Split-second decisions
The small decisions inside a phone call.
A voice agent makes several small decisions in every turn of a call. Handed to a small decision model, about 39 in 40 came out right, each in under a quarter of the time allowed.
The idea
A good receptionist decides a dozen things while you are still talking: are you finished, is this urgent, who should take it. None of it is deliberate. Split-second decisions gives a voice agent that reflex, and keeps the slow thinking for what to say.
Read the full story →Why it matters
A voice agent that asks a large model about everything is too slow to talk to. If the small decisions can be answered quickly and checked one by one, the time that is left goes to the reply.
Voice agentsTurn-takingDecision modelsResponse time
How it works
- In
- What the caller has said so far, and one closed question about it.
- Out
- A typed answer with a probability, in a fraction of a conversational pause.
Closed source · results onlyWhich models were used, how the questions are worded and how the pieces are wired are not published. The measured figures are in the full results paper, shared on request.
Explore deeper Each path is a different view of the same work.
Why it exists
On a call, the time to reply is a conversational pause. A voice agent has to decide what to do and start talking inside it.
Much of that deciding is small: has the caller finished, is that an interruption, does this need a person, is this action one that cannot be undone.
What it is judged by
Accuracy on labelled cases, one kind of decision at a time, averaged across the kinds. A time allowance for each decision, and the cost of a thousand of them.
Separately, the caller's wait from their last word to the agent's first, on a scripted call.
What it found
On the first run, with plain instructions and nothing tuned, about 39 decisions in 40 were right. A typical decision took under a quarter of the time allowed, and two kinds had no errors at all.
Checking a drafted sentence against a policy was the weakest, at about nine in ten. End to end the agent still answered in nearly three times the target, and the decisions were only about a tenth of that wait.
What it does not claim
The test is machine-made, machine-labelled clean text, and on the easy side. It is one run in one setting, before any tuning.
The end-to-end timing is from a scripted call, not from real callers.
What was measured
Each number is read from the file named under it. Nothing here is rounded up.
private research log: voice agent labprivate research log: voice agent labprivate research log: voice agent labprivate research log: voice agent labprivate research log: voice agent labprivate research log: voice agent labCommon questions
Where are the actual figures?
In the full results paper: the accuracy for every kind of decision, the timing of every stage and the cost. It is shared on request.
Which models were used?
A small decision model for the decisions and a language model for the reply. Which ones is not published.
If the decisions are fast, why is the call slow?
Because the decisions are about a tenth of the wait. The largest part is the reply's first word.
Is this tested on real calls?
No. The accuracy is on labelled text cases and the timing is from a scripted call.
Can this be done for my voice agent?
Ask through the request form and say what your callers ask for.
Ask a question
Ask about what it does and what it measured. Answers come from what is written on this site.
Carry thisGive the reflexes to something small and fast, and spend the pause on the reply.
proprietary · the result is public in relative terms · the figures on request