> For the complete documentation index, see [llms.txt](https://docs.voiceb.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.voiceb.ai/blog/your-voice-agent-sounds-great.-so-why-isnt-it-selling.md). # Your Voice Agent Sounds Great. So Why Isn't It Selling? In the last eighteen months, thousands of companies built AI voice agents. The tools made it easy: ElevenLabs gave everyone broadcast-quality voices, Retell and Bland made it possible to stand up a working agent in days instead of months. The demos were stunning. Executives heard an AI handle objections in a natural voice with sub-second latency, and budgets got approved on the spot. > Then the pilot ended, and somebody in finance asked the only question that mattered: > > Did it sell? > > And nobody could answer. There were transcripts. There were minutes consumed. There were sentiment scores, maybe a dashboard with call volumes. But there was no number anyone could put in front of a CFO that said: this many calls produced this many sales, this much pipeline, this cost per acquisition. The pilot didn't fail loudly. It just quietly produced no business case, and the project stalled. If this sounds familiar, the problem is probably not your voice agent. And it's almost certainly not the platform you built it on. The problem is structural — and once you see it, you can't unsee it. ### Voice is solved. Outcomes aren't. Let's give credit where it's due. The voice AI platforms are genuinely excellent at what they do. ElevenLabs made voice quality a commodity — we use their speech technology ourselves, and it's superb. Retell, Bland, and Vapi solved real-time orchestration: latency, interruption handling, turn-taking. Five years ago, each of these was a research problem. Today they're an API call.
**That's exactly the point. The conversation layer is solved. Any competent team can ship an agent that sounds human in a week.** But the conversation layer was never designed to answer commercial questions. It can transcribe a call perfectly and still have no idea what the call was. Was it a sale? A warm lead that needs a human? A wrong number? An existing customer asking about their bill? A competitor doing reconnaissance? \ A transcript is not an outcome. A sentiment score is not a revenue event. And a per-minute invoice tells you exactly one thing: how long people talked. It tells you nothing about whether any of it was worth paying for. \ This is the missing layer. Between a great-sounding agent and an actual revenue system sits a piece of infrastructure most voice deployments simply don't have: deterministic outcome classification, connected to billing, the CRM, and the ad platforms that generated the leads in the first place. ### The blind funnel Here's why this matters beyond reporting. Without outcome classification, you can't even diagnose why your pilot underperformed. A real example from our own production data. SHV Energy's Primagas brand has run autonomous sales agents with us for over a year, with a 67.8% autonomous close rate across twelve months. But there were specific periods where performance dropped sharply. If all we'd had were transcripts and minutes, the obvious conclusion would have been: the agent is failing. Tune the prompts. Change the voice. Maybe kill the project. Because every call carried a deterministic outcome label, we could see what was actually happening: the agent's close rate on valid calls hadn't moved. What had changed was the ratio of valid to invalid traffic. A specific batch of Meta Ads campaigns was sending low-quality leads — wrong audience, wrong intent. The agent was fine. The leads were wrong. That diagnosis took minutes, because the data existed. Without it, the same situation looks like AI failure, and that's exactly how most stalled voice pilots get read. The company concludes "voice AI doesn't work for us" when the truth was a fixable traffic problem nobody could see. This is worth saying directly to everyone whose pilot disappointed: your team probably didn't fail. You were flying blind, because the infrastructure to see the funnel didn't exist in your stack. #### Twelve months of outcome-classified traffic Primagas / SHV Energy, vs prior period

Primagas - SHV Energy live production. All figures deterministic and auditable.

