A voice-agent platform with a visible cost ledger

Why "the engine, not a prompt" matters

Most voice-agent platforms hand a language model one long, growing instruction and let it decide every turn. That single architectural choice is the common root of the problems teams report after a pilot: inconsistent behaviour on the same input, invented numbers, a pause before every reply, and a bot that quietly slides back into English mid-call.

VARTA is built the other way. Cheap, deterministic logic decides first; the language model is called only when nothing else can resolve the turn, and even then it can only choose from options the workflow itself defines -- it never composes a free-form answer.

What that buys you

• No invented prices, dates, or IDs. Logic, not the language model, owns anything that goes into a system of record. When the model does run, it picks from a shortlist the workflow defines; a low-confidence pick falls through to a re-ask instead of a guess.
• Human-speed responses. Most turns are answered without a model call at all, and cached audio plays back instead of being resynthesised live.
• A cost ledger per call, not a blended rate. Every model call, every synthesised character and every second of silence is priced and attributed, so you can see what a specific call cost -- not an average across your whole account.
• Caching that survives personalisation. A caller's name, order number and amount don't have to break the cache the way naive hash-the-sentence caching does -- see how the caching layer works on the platform page.
• Every language your callers actually speak, including how they mix them. The engine itself contains no language-specific text, so there is no English default for it to fall back to under pressure.
• Callers can interrupt, backtrack, and change their mind, and the call flow keeps track of where it needs to return to.
• Empathy your brand can govern. What the agent can, cannot and must never say to an upset caller is written down as policy, not left to a prompt's good judgment.
• A workflow a non-engineer can author, described in plain English, with an AI co-author drafting the flow, the rules and the language variants -- and an engineer can still read and review every step, because the flow is a structured graph, not a prompt.
• A decision trace, not just a transcript. Every turn records which layer of the engine answered and why, on real calls as well as test calls.
• A regression gate for conversation quality, the same way a codebase has one for correctness: before-and-after runs, not "we changed the prompt and it sounded fine on a few calls."

Live today, not a lab demo

humanise.ai runs as a live, multi-tenant SaaS product on this exact engine, handling real calls in production. Read more about how the platform is put together, or see what running it inside your own environment looks like.