Sign inBook a demo

Blog

Notes from the engineering team.

What we've learned shipping voice agents into production, written in long form with the math, the tradeoffs, and the things that did not work.

CompanyJuly 5, 20267 min read

Outbound AI that isn't a robocall: triggers, not judgment.

Proactive outreach only works if it is automatic: the lead who fills a form at 9 PM gets called in seconds, the no-show gets the reminder without anyone remembering to send it. So the system dials on its own. What it never does is decide on its own. Every call traces to a trigger the business configured, and every dial re-runs a fail-closed compliance gate stack. The line that separates outbound AI from robocalls is not automation. It is improvisation.

ProductJuly 4, 20265 min read

Dialect is not a locale code.

Setting language: 'ar' tells your voice agent which script to write, not what a caller in Amman actually sounds like, or how weird it feels when the reply comes back in newsreader Arabic. What it takes to build a voice agent that speaks the way its callers do, and why the same problem exists in English.

EngineeringJuly 3, 20265 min read

How do you test an agent that writes to your CRM?

You can't rehearse with an agent whose every conversation creates contacts, books slots, and posts notes into a live system of record. Mock the model and you've tested nothing; clone the CRM and you've tested a different system. The dry-run architecture that lets you exercise the real pipeline with the side effects intercepted at the last possible layer.

EngineeringJuly 2, 20265 min read

When the model hedges, force the tool call.

Ask a voice agent 'do you have Thursday at 2?' and a surprising fraction of the time the model answers 'we should have availability then' without checking the calendar. That is a hallucination with good manners. Why prompting doesn't fix it, and the mechanical guardrail that does.

EngineeringJuly 1, 20265 min read

The tool-call tax: every lookup costs a second trip through the model.

A voice turn that calls a tool runs the LLM twice: once to decide what to look up, once to read the result and answer. That second trip, not the lookup itself, is why tool turns feel twice as slow as plain ones. Where the tax comes from and the five levers that actually reduce it.

EngineeringJune 30, 20264 min read

The transcript table is append-only. The database enforces it, not us.

Every dispute about what an AI agent said gets settled by the transcript, which makes the transcript worth protecting from everyone, including the vendor. Why conversation logs should be append-only at the database layer, what the column-whitelist exception looks like, and the one-line question that tests whether a vendor's 'immutable logs' claim is architecture or policy.

EngineeringJune 29, 20265 min read

A filler phrase is a promise. Most voice agents break it.

'Let me check that for you' means work is happening. When a voice agent says it reflexively, before every turn, callers learn within three exchanges that the agent's words don't mean anything. The rules we ship for when an agent is allowed to speak a filler, and when silence is the honest answer.

ProductJune 28, 20266 min read

Why we don't stream the agent's reasoning live.

Every CX-AI vendor now ships a 'thinking…' panel that narrates the agent's reasoning token by token. We don't, on purpose. A live reasoning stream is a UX pattern borrowed from slow agents: by the time you finish reading the first line, our agent has already acted and moved to the next turn. The reasoning still gets recorded. It just lands where it is useful, on the durable audit row, not in a panel that vanishes when you close the tab.

CompanyMay 24, 20269 min read

What 'enterprise-ready' actually means in CX-AI.

Every CX-AI vendor claims to be enterprise-ready, and the phrase has stopped meaning anything. Here is a working definition across seven commitments (procurement, region, BAA, identity, audit, approvals, exit) plus the ten-question checklist a buyer can hand to any vendor in week one.

ProductMay 24, 20269 min read

Voice AI vs voice infrastructure: the category buyers conflate.

ElevenLabs and Sierra both ship under the "voice AI" banner, but they sell different products to different buyers. One is infrastructure a developer assembles into an agent. The other is a finished platform an operator buys to pick up the phone. The conflation costs buyers real money. Here is the map.

EngineeringMay 24, 20269 min read

ASR confidence in production: when to trust what STT says, and when to ask again.

Every speech-to-text provider returns a confidence score with each transcript. Most voice agents ignore it. That is a mistake. Confidence is the single best signal for deciding whether the agent should act on what it heard, ask the caller to repeat, or escalate. Here is the production logic we use, including per-intent thresholds, per-tenant tuning, and the graceful re-ask script that never blames the caller.

EngineeringMay 24, 202610 min read

Grading every voice turn. The production loop that catches hallucinations before they ship.

