Run it from chat

Turn voice notes into tasks an AI agent can act on

To turn voice notes into tasks reliably, say who it is about, what you want, by when and any limits, in under a minute — and have the agent reply with a short written brief and its questions before it does anything outward-facing.

person holding notepad and pen flat lay photography
Photo by Marissa Grootes on Unsplash

Part of our guide to Run your business from chat: a practical guide to chat-first operations

On this page · 8 sections
  1. How to turn voice notes into tasks: the four-part note
  2. What the agent should send back before it acts
  3. From a site visit: photo plus voice
  4. On the road: record before you drive
  5. For practice notes and client admin
  6. Common voice note mistakes and fixes
  7. When to type instead of talk
  8. Frequently asked questions

Voice notes are the fastest way for a busy owner to hand work to an AI agent: you think out loud for 40 seconds and the agent does the typing, the looking-up and the drafting. They are also the easiest way to give a vague instruction. The fix is a small habit on each side. You structure the note — who, what, when, limits — and the agent sends back a written brief with any questions before it acts.

This is one of the practical habits behind running your business from chat. Cortex accepts voice notes, photos, documents and phone videos as instructions in Telegram, WhatsApp, Slack or Discord; the method below works in any of them.

How to turn voice notes into tasks: the four-part note

You don’t need a script, just an order. Say these four things, roughly in this sequence, and the agent has almost everything it needs:

  1. Who or what it’s about. A name, an address, a job number: “The Hendersons at 22 Wattle Rd.”
  2. What you want done. One outcome per note: “Send them a quote for two downlights and a new circuit.”
  3. By when. “Tonight” or “before Friday.” Without a time, the agent has to guess priority.
  4. Limits and specifics. “Same rate as last time. Don’t book a date yet — they’re checking with their builder.”

One outcome per note is the rule that matters most. A four-minute ramble covering six jobs is hard for anyone to turn into clean tasks, human or AI. If you have six things, send six short notes, or start with “Three things” and number them as you go.

What the agent should send back before it acts

The agent’s reply is the safety net. Speech is ambiguous — names get misheard, “Thursday” could be this week or next — so the agent should restate the task in writing and ask about anything uncertain. Ask your agent to reply in this shape:

Three things make that reply useful. It shows the figures it pulled from your records, so you can catch a wrong assumption. It says what it won’t do. And it asks one specific question rather than “anything else?”. If the reply is right, “yes, 5 hours” is all you need to send back.

Anything outward-facing still goes through approval after the brief is confirmed. Confirming the brief is not approving the message; the customer email still arrives in the thread for your OK. That distinction is covered in approval gates for AI agents.

From a site visit: photo plus voice

On site, a photo and a voice note together carry more context than either alone. Take the photo, then send a note straight after: “That’s the fence on the east boundary at 8 Ash Ave. Two palings missing, post leaning. Add to the inspection report as tenant damage unless the entry report shows it already. Get a quote from the usual handyman.”

Notice the conditional — “unless the entry report shows it already”. Good agents can check that against the record before drafting anything, which is exactly the kind of tedious cross-checking you want off your plate. Short phone videos work the same way. Tony Lorenzo at Elite Partners Real Estate says his agent turns his phone walk-through videos into NSW ingoing condition reports; we cover that workflow in condition reports from walk-through videos.

On the road: record before you drive

The drive between jobs is when many owners remember what needs doing. Be careful here: mobile phone rules for drivers are strict in every Australian state and territory, and they differ. Record your notes before you set off or once you’ve parked, or use only the hands-free methods your state’s rules allow. No task is worth a fine or a crash.

A good pattern is the “car park debrief”: before leaving a site, sit in the car for two minutes and send one note per thing that needs doing. The agent turns them into tasks while you drive, and the written briefs are waiting when you next stop.

For practice notes and client admin

Sole practitioners use voice notes to capture the admin around a session while it’s fresh: what was delivered, how long it took, what to book next. Irene Gyane of Fourth Woman Therapy describes sending quick voice notes or photos from her phone that Cortex turns into invoices with linked progress notes, bookings, client files and emails, and says she sees the exact draft before anything goes out.

Clinical and NDIS notes carry sensitive information, so the rules tighten: keep the content of notes in the client file rather than in a shared thread, use a private chat with the agent rather than a group, and check every draft. The detailed workflow, including what a progress note should contain, is in progress notes from voice notes.

Common voice note mistakes and fixes

Mistakes we see, and the fix
MistakeWhat goes wrongFix
Several jobs in one long noteTasks get merged or one is missedOne outcome per note, or number them aloud
“Him”, “that place”, “the usual”The agent picks the wrong person or supplierSay the name or address once at the start
No deadlineUrgent work waits behind routine workAlways say when
Spelling names or numbers by voiceMisheard emails, phone numbers and amountsType those, or send a photo of the card or invoice
Background noiseTranscription errorsStep away from machinery or traffic for 30 seconds
Approving the brief and assuming it’s sentConfusion about what went outTreat brief confirmation and message approval as separate steps

When to type instead of talk

Voice isn’t always better. Type when exact characters matter — email addresses, account numbers, legal wording — and when you are somewhere people can overhear sensitive details. And for anything you will need to quote back precisely later, such as a promise to a customer, a typed instruction leaves a cleaner record than a transcript.

Voice notes work well in both Telegram and WhatsApp; our set-up guides cover file sizes and privacy for each. The channels page shows how Cortex handles voice, photos and video in each app.

Frequently asked questions

Can AI turn a voice note into a to-do list?

Yes. An AI agent can transcribe a voice note, split it into tasks, look up the people and records it mentions, and draft the work. It works best when each note covers one outcome and the agent confirms its understanding in writing before acting.

How long should a voice note to an AI assistant be?

Usually under a minute. Say who it’s about, what you want, by when and any limits. If you have several things to hand over, send several short notes rather than one long one.

Is it safe to send client details in voice notes?

Keep sensitive details to a private chat with the agent and store them in the client file, not group threads. Avoid recording health, financial or identity details where others can hear, and check every draft that uses them.

What if the AI mishears my voice note?

That is why the agent should send back a written brief before acting. Check names, amounts and dates in the brief, and type anything that must be exact, such as email addresses or account numbers.