Dialogus wants to rebuild the call center, one resolved call at a time

The gap in enterprise AI is rarely where the marketing points. It is not the model, and it is not the voice. It is the moment a caller says something the script did not anticipate - a wrong account number, a refund request that touches three systems, a payment that needs confirming - and the whole polished demo has to become software that actually works. Dialogus, a San Francisco company in Y Combinator's Summer 2026 batch, built itself around that unglamorous moment.

Its pitch is deliberately plain: "Infra for enterprise voice agents." Not a chatbot, not a persona, not an assistant with a name. Infrastructure. The three founders - Rodrigo Teran Hernandez, Hans Ibarra and Juberth Rodriguez - looked at the contact center, one of the oldest and least loved corners of enterprise operations, and decided the interesting problem was not automating a single call. It was rebuilding the thing that routes, records and resolves all of them.

The ProductA voice operating layer, not a voice

Dialogus describes what it makes as a voice operating layer. In plain terms, that means three things happen in sequence on every call: real-time speech comes in, a constrained workflow runs in the middle, and structured evidence comes out the other side. The agent listens with turn detection and silence handling, reasons within the boundaries of a company's policies and customer data, executes actions - looking up a record, updating it, taking an order, confirming a payment - and asks the customer to confirm before it commits to anything consequential. When it gets stuck, it hands off to a human. When it finishes, it leaves a transcript, latency metrics and a recorded outcome behind.

That last step is easy to skim past and it is arguably the whole company. Plenty of voice systems can hold a fluent conversation. Far fewer can tell you, afterward, exactly what they did, how long it took, and whether it was allowed. Dialogus treats that receipt as a first-class product feature rather than a logging afterthought, and it built to SOC 2 compliance so the receipt means something to a security team.

The design choice that follows from this is restraint. A general-purpose language model, left to its own devices, will happily improvise. On a live customer call about a late payment or a missing order, improvisation is a liability. So Dialogus wraps the model in workflow constraints: the agent can only take the actions a company's policies permit, it has to check what the caller wants against real records, and it pauses for a human confirmation before doing anything that moves money or changes an account. The result is closer to a well-supervised employee than to an open-ended assistant. That is a feature, not a limitation, in the settings Dialogus is chasing.

It also connects to the systems that make a call useful. A voice agent that cannot see a customer's order history or update a CRM is a novelty. Dialogus plugs into phone systems, internal tools, databases, CRMs and messaging channels, so that the conversation and the underlying records stay in sync. When the call ends, a follow-up message can go out and the record reflects what happened. The plumbing is the product, and the company seems comfortable saying so.

Real-time speech in, controlled workflow execution in the middle, structured evidence out.Dialogus, on its own architecture

The CustomersWhere the agents already answer

The tell that Dialogus is past the demo stage is its customer list. The company says it is powering AI voice operations for Fortune 500 enterprises and handling thousands of calls in production. The named users are recognizable: Papa Johns and KFC on the quick-service restaurant side, and Totalplay, a large Mexican telecom operator. These are not low-stakes lines. A wrong order goes out for delivery; a mishandled telecom call becomes a churned subscriber.

The founder emails run on riko.mx, a detail that hints at the team's roots in Mexico's restaurant-ordering and consumer-tech scene - useful context for why order-taking and telecom support were early proving grounds. The product's three headline solutions map onto exactly those workflows: customer support that resolves a live call end-to-end, order calls that drop straight into a point-of-sale system, and collections that pursue past-due accounts within policy.

Each of those three is a different kind of hard. Order calls are latency-sensitive and unforgiving: a customer rattling off a modified order expects the agent to keep up, and the order has to land in the point-of-sale system correctly or a store makes the wrong food. Support calls are open-ended, where the challenge is knowing when to resolve and when to escalate. Collections are the most regulated of the three, bound by rules about what an agent can say, when it can call, and how it must document each contact. Building a system that can move across all three without breaking policy is a good stress test for the claim that Dialogus is infrastructure rather than a single-use bot.

The ProblemWhat breaks in the old contact center

The case for a company like Dialogus starts with the pain its buyers already know. Contact centers run on high labor costs, long hold times and inconsistent service. Multi-location operators - a restaurant chain with hundreds of stores, a telecom with millions of subscribers - face the added burden of hiring, training and retaining thousands of agents who each deliver a slightly different version of the brand. Turnover is brutal. Quality drifts. The software underneath was mostly built for a pre-AI era.

Dialogus's answer is a single voice layer that scales across every location and does not have an off day. The argument is not that AI is cheaper, though it is. It is that one controlled system produces a more consistent customer experience than an army of people reading from a script - and that it can prove, call by call, what it did.

There is a second problem the company is quietly built to solve: the gap between what an AI voice agent looks like in a sales demo and what it does under real load. Demos are curated. Production is not. Real callers interrupt, mumble, change their minds and ask for things the system was never told about. A pilot that dazzles in a boardroom can quietly fail the first week it meets actual customers. By designing for escalation, confirmation and evidence from the start, Dialogus is trying to make the demo-to-production drop-off small enough that a large enterprise will trust it on a live line.

The DifferenceControl as the moat

The voice-AI field is crowded. Synthflow, Balto, Sierra, Decagon, PolyAI, Parloa and Cognigy are all chasing enterprise phone workflows, and incumbents like Five9, Genesys and NICE are bolting AI onto existing contact-center suites. In a market that size, the underlying speech models are increasingly a commodity. Dialogus's bet is that the durable advantage is not the model but the control around it: constrained reasoning, mandatory confirmations, clean escalation and an audit trail a compliance officer will accept.

The interesting problem was never automating one call. It was rebuilding the system that resolves all of them.

The TeamWho left big tech to answer the phone

The founders bring resumes from Google, Meta and Microsoft. Hans Ibarra's background includes work at Google touching Gemini, Gmail and YouTube; Rodrigo Teran's spans Google, Meta and Microsoft; Juberth Rodriguez is listed as co-founder and CEO. There is a signal in where experienced engineers go when they leave companies used by billions of people. This team went to the contact center - not the obvious glamour bet, which is often the point.

At a team of three, Dialogus is early. It has no announced priced round beyond its YC backing, and much of its traction is self-reported. But the shape of the company is unusually legible: a clear problem, a named set of demanding customers, and a product thesis - control and evidence over conversational flash - that is easy to test against reality every time a phone rings.

Being small also cuts the other way. A three-person team serving Fortune 500 phone operations is stretched thin the moment a customer wants a new integration or a regional compliance quirk handled. The company's design choices - constrained workflows, standard connectors, audit by default - read partly as engineering discipline and partly as survival strategy for a tiny team that cannot afford to babysit every edge case by hand. The system has to hold up on its own, because there are not enough founders to hold it up for it.

The BetWhere it fits in the market

Where Dialogus sits in the market is, in a sense, underneath it. The visible layer of voice AI is the agent a customer hears. The layer Dialogus is building for is the one an enterprise's operations and security teams care about: the workflow engine, the integrations, the compliance posture and the record of what happened. Owning that layer is less glamorous and harder to demo, but it is also stickier once an enterprise commits, because ripping out the plumbing is far more painful than swapping a voice.

If the voice-AI market grows the way analysts expect, the winners will not all be the companies with the best-sounding agents. Some will be the ones enterprises trust to run revenue-critical calls without surprises. Dialogus is building for that second category. Whether it becomes the operating layer for enterprise voice or one capable player among many, its wager is refreshingly concrete: make the boring parts - compliance, integrations, proof - the product, and let everyone else fight over the demo.