Whether AI pays off in the contact center depends on three things around the model: the knowledge it draws from, the systems it connects to, and the people who own it.
That was the through line at Level Up 2026, where over 80 CX and contact center leaders from 33 companies spent three days at Hayes Mansion in San Jose. Across two panels, a keynote from Level AI Founder and CEO Ashish Nagar, two product launches, and a set of hands-on breakouts, they compared what’s working, what’s stalling, and what they’re planning for 2027.
Here are the five shifts we heard, and what they mean for your team.
1. Systems, data, and people now set the pace for AI
In his keynote, Ashish shared what a year in customer contact centers taught him. The teams that won with AI moved deliberately. What separated them was how AI connects to existing systems, how it hands work to people, who owns it, and how it improves over time.
Ben Huber, Director of Business Performance Analytics at Topcon Positioning Systems, described the same pattern from the inside in a breakout he co-led with Erik Katzen, Director of Sales Engineering at Level AI. Topcon supports precision positioning equipment with a six-figure SKU catalog. Its knowledge base only surfaced an answer when someone searched the author’s exact wording. Ben’s lesson from deploying a virtual agent on top of that content was blunt: AI on top of bad data produces bad answers faster. Knowledge quality had to come first.
2. Leaders are tying AI’s return to quality, loyalty, and lifetime value
Thursday’s opening panel took on how contact centers should measure what AI returns. Rob Dwyer, Senior Technical Account Manager at Level AI, moderated a conversation with Derek Mathews, Vice President at Idaho Central Credit Union; Benjamin Wedge, Head of Post Delivery, Global Partner Support at Wayfair; Paul Harraghy, Vice President of CARE at Vistaprint; and Ryan Burke, a CX leader formerly of Ulta, Roblox, and Chime. The panelists pointed to customer spend, lifetime value, and loyalty as the return that matters most. When AI works well, customers often engage more, so those outcomes show its value more fully than cost alone. The hard part, they said, is connecting day-to-day contact center operations to those outcomes.
Quality came up as one of the clearest levers. One leader’s team runs at a quality score near 65% against a target of 85% and sees closing that gap as a major gross-profit opportunity. [confirm speaker and approval] Derek Mathews’ team at Idaho Central Credit Union, which serves more than 850,000 members who prefer to call, measures AI by customer effort and compares it directly against human agents. [confirm speaker]
The right metric depends on the business. Jane Griener, Director of Operations, Customer Care at Purple, explained on Friday’s closing panel that Purple routes sales conversations to a live agent as fast as possible. Live agents drive higher cart value and conversion, so containment is the wrong target for those contacts. For leaders building a business case, Jane advised starting with the value that is simplest to prove, such as automated QA, and then making ongoing savings visible to finance so renewal is straightforward.
That approach is already working for some teams: one attendee’s team built its renewal case on a 60% efficiency gain in QA.
3. Human work gets harder, and coaching has to scale
Attendees described the same change across their operations: as virtual agents take on simpler contacts, the calls that reach people are more complex, more ambiguous, and more emotional, and new agents face that complexity from their first week. Benjamin Wedge of Wayfair described a second pressure: AI lets engineering teams ship far more changes to agent tools, and thousands of agents have to absorb each one while a frustrated customer is on the line.
Ashish expects this division of labor to deepen. In his keynote, he described AI workers moving from answering questions to planning and completing work on their own over hours or days, across the front office and the back office. As that happens, people spend less time doing the work and more time directing it. Theganizations getting the most from AI, he said, put AI-forward people into every function: people with the judgment to spot weak output, the framing to define a problem well, the systems thinking to turn one-off tasks into repeatable workflows, and the ownership to stand behind the result.
Those skills come from practice, and coaching still waits on a coach’s calendar, sometimes for weeks.
Level AI launched Coaching Simulation at Level Up so agents can work on a skill gap the same week it shows up. Sumeet Khullar, CTO at Level AI, showed how an AI Coach Worker spots an agent whose scores are slipping and builds practice scenarios from the team’s own conversations, with customer personas drawn from real calls. Each practice call is scored within seconds by the same AutoQA model that grades live calls, so practice counts toward the agent’s real KPIs. Read more about Coaching Simulation
4. Virtual agents are products, and CX teams will build them
Voice agents will keep getting better, Ashish said, and enterprise adoption will still move slower than many expect. Integrations with enterprise systems and handoffs to people will set the pace. He urged teams to treat go-live as the start of a voice agent’s life. Agents that still deliver five years from now are planned as a loop of design, guardrails, testing, launch, and improvement, with a clear owner for each integration and a path for customer and agent feedback to flow back in.
Topcon’s rollout showed who these agents end up serving. Ben said veteran equipment operators adopted the voice agent faster than his team expected, while younger operators assumed they didn’t need it. Technicians troubleshooting in the field use it hands-free, with earbuds in and both hands on the equipment. [confirm with Ben] The agent took overflow at the back of the IVR, and 40% of those callers chose it over waiting or leaving a voicemail. [reconcile with awards tab figures] Every escalation hands the human agent a full summary. As routine questions moved to the agent, Topcon’s team had time to improve knowledge content, which expanded what the agent could answer. Support that used to close at 8 p.m. ET now runs around the clock.
Teams often stall at the build, sometimes for weeks of specialist prompt tuning. Level AI’s new VA Copilot lets CX teams stand up a working virtual agent without waiting on specialists. Teams describe the agent in plain language and attach their SOPs, and the Copilot builds it. Once the agent is live, the Copilot reads past conversations and recommends instruction changes.
5. Trust becomes the deciding factor in every AI decision
Ashish expects enterprises to run their own AI systems, an approach he calls sovereign AI. Costs stay predictable as volume grows. Customer data stays inside the company’s own environment. The learning from the company’s own usage stays with the company. He gave the room four questions to put to every AI vendor:
Can we run your AI inside our environment?
Do improvements from our usage stay with us?
Does it integrate with what we already run?
Can we change models without re-platforming?
Friday’s closing panel added a buyer’s checklist. Rob Dwyer moderated a conversation with Jane Griener of Purple; Corinne Flanagan, Senior Manager of Training, Quality and Operations at Smartsheet [confirm title punctuation]; and Luke Jamieson, CX Evangelist at Operata. When a vendor says the product can do something, ask them to show you where. Corinne added a second check: talk to existing customers about what went wrong and how the vendor responded. Luke, speaking from the vendor side, urged buyers to push security reviews past SOC 2 to data storage, penetration testing, and ethical hacking. Corinne noted Smartsheet now buys nothing without a proof of concept. [confirm Smartsheet approval] The panel also leaned toward buying over building, pointing to the guardrails, data integrity, and maintenance that internal builds underestimate as teams get leaner.
Trust extends to customers too. Ryan Burke raised a risk that’s easy to underplan: a customer who agrees to be recorded for quality has not agreed to a profile built from that recording. Agentic commerce in regulated areas like skincare and lending forces teams to decide where AI should act at all.
What this means for 2027
Ashish described two paths for the year ahead. One leads to an open, sovereign-ready architecture, AI-forward people in every function, and a planned loop of feedback and improvement. The other leads to a single locked-in stack, tools nobody owns, and agents rebuilt every time something changes. The decisions teams make in the next 12 months set which path they’re on.
He closed Level Up by asking customers to measure Level AI by its actions over the next six months, on product, support, and the overall experience. Level AI certifications went to 34 attendees, and 31 customer teams were recognized at the Customer Awards, including Extra Space Storage as Unified CX Intelligence Leader of the Year and Zillow as AI Innovator of the Year.
