context, quickly: I've been building a thing where a technician takes a photo, says what he found, and a standardized report writes itself before he leaves the site.
this week I connected it to ElevenLabs' agent API. so now it actually holds a conversation – it asks, you answer, it asks the next thing, you can talk over it. no buttons, no tapping.
you'd think the voice was the hard part. it wasn't. it took an evening.
the hard part was the opposite of talking.
an agent is built to be helpful, and helpful means it will happily fill in "alvorlighedsgrad: kritisk" because the sentence sounded serious. nobody said it. it inferred it. and that is the one thing a report must never do – because a human signs that document, and a plausible guess is worse than an empty box.
so now it can do exactly two things: ask about a field that's still empty, and read a part number back to check it heard right. it cannot propose a value. ever. and everything it claims still goes through my own check afterwards, where anything you didn't actually say gets emptied out again.
that took three times longer than making it speak.
it's not perfect - but it works! ⚙️ - AI voice is fun to work with..
what's the document in your week that you fill in afterwards, from memory? (potential solution is right here...)
– could you imagine this approach for your case? your experience, multiplied 💛