A voice agent can hallucinate on one turn out of fifty and still feel pretty good in aggregate. The problem is that one turn quoting a wrong price, inventing a hold time, or claiming a feature that does not exist is the turn that gets escalated, screenshotted, and tweeted. We grade every agent turn in production against a small list of factual constraints, and route any unsupported claim straight into the operator queue. Here is the architecture, the prompt, and why a per-turn grader beats a weekly eval.

CompanyMay 24, 20268 min read

Whose name is on the audit row? The question that splits CX-AI vendors in 30 seconds.

Every procurement evaluation of a CX-AI vendor should start with one question, and you can fit it on an index card. The vendor's answer (and how quickly they can answer it) tells you whether their architecture is operator-led or bot-first. For regulated work, only one answer is defensible. Here is what each answer sounds like and what to ask next.

EngineeringMay 24, 20268 min read

Streaming the first audible word. Why perceived latency beats actual latency in voice.

The caller does not measure latency. The caller measures when they first hear the agent say something. If the first audible word lands in 350ms, the caller feels heard even when the full reply takes two seconds. Here is the streaming-first-token architecture, the constraints, and why it is the single biggest perceived-latency win in voice AI.

EngineeringMay 24, 20269 min read

Designing the audit row: what every voice agent should write back to your CRM.

The audit row is the single most important artifact a voice agent produces. It is not the transcript and it is not the reasoning trace. It is a structured record, written into the operator's system of record, that answers 'what happened on that call?' without anyone having to listen to it. Here is the field model we ship and the design choices behind it.

CompanyMay 24, 202610 min read

The bot-first trap. Why every CX-AI vendor names the wrong protagonist.

Sierra's Agent OS. Decagon's AI Concierge. Ada's brand-native AI Agents. Every major vendor in this category names the AI as the protagonist of the system, and in regulated work that framing is a liability. Here is what the inverse stance, operator-led architecture, actually looks like.

EngineeringMay 24, 20268 min read

Barge-in: how the agent shuts up the moment the caller starts talking.

Real humans interrupt each other. Voice agents that finish their sentence anyway feel robotic. Handling barge-in well requires explicit interrupt detection, a decisive mid-syllable stop, partial-transcript handling, and conversation-state recovery that remembers what the agent had said before it got cut off. Here is how the production architecture actually works.

EngineeringMay 24, 20268 min read

How prompt caching cut our voice LLM bill by 70%.

Anthropic shipped prompt caching for Claude. Voice booking flows have a huge stable prefix and a tiny variable tail, which is the exact shape caching was built for. Here is the production math, the cache-write trade-off, and where it stops paying off.

EngineeringMay 24, 20267 min read

Skip the LLM: templated short-circuits for the easy 30% of voice turns.

A lot of voice turns don't need a frontier model. Booking confirmations, hours lookups, 'let me check that for you' filler: the slot data is in the tool result, the template is fixed, the model adds zero reasoning value. We route those turns to templated short-circuits, return in under 250ms, and pay $0 in inference cost.

EngineeringMay 23, 20266 min read

Why we're not shipping speech-to-speech yet.

Gemini's Live API collapses STT, LLM, and TTS into a single audio-in audio-out model. The latency win is real. The vendor-lock-in cost is bigger. Here is the engineering math behind why we said no for v1.

EngineeringMay 22, 20265 min read

Voice latency is the LLM. Everything else is a rounding error.

When a voice agent feels slow, it's almost never the network. We decomposed where the seconds go in a production pipeline, and the answer changes how you should reason about latency at every layer.

ResearchMay 18, 20264 min read

Why we built a benchmark we don't control.

CXBench is a public, multi-vendor benchmark for CS/voice AI agents. It was built with one design principle: vendors cannot tune to win. Here is what that means and how we enforce it.

EngineeringMay 15, 20264 min read

CRM-as-source-of-truth is not a feature. It's a tax bill.

Every CX-AI vendor's data layer is a vendor lock-in. The architectural choice to keep your customer record in your CRM, and have the agent write into it natively, costs the vendor real engineering and saves you real money. Here is why it matters more than buyers realize.

Newsletter

Want this in your inbox?

We send one short email when a new post lands. No marketing drip. No re-engagement sequences. Just the posts, when they ship.

The next call doesn’t have to go to voicemail.

Book a thirty-minute demo. We point Vorel at one of your real numbers on the same call.

Book a demo See the console