Short answer: An AI caller listens to the menu the way you do, works out which option matches the goal, and either presses the key (the phone sends a keypad tone) or says the answer out loud for menus that understand speech. Then it waits through the hold music, listening for a real person, and starts the conversation when one picks up. The tricky parts are menus that change, menus that want an account number, and telling a recording from a human.
Phone menus, called IVRs, accept keypad tones, spoken answers or both.
The AI picks options by meaning, not by memorising a fixed path.
Waiting on hold is easy for an AI; spotting the moment a human answers is the hard part.
OsmO works phone menus and waits on hold, and a long hold costs the same 3 credits as a short call.
How it works
- Sign in with your mobile numberYou get 50 free credits. Nothing to pay.
- Add OsmO to your AI appChatGPT, Claude, Grok and other apps that take connectors.
- Ask it to callOsmO calls, talks, and tells you what was said.
The full guide
What a phone menu is
Most large businesses answer with an interactive voice response system, or IVR: the recording that says "for billing, press 1". It is a decision tree. Each answer sends you down one branch until you reach a department, a recorded message or a queue. Older menus only listen for keypad presses; newer ones also understand short spoken answers such as "billing" or "agent".
How an AI gets through one
- Listen. Speech recognition turns the menu recording into text as it plays.
- Match. The model compares each option with the goal. "Change a flight" matches "existing reservations", even if the words differ.
- Answer. It sends the keypad tone for the chosen number, or speaks the option if the menu asks for words.
- Fill in. When the menu asks for a booking reference or a postcode, it enters the details you gave it.
- Repeat. It keeps going level by level until it reaches a person or a queue.
Because it chooses by meaning, an AI copes with a reworded menu better than a script that hard-codes "press 2, then 4". It can still go wrong when two options sound alike, so a good caller listens to the whole list before choosing rather than jumping at the first match.
Waiting on hold
Hold is where people give up and where an AI shines, because waiting costs it nothing. The challenge is recognising what it hears. Music, silence, "your call is important to us" and a queue position are all recordings. A person says something new and expects a reply. The caller has to tell them apart quickly, because an agent who hears silence after "hello?" may hang up.
| Sound on the line | What the AI should do |
|---|---|
| Menu options | Pick the matching option |
| Music or a looped message | Keep waiting quietly |
| "Press 1 to get a call back" | Follow the instructions it was given about call-backs |
| "Hi, this is Dana, how can I help?" | Introduce the call and ask the question |
| Voicemail beep | Leave a message or end the call, as instructed |
Phone makers have built the waiting part into the phone itself. iPhone's Hold Assist and Pixel's Hold for Me alert you when a person returns, so you can step away. They stop there: you still do the talking. An AI caller carries on into the conversation.
If you make the call yourself, the free hold tools on iPhone and Pixel are worth turning on. Hand the call to an AI when you also want it to do the talking once someone picks up.
The hand-off to a person
When a person answers, the caller explains why it is calling and asks the question. Agents often want to confirm who you are: a name, a booking reference, a date of birth. If the business needs something the AI was not given, or something it should not share, such as a card number or a password, the right move is to say so and report back, not to guess.
Where it still goes wrong
- Menus that only accept a spoken account number read in a particular rhythm.
- Security checks that want you, in your own voice.
- Very long holds, where the line drops before anyone answers.
- Businesses that close the queue at a set time and play a closed message.
How to help an AI get through
You can make a menu call far more likely to succeed with a few lines in your request:
- Give the reference numbers up front: an account, a booking or a policy number the menu may ask for.
- Say which department you want, in your own words: "baggage claims", "billing", "new patients".
- Say what to do with a call-back offer. Some lines offer to ring back instead of holding; decide in advance whether that is fine.
- Say what counts as done. "Get a claim reference" is clearer than "sort out my bag".
- Pick the time. Many queues are shortest soon after opening, and some lines close earlier than the shop.
If a menu insists on something only you can provide, such as a voice check, a good caller stops and tells you what it found so far, so your own attempt starts further along.
Trying it with OsmO
OsmO works phone menus and waits on hold for you from ChatGPT, Claude, Grok and other AI apps that take connectors. Include in your request the details a menu might ask for, such as a booking reference. OsmO For my AI starts with 50 free credits when you sign in with your mobile number, with no card. An answered call uses 3 credits however long it lasts, and a call nobody answers is free, so the free credits cover about 16 answered calls. When they run out you can keep calling from the OsmO app.
For hold queues on your own calls, see AI that waits on hold for you.
Questions people ask
What is DTMF?
Dual-tone multi-frequency signalling: the tones your phone sends when you press a key during a call. Phone menus read them to know which option you chose.
Can AI get through "press 0 for an operator" menus?
Yes, when the menu offers that path. Some menus hide or disable it, and then the AI has to follow the menu like anyone else.
Does a long hold cost more with OsmO?
No. An answered call uses 3 credits however long it lasts, and a call nobody answers is free.
Sources
Checked November 13, 2026. Vendors change their products often; follow the links for the current version.