{% hint style="success" %} The same outcome data that diagnosed the traffic-quality problem now prevents it: every classified call feeds the diagnosis, the CRM, and the ad platform. {% endhint %} ### What the missing layer actually is At VoiceB we call it the [Outcome Engine](https://voiceb.ai/outcome-engine), and it does three things the voice layer can't. First, it classifies every call deterministically. Not a probabilistic sentiment guess — an auditable classification. Every call gets a Status (Valid or Invalid), then an Outcome. A Sold outcome means the customer accepted the offer and provided the required data. A Talk to Agent outcome means warm intent that needs human follow-up. Invalid outcomes — not interested, voicemail, dropped, customer care, unreachable — are identified with 96% accuracy and separated from the sales flow entirely. On top of that, custom Tags apply your own business logic in plain language: high intent, price driven, SME, existing customer. Second, it makes the CRM trustworthy. Outcomes push to Salesforce or HubSpot in real time, structured and validated — and only the outcomes you specify trigger CRM writes. No more polluting your pipeline with voicemails and wrong numbers. Your sales team sees only sales-ready demand, with full context attached. Third, it closes the loop with your ad spend. This is the part almost nobody has. Every classified outcome is a conversion signal. Fed back to Meta and Google, it means your campaigns stop optimizing for form fills and clicks and start optimizing for what actually sold. The same outcome data that diagnosed the Primagas traffic problem now prevents it: the ad platform learns which audiences produce Sold outcomes and reallocates budget automatically. A lead arrives from a Meta ad. The agent qualifies it, closes it, writes the confirmed sale to the CRM, and tells Meta the lead converted — all within minutes, without a human touching anything. That's not a voice bot. That's a revenue system. #### What outcome classification reveals in real traffic Holaluz, live production

Holaluz, live production. All figures deterministic and auditable.

{% hint style="success" %} Every segment above is a deterministic, auditable classification — not an estimate. Invalid calls never reach a human agent and are never billed, on any plan. {% endhint %} ### The pricing model is the proof Here's a simple test for any voice AI vendor: look at how they charge you. If a vendor charges per minute or per call, they're telling you something important — they don't know what happened on the call. They can't distinguish a closed sale from a wrong number, so they bill both identically, and you carry all the risk. #### The pricing test
{% hint style="success" %} A vendor that can classify outcomes can afford to be paid on them. {% endhint %} We charge €0 for every invalid call. Not interested, voicemail, dropped, busy, unreachable — all free, on every plan. We only get paid when a real revenue event occurs: a qualified handoff to your team, or a sale closed autonomously. We can offer this for exactly one reason: the Outcome Engine classifies every call, so we know which ones created value. The pricing model isn't a marketing gimmick. It's the architecture made visible. A vendor who can classify outcomes can afford to be paid on them. A vendor who can't, can't. The results in production: at Holaluz, 30.4% of valid inbound calls now close autonomously, with 78% of total inbound traffic auto-filtered before a human is ever involved. At Primagas, the twelve-month numbers came alongside a 153% increase in sales and a 37% reduction in cost per acquisition — because for the first time, every euro of ad spend could be traced to an actual outcome. ### If your pilot stalled, start here I've sold conversion infrastructure to this exact buyer for fifteen years. At my previous company, Whisbi, we put human agents on video for AT\&T, Vodafone, Toyota, and thirty other enterprise brands — and we proved that live conversion of inbound demand works at industrial scale. The one thing we could never do was close the loop: prove the sale, attribute it, and feed it back to the ad platform automatically. That's the layer VoiceB was built to be. So if you've built a voice agent that sounds great but isn't moving the revenue line, don't start by rebuilding the agent. Start by asking what happened on the calls you already ran. We'll show you. Bring a sample of your agent's calls — or your call center's — and in twenty minutes we'll run an outcome classification on that traffic: how much was invalid, what was transferable, what was closeable. No demo, no pitch deck. Just your funnel, visible for the first time.
[Book your outcome audit](https://cal.com/team/voiceb-demo/voiceb-demo) — or [read the technical docs](https://docs.voiceb.ai) to see how classification works under the hood. *** VoiceB.ai is the Autonomous Sales Voice Agent (ASVA) platform. Live in production with Vodafone España, Holaluz, and SHV Energy/Primagas.
--- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://docs.voiceb.ai/blog/your-voice-agent-sounds-great.-so-why-isnt-it-selling.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.