The demo always works beautifully. An AI agent fields a customer complaint, navigates a menu order, maybe even cracks a joke. Executives lean forward. Then someone asks the inevitable question: Can we put this on our 1-800 line tomorrow?
That's when the conversation gets uncomfortable.
A small team of former Google engineers believes the real distance between a polished voice AI prototype and an enterprise deployment isn't about making the conversation sound more human. It's about everything else—the audit trails, the escalation protocols, the SOC 2 checkboxes, the operational scaffolding that lets a company route thousands of actual customer calls through an AI system without executive-level panic.
Dialogus, a three-person startup that went through Y Combinator's accelerator program, is building infrastructure designed specifically for that gap. Not the conversational flourish that lands well in controlled environments, but what co-founder Hans Ibarra describes as the "auditable, policy-constrained machinery" required when the stakes are production customer calls, not internal experiments.
The company launched its platform recently with a pitch that skips past natural language polish entirely. Instead, it focuses on workflow integration: executing pre-approved tools, logging every decision the AI makes, failing gracefully when things go wrong, and escalating to human agents when guardrails are breached. The target customer is the enterprise stuck between an impressive proof-of-concept and a live phone system.
Constraints as a Feature
At the heart of the platform sits what Dialogus calls "Wildfire"—a runtime engine that constrains the underlying language model to the current workflow state, customer context, and a whitelist of valid actions for each conversational turn. The system doesn't simply listen and respond. It cycles through four stages the team labels Listen, Reason, Act, and Report.
Listening handles the mechanics of turn detection, barge-in interruptions, and first-audio latency—the milliseconds that determine whether a conversation feels natural or robotic. Reasoning runs inside Wildfire's constraint engine, which limits what the model can propose based on company policies and available tools. Acting means calling approved APIs: updating a CRM record, modifying an order in a point-of-sale system, sending a WhatsApp confirmation, or handing the call off to a human.
Reporting generates structured evidence. Full transcripts. Tool execution logs. Latency measurements. Error traces. Every call becomes a replayable test case, which production failures can be fed back into what the company describes as a "voice lab" for regression testing.
The idea—perhaps more ambitious than it sounds—is to make voice agent operations debuggable in the way backend services are, not the black-box affair conversational AI has been historically.
A Claim That Needs Context
Dialogus claims on its homepage to be handling thousands of calls for Papa Johns, though this has not been independently verified. There's an asterisk worth noting: Papa Johns has publicly rolled out an AI ordering agent, but tied it to Google Cloud's Food Ordering Agent platform, not Dialogus. Whether the startup is serving a regional franchise, a specific workflow layer, or something else entirely remains unclear. The company hasn't released a case study.
The initial vertical focus splits between customer support resolution and phone order intake—workflows that demand deep integration with point-of-sale systems, CRM platforms, and telephony infrastructure. These are table stakes if the target customer is a restaurant chain or call center operator, not a developer tinkering with OpenAI's Realtime API on a side project.
Crowded Timing
Dialogus enters a market in the middle of what looks increasingly like a land rush. Five9 launched Voice AI Agents with an "AI Agent Studio" this past June, explicitly positioning the release around moving enterprises "from pilots to production." Twilio made Agent Connect generally available in May, framing infrastructure complexity as the primary blocker to scale. IBM embedded Deepgram into its agent builder earlier in the year. PolyAI opened its dialog platform to all builders in the spring. LiveKit raised $100 million at a $1 billion valuation to scale real-time voice and video infrastructure.
The common thread: everyone is racing to close what the industry has started calling the operationalization gap. The first wave of voice agent platforms prioritized developer experience and conversational quality—how natural the AI sounded, how well it understood intent. The second wave is adding compliance controls, observability tooling, and the trust mechanisms enterprises actually require before handing customer relationships to software.
Dialogus' bet is that a purpose-built operations layer will matter more than raw conversational ability. It's infrastructure that assumes the language model will eventually work well enough and focuses on everything that comes after.
The Team and What's Public
The founding team includes Ibarra, who worked on Gemini, Gmail, and YouTube at Google, along with engineers Juberth Rodriguez and Rodrigo Terán, who has stints at Google, Meta, and Microsoft on his résumé. All three are engineers. The company is not actively hiring, according to its careers page.
Beyond Y Combinator's standard investment, no additional funding has been disclosed. The team remains at three people, though that figure comes from the company's accelerator profile and may not reflect recent movement.
What Enterprises Will Ask For
The company's SOC 2 compliance claim appears on its security page, but there's no publicly available attestation report. For enterprises evaluating the platform—particularly those in regulated industries—requesting a current Type II letter and understanding the audit scope would be standard practice before signing anything.
The Papa Johns relationship, if it exists in the form Dialogus suggests, would be the most concrete proof that the startup has crossed from infrastructure-in-theory to infrastructure-in-production. Until there's independent confirmation or a public case study, it remains a data point with an asterisk attached.
Still. The architecture Dialogus describes does address the specific pain points enterprise buyers have been vocal about: auditable tool execution, human escalation paths, policy guardrails, and replay-driven debugging. If the team can deliver that at scale, the question won't be whether the market exists—it clearly does—but whether a three-person startup can move fast enough in a space where billion-dollar incumbents are shipping new features monthly.
In enterprise software, the unglamorous problems often matter more than the elegant demos. Dialogus is betting everything on that being true for voice AI, too.
