AiThority Interview with Brooke Hopkins, Founder of Coval

Brooke Hopkins, founder of Coval discusses the deepening impact of Voice AI and how modern adopters should pay more attention to their respective infrastructure norms to drive optimized business processes and calls:

_____

Hi Brooke, please tell us about Coval’s growth journey so far.

My background is from Waymo, where I led the evaluation infrastructure team. My job there was answering the question: how do you know a self-driving car is safe enough to put on a public road? It took millions of simulated miles, regression suites, and production monitoring for the probabilistic system to earn trust.

When I left, I saw the same problem arriving for voice AI. Every company is going to have an autonomous voice agent the way every company today has a website. And almost none of them had the infrastructure to deploy one with confidence. I founded Coval in 2024 to build that infrastructure, and the growth has validated that the timing was right. We’re now working with Fortune 500 enterprises across largely regulated industries like financial services, healthcare, and insurance. We recently closed a $28M Series A led by Norwest, and we’ve been growing the team quickly across engineering, sales, and partnerships.

What about enterprise voice AI (goals / adoption / other aspects) should modern adopters exploring these technologies keep in mind?

Evals are the product – you can’t build a great product if you don’t have clear requirements for what it can and can’t do. In the AI world when models can do anything, features aren’t a differentiator anymore. How well you do a task is, and so is being able to clearly define what you don’t do. You can’t make this happen without evals.

Evals are what allow you to evolve and keep pace with the market. Even though evals take more time to set up the first iteration, they make every subsequent iteration faster and systematically improve the product. This is what allows you to avoid being stuck with a legacy model two years down the line that you are too scared to swap out.

It’s better to have one killer feature backed by evals, than twelve brittle features that create AI slop for your users.

How would you break down the basics around infrastructure norms for voice AI?

A production voice agent is really six layers working together:

• Telephony handles the actual phone connection.
• Speech-to-text (STT) turns audio into words.
• LLMs decide what to say and do.
• Text-to-speech (TTS) turns words back into a voice.
• Orchestration is the framework wiring it all together.
• Evaluation is how you know whether any of it is actually working.

There is an interesting parallel here to self-driving.

• Transcription (STT) = Perception → “What is happening in the world around me?”
• LLM Reasoning = Planning → “What should I do next?”
• Text to Speech (TTS) = Controls → “How do I take that action?”

Most teams pour almost all of their attention into the first five layers and treat evaluation as an afterthought. You can swap any other layer you want, but if you can’t measure how your agent performs across a wide range of real conversations, making changes become too risky and time-consuming to do so and you’ll quickly fall behind. You won’t know what broke, when it broke, or how often it happens.

The best and fastest teams in the industry are treating evaluation not as something you bolt on at the end, but as part of the foundation that needs to be in place before you scale production.

Also Read: AiThority Interview with Gou Rao, co-founder and CEO at NeuBird AI

When testing and monitoring, what aspects would you advise modern adopters to measure?

Testing in a perfect environment is the most common mistake teams make. Your agent will sound great when the audio is clean, the caller is patient, and the conversation goes as scripted. But that’s not what real calls look like. You need to test with accents, background noise, interruptions, and callers who go completely off script. We used advanced techniques to model out how these audio combinations would sound given the physics and sound qualities of a room. Here’s what I’d measure, broken into a few areas.

• Audio quality:

Speech tempo, background noise levels, voice consistency throughout a call (volume, pitch, tone), dropouts, clipping, and awkward pauses all contribute to how delightful the call is.

• Responsiveness:

Latency sounds easy to measure but it’s actually way harder than just a Python script and a timer. You need accurate turn detection, visibility into end-to-end response time, as well as into each component: speech-to-text, model response, audio rendering. And you need to know when the agent fails to respond entirely, which is an immediate critical failure.

• Conversation behavior:

How often is the agent interrupting the caller? Is it repeating itself? Is the speaking time balanced? These signals tell you whether the conversation feels natural or whether callers are fighting to be heard.

• Compliance and workflow adherence:

In regulated industries, you need to verify that the agent followed required workflows and made required disclosures on every call, not just the ones you happened to review.

• Transcription accuracy:

If the system isn’t accurately hearing what callers say, everything downstream breaks. Measuring word error rate (WER) against an independent transcription of the same call tells you whether that foundation is solid.

A big difference between traditional software testing and autonomous agent testing is how you measure success. Instead of thinking in absolutes (pass/fail), you need to measure things in terms of probabilities (at what rate will this happen to users) and determine what is the risk or quality level you are comfortable with. A compliance violation will have a different probability risk tolerance than an awkward pause. There are always tradeoffs, and once you have a basic voice agent working, the last 20% is about making the right optimizations and tradeoffs.

As human-like voice AI agents become mainstream, what ethics and security protocols should end users be more cautious about?

You should have a passcode or safe word with your family and loved ones so that if you get suspicious calls from someone you know, you can verify it’s actually them. Phishing has been going on for years, but it’s now easier than ever to deepfake your voice, so the scam is just now easier and more realistic. If you aren’t sure if it’s AI, just ask. Reputable companies will tell you (they’re actually required to by law for outbound calls).

From there, the same basic instincts that serve you well on any phone call apply here. Voice agents in financial services, healthcare, and other regulated industries will often ask for personal information to verify your identity or complete a transaction, which is completely normal. But if something feels off, it’s always fine to hang up and call back on a number you know is real. It’s also worth knowing how the companies you interact with handle your data. SOC 2 compliance and a clear privacy policy are good baseline signals that a company takes data security seriously, and most reputable companies will be upfront about this if you ask, or have it on their site.

Since the voices sound so real, it’s easy to forget that they’re not perfect and can make mistakes.  A voice agent might mishear you or give you incorrect information confidently. That’s why Coval exists, but for anything high-stakes, it’s worth verifying through another channel before acting on it.

Do you have any thoughts on the future of AI?

The interface is changing in a way that I think people are underestimating. For decades we interacted with computers in the same way through screens on big devices. Now we’re just going to talk to our computers the way we talk to our coworkers wherever we are: in a truck, on a site, in the grocery store or in a car. That shift is already happening inside major enterprises, and when the interface changes at that scale, everything gets rebuilt on top of it.

What follows from that is agents that don’t just answer questions but actually do things, reliably. Book appointments, resolve disputes, handle complex multi-step tasks end to end. Customers will expect that they can interface with products conversationally.

We already see this at Coval: we have agents running alongside our team every day. One focused on compliance, one on GTM, one tracking our hiring pipeline. They’re doing real work that would otherwise fall to a person. And soon, instead of interacting with them through a chat interface or a terminal, you’ll just talk to them out loud the way you would any other coworker.

That said, I don’t think the human goes away. The human sets direction, makes judgment calls, and stays in the loop on anything that matters. The agent handles the volume. That’s actually a more interesting version of work for most people, not a threatening one.

The part I think about most is what has to be true for that world to actually work. Every major technology platform eventually had to build the trust layer that made it durable: the security standards, the audit frameworks, the shared language for what “reliable” means. Voice AI is in that moment right now. Getting it right means AI becomes something people and enterprises can genuinely rely on. Getting it wrong sets the whole category back. That’s why I’m so focused on the infrastructure side, and why I think it’s one of the most consequential problems in technology today.

Also Read: ​​AI and The Future of Work: Artificial Intelligence Is Expanding Organizational Intelligence Beyond Human Limits

[To share your insights with us, please write to psen@itechseries.com]