It presses 3, then 1, then waits on hold.
Navigating phone menus means the AI listens to an automated switchboard as it speaks, decides which option matches what it is trying to do, presses the key, and repeats until it reaches a person or completes the task. It handles being transferred and being put on hold, and a human can watch, steer or take over at any point.
Calls need a card on file. US voice is about 2.6¢ per talk minute, billed in whole minutes — and four of those minutes spent in a menu is four minutes nobody on your team spent.
- Benefits and eligibility
- Member ID entered
- Retry recognised, re-entered
- "One moment, transferring you to a representative."
- Hold music
- "Benefits, this is Marta — how can I help?"
Or just ask for it
Send it into the switchboard. Go and do something else.
It presses the keys and survives the hold — you get the answer.
InboxReached a person — 4 menus, 11m hold
Ring the supplier’s main line and find out when order TR-77812 ships. Don’t settle for the automated line.
Four menu levels, one transfer, 11 minutes on hold, then a person. Ships Thursday. The whole navigation is in the transcript, key presses included.
@Octo chase the carrier about the missing pallet, their line is a nightmare
On it 🔧(runningstart_voice_call…)
In the menus now. I’ll post here when I have a human, or if I hit the wait ceiling. (edited)
🙏 23 replies · last reply 1m ago
this is the best thing we bought all year
Email octo@agent.omniocto.com · WhatsApp 1-MAN-ASK-OCTO · Slack #omniocto
How it gets through
Four things a phone menu does, and what it does back.
A switchboard is not one problem. It is a sequence of them, and most attempts fail on the third or fourth rather than the first.
The menu is read out
It is evaluated while it is still being spoken, in fragments, instead of waiting for silence and then reasoning about a transcript. Menus are long and the useful option is often first.
An option has to be chosen
The option is matched against what this specific call is trying to do, and the key is pressed as a real tone. Menus that ask you to repeat yourself are recognised and handled rather than looping.
"Transferring you now"
Treated as its own state, with a deliberate pause afterwards — switchboards routinely say this and then read another menu. Committing to hold on the first mention is how these calls get stuck.
Hold music, for minutes
Survived rather than mistaken for a dropped call, with a bounded ceiling on the wait so a stuck call ends instead of occupying a line indefinitely.
Almost no competitor can do all four, which is why this page exists — and why what follows is shown rather than asserted.
How we know it generalises
Tested against menus that disagree with each other.
Working on one switchboard proves very little — every company builds its menu differently. So alongside live calls it is run against trees with genuinely different shapes, and it reaches the right endpoint on each.
A seven-step tree
asks for a social security number partway through
A six-step tree
asks for a date of birth instead
A five-step tree
asks for no identifier at all
Different identifiers, different depths, different orders — all navigated correctly in the same run, alongside pharmacy and rebooking scenarios. That spread is the point: a navigator tuned to one tree is not a navigator.
Recorded walkthrough
You are not locked out
A human can step into any call, at any point.
It can stop and ask you
When the menu wants something the agent does not have, it raises the question on your dashboard while the call is still live. You type the answer and it goes straight into the conversation.
ask-a-human mid-call · answer injected liveYou can send it an instruction
Watching it take a wrong turn, you can tell it what to do next without hanging up and starting again. The instruction lands in the live context.
instruction injection · no restartYou can listen, or take the call
Join the call to hear what is happening, mute, or take it over entirely when a person finally answers and the conversation stops being mechanical.
join · listen · mute · hang upQuestions
The things people actually ask.
- How does it decide which option to press?
- It evaluates the menu as it is being spoken, chunk by chunk, rather than waiting for the whole recording to finish and then guessing. Each fragment of the menu is weighed against what the call is trying to achieve, and the key is pressed as soon as the right option has been heard.
- What happens when the menu says "transferring you now"?
- Transfer is detected as its own state rather than treated as more menu. There is a deliberate grace period afterwards, because switchboards very often say "transferring" and then immediately read out another menu — committing to hold too early is the classic way this goes wrong.
- Will it sit on hold?
- Yes, and there is a ceiling on it. Hold is survived rather than treated as a dead call, with a bounded wait before the attempt is abandoned, so a call cannot silently occupy a line forever.
- Does this work on real phone systems, or only in a demo?
- Real ones. It has been run against live phone trees, and it is also exercised continuously in the call simulator against menus built to disagree with each other — different depths, different identifiers, different orders. The simulator is how we know it generalises rather than working on one switchboard.
- Can I watch, or step in?
- Both. You can listen to a call in progress, send the agent an instruction mid-call, answer a question it raises, or take the call yourself. That is the same live-steering machinery used on ordinary calls.
- Why does this matter?
- Because the calls worth automating are almost never answered by a person. Verifying insurance benefits, checking a prescription price, chasing a rebooking — they all begin with several minutes of menu and hold before a human says hello. That is the part worth handing over.
Keep reading
Related jobs you can hand over.
Calls need a card on file. US voice is about 2.6¢ per talk minute, billed in whole minutes — and four of those minutes spent in a menu is four minutes nobody on your team spent.