Level AI

I came across Atonom's CX Awards last week. I nominated a number of people I thought should be included. Suddenly, I found myself on the list. Thanks, Luke Jamieson! If you want to vote, go for it - there are a *lot* of great people across 4 categories, many of whom I know personally. Being mentioned in the same breath as some of these people is all the award I need. But the real value here is for those looking to expand their network with people who provide value on LinkedIn for anyone in or adjacent to the CX space. If you're not already following or connected with people on this list, you're missing out. You're missing out on insights from people at brands doing the work every day. You're missing out on thoughtful, educational content from people who know what they're talking about. You might be missing out on the perfect person to fill that role you're hiring for, now or in the future. You're missing out on a larger CX community where people give without expectation of getting. Voting ends August 17th. https://lnkd.in/gJytBC6S

Imagine your smoke detector only went off after the 5,000th spark because the first 4,999 looked "routine." That's basically how most CX organizations operate. I've been thinking about how we define a "customer problem." Usually, it goes something like this: One angry escalation → investigate. One executive complaint → all hands on deck. One viral social post → emergency meeting. The problem is that we're looking at conversations one at a time instead of patterns across all of them. But customers don't experience your business one conversation at a time. They experience your workflows. That's the pattern that emerged when we analyzed 3.4 million enterprise support conversations. The biggest drivers of customer effort weren't spectacular outages or catastrophic failures. They were boring. "Where's my order?" "Can you check my application status?" "I'm just following up." "Nobody called me back." The kind of interactions nobody puts in a weekly executive deck because each one, by itself, looks insignificant. The problem is repetition. None of these make headlines. But together, they define the customer experience. I think that's where CX is heading. Less energy spent finding the loudest customer. More energy spent finding the emerging patterns. That's exactly why we built this report. If you're curious what those patterns look like across 3.4 million enterprise support conversations, the Visibility Gap Report is now live. Read here - https://lnkd.in/gKQVWFGq

One of my favorite customer stories recently reminded me of something important about Enterprise AI: Enterprise AI becomes powerful when you give it the right business context. The best AI workflows aren't always the ones that work perfectly out of the box. They're the ones that can learn the processes and institutional knowledge that make each business unique. Smarter models matter. But smarter context matters just as much. Problem : A customer in the auto and home insurance industry initially found our Coaching Plan AI agent underwhelming. The recommendations weren't aligned closely enough with their sales process, and supervisors had to manually filter the output—making the workflow feel counterproductive. Solution: The turning point was simple, they taught the AI Agent how they actually sell. The team uploaded their "perfect call flow" and refined the supporting PDFs used as guidance. The results? → Coaching sessions grew from ~10 to 25–28 per week → 25 unique agents coached in one week → One supervisor became the highest-frequency Coaching Worker user The most interesting part wasn't just the adoption. It was the shift from: "This isn't useful for us." to "This understands how we coach our teams." #EnterpriseAI #GenerativeAI #CustomerExperience #ContactCenterAI #CustomerSuccess

On CNBC earlier this month, Alex Karp told enterprise buyers to stop handing their data to frontier labs. When you feed proprietary data into someone else's model, you're handing your advantage to a vendor who can turn around and build a competing product. His point: technical buyers should own their compute, their models, and their data stack. That decision is hiding inside every CX platform. Every AI feature runs on a model. If that model is a frontier API, your data is leaving, your bill scales with volume, and the path from input to output is a black box. Access, pricing, and rate limits can change on someone else's schedule, and when that model sits behind your virtual agent or your QA process, losing it means gaps, brand risk, and churn. Everyone wants agentic AI. But an agent that acts on its own is only as good as your ability to see what it took in and measure what it did with it. Without access to the inputs, you can't effectively calibrate, and you're shipping a system you can't correct. That's what Latitude is for. 7 proprietary models fine-tuned for the jobs of CX: transcription, redaction, intent, summarization, inferred CSAT, QA, and Voice of Customer. They run on our own GPU clusters in a dedicated VPC, so customer data never reaches an outside model provider and never trains anyone else's. We own the models; you get full visibility into the inputs and outputs, and we recalibrate against human QA leads every month, because policies, products, and regulations move. The reason a family of smaller models beats one general-purpose LLM: each job has its own accuracy bar, latency budget, and failure mode. Transcription is judged word by word in noisy audio. Redaction is judged by what it catches before data leaves the boundary. QA is judged by whether a supervisor can trace a score back to evidence. One model can't be tuned for all of that at once. Ours are trained on more than a billion real customer interactions, so they know how support sounds, and each one is pointed at a single task. That's how they hold accuracy on par with GPT, Gemini, and Claude on live CX work, at a fraction of the size. The result: up to 49x lower cost to serve, up to 3.5x higher throughput, 4x lower latency, running at a 4 trillion token rate with 97% on our own models. Whose models are your customer conversations training right now? Level AI https://lnkd.in/duSWidBR

