AliceHQ R&D
Alice started as a passion project. It turned into Receipted AI™.
We set out to build a voice agent we could actually predict. What we ended up building was the thing that makes any AI agent safe to let act.
Why we started
Voice AI sounds great on a demo call. The problem is what happens after: it books the wrong slot, promises something nobody agreed to, or sends a message it shouldn't. It sounds right and does the wrong thing, and nobody finds out until a customer rings back angry.
We wanted a receptionist we could trust with the boring stuff. So we started testing it like we test hardware: decide what a good result looks like first, then make it prove it.
What we found
The language model on its own is never going to be fully predictable. So we stopped trying to make the talking part perfect, and gated everything it does instead.
Receipted AI™
Before an agent does anything that matters, like booking an appointment or sending a message, the action goes to the gate. Payments will go through the same gate once that part is built. The gate allows it, blocks it, or holds it for a person, and issues a receipt. No receipt, no action.
On top of that sit test gates. You define the expected outputs up front, and every result is checked deterministically against them before anything gets called done.
The conversation can still be flexible. What the agent actually does is gated every time.
Where it's going
Done
Voice agent R&D, and the Receipted AI™ gate that came out of it.
Now
The gate is the core of AcesOps, our business operations system. It already holds AcesOps' own build agents. Voice R&D carries on inside AcesOps.
Next
On-demand voice agents built by AcesOps, with every action gated and every result tested.
Later
All of it running on hardware here in New Zealand.