Level AI

I've talked to a lot of contact center leaders over the years but recently, the conversations have shifted. It used to be, "Should we deploy an AI agent?" Now it's, "Well, we've deployed one but we're not completely sure what it's doing out there." 👀 If you've deployed any kind of AI Virtual Agent and think, "We have containment rate. We have CSAT where customers respond. But neither tells me what the agent actually did on any given call," then I'm talking to you. The gap between the metrics most teams have and the visibility they actually need is what I'm unpacking with Sumeet Khullar on May 7. Sumeet is co-founder and CTO of Level AI. He's spent years building quality infrastructure for contact centers. On May 7, he's walking through what closing that gap actually looks like, live in the platform. 45 minutes. Free. I'll be asking the questions your ops team is probably already asking internally. :)

Toronto ✅ The Five9 AI Roadshow did not disappoint, stay tuned for what's coming next! Ian Storm Donal Hennessy

Join Sumeet Khullar and Rob Dwyer on May 7th for 45mins to understand how to build a top-notch AI Agent that is built for real life complexity and compliance! Register now: https://lnkd.in/dksMJy-Q

"Every feedback has to be associated with a decision. Do something or decide not to do anything but have a decision." Vishal Anam, Head of CX Consulting at Datamatics, is not talking about response templates or closed-loop ticketing systems. He is talking about a structural requirement that most CX programs never put in place. Letting feedback accumulate, aggregate, and surface in a dashboard is not a process. It is autopilot. And autopilot is why most programs produce no real change. The distinction between listening and acting starts with making a decision mandatory, even if that decision is to do nothing. Full episode: https://lnkd.in/dtXvHx7q

Many teams running an AI support agent have the same gap: they can see containment, but they still lack the boundaries, runtime controls, and inspection mechanisms needed to run that agent safely in production. Containment only tells you whether the customer reached a human. It does not tell you whether the agent made the right decision, followed policy, used the right system correctly, or actually resolved the problem. A conversation can be contained and still fail the customer. That is the theme of my new three-part series: What building AI agents taught me about silent AI failures. Part 1 is live now: The failures hiding inside your AI agent containment numbers. Link in comments! Part 2: Four AI agent failure types that will not show up in your QA reports Part 3: What breaks when your AI agent and your QA tool are separate systems On May 7, Rob Dwyer and I will walk through how to solve all three live in the Level AI platform. https://lnkd.in/gzZviDM2

I've talked to a lot of contact center leaders over the years but recently, the conversations have shifted. It used to be, "Should we deploy an AI agent?" Now it's, "Well, we've deployed one but we're not completely sure what it's doing out there." 👀 If you've deployed any kind of AI Virtual Agent and think, "We have containment rate. We have CSAT where customers respond. But neither tells me what the agent actually did on any given call," then I'm talking to you. The gap between the metrics most teams have and the visibility they actually need is what I'm unpacking with Sumeet Khullar on May 7. Sumeet is co-founder and CTO of Level AI. He's spent years building quality infrastructure for contact centers. On May 7, he's walking through what closing that gap actually looks like, live in the platform. 45 minutes. Free. I'll be asking the questions your ops team is probably already asking internally. :)

See you tomorrow, Toronto! 🧡

Most enterprises treat voice authentication like a gate. You're in, or you're out. But as we move toward autonomous AI agents capable of moving funds and accessing PII, that pass-fail logic is exactly why fraud is projected to hit $415B by 2028. When we architected our latest voice stack, we stopped looking at authentication as a hurdle and started looking at it as a streaming data problem. Here’s the reality of what we’re solving for: 1️⃣ Precision vs. Hallucination: General-purpose STT/TTS models are built for conversation, not rigid alphanumeric data. We’ve implemented specialized models to eliminate the phonetic drift that causes false rejections. 2️⃣ Latency vs. Integrity: A security handshake that exceeds 200ms creates dead air. We’ve optimized our SIP stack to ensure liveness detection and CRM lookups happen mid-stream, without breaking conversational flow. 3️⃣ Friction vs. Risk: We advocate for the Principle of Least Privilege (PoLP). Security should be proportional to risk, escalating to Multi-Factor (MFA) only when the transaction context demands it. The goal? Zero-trust security with zero-friction CX. Explore our technical blueprint for a resilient voice authentication strategy here: https://bit.ly/3OeKUJn

Most AI agent deployments have no visibility into what the agent actually did on a live call — which systems it contacted, what parameters it sent, whether the outcome processed. Our Co-Founder and CTO, Sumeet Khullar has spent a decade building quality and governance infrastructure for contact centers. In a new 3-part blog series, he breaks down the governance problem most AI agent deployments leave unsolved. Part 1 (April 22): The failures hiding inside your AI Agent containment numbers Part 2 (April 27): Four AI agent failure types that won't show up in your QA reports Part 3 (April 30): What breaks when your AI agent and your QA tool are separate systems On May 7, Sumeet Khullar walks through all three live in the Level AI platform — including a real failed conversation being caught, explained, and fixed. Register now if you're running an AI support agent in production: https://lnkd.in/dkFy_7aB