Your brand guide is probably 40 pages long. Your QA process checks maybe 3% of the calls where customers actually experience that brand. For most DTC brands, there's no store. No fitting room. No cashier saying hi. The contact center is the brand experience. It's the one place a real human represents you in real time. And it's the one place almost nobody is watching closely. Here's the math that should bother any CX or brand leader: 500,000 contacts a year. 3% QA coverage. That's 15,000 calls scored. 485,000 calls where nobody knows what your brand actually sounded like.❓ At a big retailer, one bad call barely matters. There are a thousand other touchpoints that day. At a DTC brand, that call is the brand that day. For that customer, it might be the only impression they get. 🎯 So why does visual identity get a whole team and zero tolerance for inconsistency, while service tone gets a spot check every once in a while?🤔 The brands that actually protect their brand score every interaction, not a sample. All... Of... It...💯 That's the gap we built Level AI's QA around. 100% coverage, brand specific criteria, so a call gets held to the same standard as your packaging. Your brand shouldn't come down to a coin flip. 🪙

Somewhere right now, a travel CX leader is looking at a rising handle time chart and getting ready to coach an agent who's doing harder work than anyone on that team did three years ago. The chart is right. The read on it is wrong. Here's what happened. Mobile self-service quietly took over the easy stuff. Check-in, seat changes, standard rebooking, baggage status. All the four-minute calls left the building and moved into the app. So think about what's actually left in the human queue. Loyalty disputes. Itineraries that fell apart mid-trip. Cancellations where someone's angry and stranded. Compensation conversations after the airline already dropped the ball. A 15-minute call in 2026 is doing the work a 4-minute call did in 2019. The volume didn't get slower. The mix got heavier. But most AHT benchmarks were set before any of that shifted. So teams are grading today's harder queue against a number built for a job that no longer exists, and it flags the wrong agents every time. The question worth asking isn't why handle time went up. It's what the escalations are actually telling you: → Which app failures push the most people into the human queue → Where the handoff from digital to agent keeps breaking, by journey type → Which problems show up again and again on fallback calls → How many of those calls a better app flow would have caught upstream An AHT report can't answer any of that. It only counts minutes. The answers live inside the escalation conversations themselves. At Level AI, we read every mobile-to-human escalation and trace it back to the digital gap that caused it, so CX and product teams can fix the thing generating the calls instead of coaching the people absorbing them. If your handle time stopped telling you anything useful once self-service took off, I'd love to talk shop.

Ever wonder why most virtual agents feel like they have a five-second memory span? It happens because conventional bots drop customer context the moment they move to the next step of a conversation. True automation maturity requires a different approach to infrastructure. Introducing  Level AI’s  Variable Management, a central memory system built directly into our platform architecture, that keeps customer data active across the entire call lifecycle. What this means for enterprise automation: - Zero AI guesswork: Direct database mapping connects variables to verified software records. - Faster workflows: Developers can build and assign data fields right inside their active canvas. - Built-in safety: Automated guardrails scan for risks before updates go live. It is time to bridge the gap between complex backend systems and seamless customer experiences. Read the full breakdown here: https://lnkd.in/da2HATBj

About a third of all customer interactions contain sensitive data. Name, address, social security number, payment info, etc. So if AI touches your customer conversations, find out how data custody is managed before asking "which model performs the best?" or "build vs. buy?" Once customer data leaves your environment in any way, shape, or form... your risk now lies in the hands of any downstream AI vendor. Our new whitepaper "The Operating Standard for Enterprise AI" outlines the risks, tradeoffs, and economics behind your tech stack selection and why things like built-in redaction are critical for security and compliance. #goodread for anyone working with customer-facing AI tools. Get the full whitepaper: https://lnkd.in/gGrM9ZGi