How AMI actually works
The whole pipeline, where each part runs, and what it cannot do. If you are going to rely on this during an interview, you should know how it behaves.
The pipeline, step by step
Four stages. Two happen entirely on your computer.
-
Capturing the conversation
AMI captures your system audio output — what the interviewer's voice is coming out of — plus your microphone. It does not join the call, install a bot, or appear as a participant. As far as the meeting software is concerned, nothing extra is there.
Runs on your computer -
Turning speech into text
A speech-recognition model ships inside the app and runs on your own CPU. It buffers speech, waits for a natural pause, then transcribes that chunk and decides whether it looked like a question worth answering.
Runs on your computer — the audio is never uploaded -
Writing the answer
The question text, along with the background you gave AMI and your chosen tone, is sent to our server. The server holds the AI credentials — that is why you never need an API key — and streams an answer back token by token.
Runs on our server — question text only -
Putting it on screen
The answer appears on a compact overlay that is excluded from screen capture, kept out of the taskbar, and kept out of alt-tab. It streams in as it is written, so you can start speaking before it finishes.
Runs on your computer
The answers
Written for you, not generic
- Your background. You tell AMI about yourself once, and answers are written in the first person from that.
- Real context. Because it follows the whole conversation, an answer can build on what was already said instead of restarting.
- Your tone. Formal or conversational, concise or detailed — switchable per session.
- Points, not a script. Answers come as a few specific points you can speak from, rather than a paragraph to read aloud, which sounds like reading aloud.
Speed and models
- Streaming. Text appears as it is generated, so you can begin talking after the first point.
- Trial model. The free trial uses a smaller, faster model.
- Paid model. Every paid plan uses a 70B-class model — better on unusual questions and at holding context.
Where your data goes
Split by what leaves your computer and what does not.
| Item | Leaves your computer? | Notes |
|---|---|---|
| Interview audio | No | Transcribed locally by a model bundled in the app |
| Question text | Yes | Sent so an answer can be generated |
| Your background / profile | Yes | Sent with the question so the answer is about you |
| Display name & language | Yes | Stored with your device record |
| Trial time used | Yes | Counted server-side so it is consistent |
| Email address | Yes | Only if you create an account |
| Card details | No | Handled entirely by Paddle; they never reach us |
What it cannot do
Stated up front, so nothing here is a surprise during an interview.
It needs an internet connection
Transcription is local, but answers are generated on our servers. With no connection, AMI listens and transcribes and cannot reply. There is no offline answer mode today.
Windows only
Windows 10 and 11, 64-bit. No macOS, no Linux, no web version. The screen-capture exclusion it depends on is a Windows feature, so a port is not a recompile.
It is not a lie detector, and it does not know the company
Answers are built from the background you provide and the conversation. AMI does not research the employer, and it can be confidently wrong about facts. Read what it suggests before you say it.
Audio capture depends on your setup
AMI captures your system output and your microphone. Unusual routing — virtual audio devices, some headsets, per-application volume set to silent — can mean one side is not picked up. Test a session before it matters.
Whether you are allowed to use it is on you
Many employers and assessment platforms prohibit assistance. AMI does not check the rules of your interview, and it cannot. That call, and its consequences, are yours.