On-device voice agents, built in Helsinki
Voice AI that can explain itself.
Better Labs voice agents run on the device. Decision models and deterministic code make every decision; the language model only puts it into words.
- Runs on
- The device
- Decides with
- Rules and models
- Company
- Helsinki, Finland
Demo video coming soon
Hear an agent take a real call.
Every reply shows the rule behind it.
Step through an example call. Each answer comes from a named rule or model step, and you can read why it fired.
- Caller
Hi, I need to change the address on my order.
- Agent
I can help with that. What is your order number?
model: intent → change_address - Caller
It’s BX-4471, under Laine.
- Agent
Thanks, that matches. What is the new address?
rule: verify_identity → pass - Caller
Mannerheimintie, in Helsinki.
- Agent
Your order has already shipped, so I can’t change it here. I can text you the carrier’s redirect link.
rule: order_shipped → offer_redirect - Caller
Yes, please.
- Agent
Done. The link is on its way.
action: send_redirect_link → ok
Decision trace
- Heard
- “change the address on my order”
- Decided by
intent model → change_address- Checked
- The intent is on the allowed list for this line.
- Language model was asked to
- Confirm the request and ask for the order number.
How a reply is made, on the device.
Speech comes in, rules and models choose the next step, and only then does a language model write the sentence.
On the device
- 01
Hears the caller
Speech recognition turns the caller’s words into text.
audio → “change my address”
- 02
Decides what to do
Decision models and deterministic code choose what happens next.
rule: order_shipped → offer_redirect
- 03
Words the reply
A language model writes the sentence for the decision. It can’t change the decision.
offer_redirect → one sentence
- 04
Speaks it back
The reply is spoken back to the caller.
text → audio
The same input always reaches the same decision, because step 02 is code.
What on-device decisions give your company.
Audio stays on the device
Speech becomes text on the device. Recordings don’t go to a cloud model.
Rules an auditor can read
Each decision points to a named rule you can test and show.
Agents the language model can’t overrule
The language model can’t promise a refund or skip a check.
Agents that cost less to run
No hosted-model round trip per turn, so no network delay or per-call cloud fees.
rule order_shipped when order.status == "shipped" intent == "change_address" decide offer_redirect say explain_locked, offer_carrier_link never change_address, promise_date
The trade-off: we write your process down as rules before launch, together with your team.