “really show you where there’s product market fit.” A company announcing hardware because people are strapping bird beaks to their faces to talk to ChatGPT in peace — that’s a signal you can’t unsee.

Do this yourself

Here’s the setup I landed on after the experiments. Total time: about 20 minutes, all free.

Step 1: List your three most-repeated AI prompts. Open your chat history and count. Mine were email drafting, article summarizing, and content brainstorming. Write yours down — these become your voice workflow.

Step 2: Turn your top task into a custom GPT. If you’ve never built one, it’s genuinely a 10-minute job with no code — fill in instructions, paste an example of the output you want, save. Custom instructions do most of the work; my daily toolkit runs on two of these. Voice input into a well-configured custom GPT beats voice input into a blank chat every time.

Step 3: Wire up voice input on your phone. ChatGPT’s voice mode handles conversational tasks well. For phone-based dictation into any app, iPhone users can build on the Shortcuts approach I covered in the iOS voice workflow guide — one tap, speak, done.

Step 4: Run a one-day voice trial. Pick tomorrow. Do your top three tasks by voice only. Awkward at first, then it clicks — you’ll notice you phrase things differently when speaking, more context, less prompt-engineering gymnastics.

Step 5: Note where voice broke down. This list becomes your filter later. When OpenAI’s devices actually ship, you’ll know exactly which tasks belong on dedicated voice hardware and which ones you’ll keep typing forever.

What failed

Voice dictation mangles proper nouns relentlessly. Client names, product names, anything not in a dictionary — the AI guesses, sometimes hilariously, sometimes expensively. One client email went out with the wrong company name because I didn’t proof the dictated draft.

Long-form writing by voice drained me. Summarizing an article by voice worked great. Drafting a 1,000-word essay by voice did not — I’d lose the thread halfway through a paragraph. My voice worked fine; my brain’s outlining process apparently needs a cursor.

And the open-plan office problem is real. I work from home, so talking to my laptop is harmless. If you sit in a shared office, you already know why those beak masks exist. Brockman’s right that they signal demand — they also signal that nobody’s solved the social weirdness yet.

Who should skip this

If your AI work is mostly editing and refining existing text, keep typing. Precision editing by voice is slower and more error-prone, full stop.

If you work in a shared space and can’t use headphones comfortably, park this until quieter hardware shows up. That’s arguably what OpenAI’s devices are for.

And if you haven’t gotten value from AI in text form yet, voice won’t fix that. Start with what AI can actually do for you before optimizing the interface.

The bottom line

OpenAI building a family of devices confirms the direction: AI is becoming something you talk to. The tools for practicing that shift exist today, they’re free, and twenty minutes of setup puts you ahead of most people who’ll wait for hardware to force the change. If you want the fastest possible start with this stuff, start here and build your first workflow today.