Voice AI Agents Break Into Production in 2026: Retell's Conductor, Coval's $28M Raise, and the Real-Time Phone-Agent Tools Worth Knowing

Introduction: Voice Agents Stopped Being a Demo

For most of the generative-AI era, the phone call was the one channel that resisted automation. Text chatbots could bluff through a web widget, but a stilted, two-second-delayed voice bot just drove customers to smash "0" for an agent. In the final week of June 2026, that assumption finally broke. A burst of launches and funding rounds signaled that real-time voice AI agents are now a production infrastructure category, not a novelty.

Voice startup Retell AI shipped Conductor, a product pitched as a way to "industrialize enterprise voice-agent operations" and featuring what the company calls the first-ever graph-native review interface for production voice agents. Days earlier, Coval raised a $28 million Series A "to define safety and reliability for autonomous voice agents", and healthcare-focused Assort Health banked a $120 million Series C to scale its voice-agent platform across medical call centers. If you are evaluating AI tools in 2026, voice is suddenly a category you cannot ignore. Here is what shipped, why it matters, and how to choose.

Why Voice Agents Finally Crossed the Threshold

Two engineering problems kept voice agents out of production for years, and both got meaningfully cheaper in 2026. The first is latency. A phone conversation feels broken above roughly 500 milliseconds of round-trip delay, but traditional speech systems stitched together separate transcription, reasoning, and text-to-speech steps that routinely took seconds. This year's real-time, speech-to-speech models — including the streaming mode behind Google's Gemini 3.1 Flash Live — collapsed those steps, bringing turn-taking latency down to a range that no longer screams "robot."

The second is reliability at scale. A chatbot that hallucinates a refund policy is annoying; a voice agent that books the wrong appointment or quotes the wrong price on a recorded line is a legal and brand problem. That reliability gap is precisely what the new wave of infrastructure companies is racing to close — and why investors are writing large checks to fund it.

The Late-June Signal: Infrastructure for Production Voice Agents

The clearest tell that a category has arrived is when money and tooling cluster around it in a single week. Three moves landed in late June 2026:

• Retell AI's Conductor. Billed as an operations layer for enterprise voice agents, with a graph-native review interface that lets teams inspect and debug multi-step calls the way a developer traces a workflow. The pitch is that shipping a voice agent is now the easy part; operating and improving it is the hard part.
• Coval's $28M Series A. Explicitly framed around "safety and reliability for autonomous voice agents," the round is a bet that evaluation, guardrails, and testing will become non-negotiable before enterprises let voice agents act unsupervised.
• Assort Health's $120M Series C. A vertical bet on healthcare call centers, where voice agents handle appointment scheduling, prescription refills, and triage — high-volume, high-compliance calls that are expensive to staff with humans.

Read together, these point to a maturing stack: a real-time model layer (Gemini Live and peers), an orchestration and operations layer (Retell Conductor), an evaluation and safety layer (Coval), and vertical deployments (Assort in healthcare). That is the shape of a real product category, not a science experiment.

The Proof Point: Home Depot Answers the Phone With Gemini

Infrastructure only matters if someone deploys it. The reference enterprise case is The Home Depot, which replaced its traditional phone menus with AI-powered voice agents built on Google Cloud's Gemini Enterprise for Customer Experience. The company reported handling customer store-phone support roughly four times faster than before, with callers able to simply state what they need instead of navigating a keypad tree. Fellow retailers including Macy's and Ulta have moved in the same direction, and pharma giant Merck has tapped the same Gemini platform for agent development.

The Home Depot example matters because it is not a controlled demo — it is a Fortune 100 customer-support line running in production, with measurable speed gains and a clear before/after. That is the benchmark every vendor in this category now has to beat.

Latency, Evaluation, and Safety: The New Battleground

With raw conversational quality mostly solved, the competition has shifted to the unglamorous engineering that decides whether an agent survives contact with real callers:

• Latency budgeting. Every tool in the call path — telephony, speech-to-text, the model, text-to-speech, function calls — eats into your sub-second budget. The best stacks let you measure and trim each hop.
• Evaluation at scale. You cannot manually QA thousands of calls. The emerging playbook is synthetic test-call suites and automatic scoring of transcripts, the kind of capability Coval is betting on.
• Guardrails and handoff. Production agents need hard rules for when to act (book, refund, schedule) and when to hand off to a human, plus a clean audit trail for every decision.

Notice that none of these are about the model being smarter. They are about making a probabilistic system safe enough to put on a customer-facing phone line — which is exactly why a tooling layer is forming around the models.

What This Means for the AI Voice Tools You Pick

If your organization handles any meaningful volume of inbound or outbound calls, voice agents moved from "someday" to "evaluate now" in 2026. A few practical decisions to make:

• Build on a platform, not just a model. A low-latency model is necessary but not sufficient. Prioritize providers that ship the orchestration, telephony, and observability around it — the gap between Retell-style platforms and raw model APIs is where most projects stall.
• Buy evaluation before you buy scale. Lock in a way to test, score, and review calls automatically before you route real volume. Tooling that makes multi-step calls debuggable (Conductor's graph-native review is one example) pays for itself the first time you catch a systematic error.
• Match the deployment to the vertical. Healthcare, finance, and retail each have different compliance, PHI/PII, and disclosure rules. Vertical players like Assort Health exist precisely because a generic voice stack does not satisfy them.
• Keep a human in the loop — and make handoff seamless. The Home Depots of the world are not firing call-center staff; they are triaging calls so humans handle the complex ones. Plan for graceful escalation from day one.
• Verify every claim with a test call. Vendor latency and accuracy numbers are best-case. The only number that matters is how the agent performs on your actual scripts, accents, and edge cases under real network conditions.

The Honest Caveats

The momentum is real, but the category is still young. Most performance figures — including the "4x faster" results — are reported by the deploying companies and their vendors, not independent auditors, and they describe specific workloads, not yours. Voice agents also sit in a regulatory gray zone: in several regions, disclosing that a caller is speaking to an AI is already required, and recording, storing, and acting on call content triggers familiar privacy obligations. Treat any voice-agent rollout the way you would any customer-facing automation: pilot small, instrument everything, and let your own metrics — not the press release — decide whether it earns more volume.

The Bottom Line

Voice AI agents crossed from demo to production in 2026, and the last week of June drew the category's outlines clearly: a real-time model layer, an operations and review layer, an evaluation and safety layer, and vertical deployments with real revenue behind them. The practical takeaway is that voice is now a distinct tool-buying decision — one where latency, evaluation, and compliance matter as much as conversational quality. Start with a platform that lets you observe and test calls, match the stack to your vertical's rules, and keep a fast path to a human. The phone call, long the holdout channel, is finally becoming a software problem — and the teams that choose their voice tools deliberately will be the ones who solve it well.

Frequently Asked Questions