Your browser doesn’t support HTML5 audio
The gap between what enterprises want from AI-powered self-service and what they are actually getting remains significant. According to a Gartner report from November 2025, the average self-service success rate sits at just 22%, and 46% of customer experience (CX) leaders identify improving self-service as a top-three priority for 2026.
Genesys has framed its latest launch squarely around that gap, introducing what it calls the industry’s first agentic virtual agent built on large action models (LAMs). The company presents LAMs as a meaningful architectural break from earlier generations of conversational AI, rather than just an incremental upgrade.
The product, Genesys Cloud Agentic Virtual Agent, is built in partnership with Scaled Cognition, whose APT-1 LAM forms the execution engine. Genesys positions the launch as the moment AI moves from talking to doing — from understanding what a customer needs to actually resolving it across enterprise systems without human intervention.
What separates a LAM from an LLM — and why it matters
To understand the claim Genesys is making, it helps to understand what Large Language Models (LLMs) were not designed to do. LLMs are trained on text: they predict the next most probable word, which makes them powerful for conversation and explanation, but not inherently suited to executing sequences of actions across enterprise systems. Mike Szilagyi, SVP and GM of Product Management at Genesys, explains how the limitations surfaced in practice:
When it actually came to executing something from the front office to the back office and doing it in a repeatable way, that's where we found that LLMs really started to suffer. They can handle one or two steps just fine. But as you started to add more and more steps, things were more long-lived — maybe they took days to resolve. It started to lose that context and it started to be less predictable.
A large action model treats actions as first-order objects in training — which Application Programming Interface (API) to call, which business rule applies, which system to engage in sequence — rather than bolting action-taking capabilities onto a text model as an afterthought. Scaled Cognition's APT-1 was purpose-built for customer service workflows. Genesys had been developing its own approach to the problem before the two companies connected, and by the Fall of 2025, the partnership had produced results Szilagyi describes as meaningfully different:
We saw great agency around the action side and being able to do more complex things with a level of repeatability.
The practical illustration is a familiar one for contact center operators. A customer calls about an incorrect charge. An LLM-based virtual agent can locate the policy, explain the process, and identify the right team — but it cannot process the refund. The LAM-powered agent looks up the account, validates the charge against policy, and executes the refund across CRM, billing, and service systems, without a human in the loop and without requiring the customer to be transferred.
On the question of what happens when the agent hits the boundary of its defined capability — whether failure is binary or graceful — Szilagyi points to configurable escalation paths. Enterprises define which actions the agent can take and wrap those with guidelines; the agent knows its scope and routes accordingly. "Really well-designed ones," he says of the implementations Genesys has worked with so far, "the virtual agents understand their limits of what they can do."
Governance as a design principle, not an afterthought
The announcement leans heavily on governance language — guardrails, explainability, auditability, policy alignment. These terms have become common currency in enterprise AI marketing, which makes specificity valuable. At Genesys, the governance architecture operates at several levels.
The first is data transparency: model cards document what data each piece of AI uses and how, and are shared with customers' information security teams. By default, Genesys does not use customer data; enterprises must opt in explicitly. The second level is immutable prompt engineering — baseline guardrails Genesys has built in that customers cannot override, covering basic behavioral constraints that the industry has established as necessary. The third level is configurable: enterprises can layer additional guidance through AI Studio, the platform's configuration environment, defining what actions the agent can take, what language it uses, and what triggers escalation.
On action-level transparency, Szilagyi describes a decision trace that goes beyond logging:
Based on this intent that I just learned from this customer, here's the string of actions that I put into use, into purpose, and this is why I'm executing on those. So sometimes in our demo, we'll show the customers what that looks like — the level of traceability that they get on the model itself. Why is it picking these actions and what's the purpose of them and what's the information that's passed into those?
Genesys also cites compliance with ISO/IEC 42001, the AI management system standard, as external validation of its governance approach — one of the first vendors to meet those requirements, according to Szilagyi.
General availability is expected by the end of April 2026. Early adopters including M&T Bank, Banco Pichincha, a Fortune 500 healthcare company, and a Fortune 50 North American retailer are already working with the capability ahead of that date — a measured rollout that reflects Genesys's stated preference for building confidence incrementally rather than making claims ahead of the evidence.
The enterprise adoption reality
For existing Genesys Cloud customers, the barrier to a pilot is relatively low, as the platform, governance tooling, and integration points already exist. A contained use case — password resets, order status, straightforward billing queries — can be scoped and monitored in weeks. At the other end of the spectrum, a genuine contact center transformation — mapping customer journeys end-to-end, integrating with multiple enterprise backends, redefining human agent roles — is a program measured in months to years.
Szilagyi is direct that the technology challenge is frequently not the hardest part. Where he expects the real complexity to land is organizational, so when an autonomous agent resolves calls that previously went to human agents, those agents need a new job description. That is change management, not product configuration.
The Jevons' Paradox dynamic Szilagyi has articulated in previous contexts is relevant here, and he now has data to support it. Genesys analyzed customers that had been using its AI for more than a year, focusing on those at a steady state rather than actively expanding. The result:
Their conversation volume has grown over 100% since they've been using AI. But their agent volume also grew, but in the single digits — so like around 6%. So that says to me AI is working. It's meeting the unmet demand. 100% increase in conversations prior to AI would be 100% increase in your agent pool as well to handle that. So the fact that it only grows 6% while conversations are north of 100% says AI is doing its job.
The implication is that autonomous resolution does not reduce contact volume — it enables organizations to serve demand they were previously unable to meet. The 24/7 availability, multilingual capability, and absence of hold times that AI agents provide are, in Szilagyi's framing, not cost-reduction tools but demand-unlocking ones. Whether that creates net new business value or compounds operational complexity at scale is an empirical question the data is only beginning to answer.
Standards, ecosystem, and lock-in
The announcement flags planned support for Model Context Protocol (MCP) and Agent-to-Agent (A2A) standards — protocols intended to enable the Genesys agent to collaborate with AI agents from other vendors in multi-agent workflows. These are not yet available; they are on the roadmap.
Genesys is not alone in this space — NICE, Five9, and others are all advancing autonomous agent capabilities, and Salesforce and ServiceNow, both $1.5 billion investors in Genesys, are developing their own agent frameworks in Agentforce and AI Control Tower respectively. The commitment to open standards is therefore notable: rather than building a proprietary wall, Genesys is signaling that it intends to compete on the quality of its architecture and orchestration, not on lock-in.
For enterprises evaluating platforms, that matters. Configuring governance policies, guardrails, and workflow definitions across hundreds of use cases creates real switching costs regardless of vendor. MCP and A2A, if they mature as intended, give buyers more flexibility over time — and Genesys's early commitment to both is a meaningful signal about the direction of travel, even if the ecosystem is still developing.
Genesys's technical differentiator — a purpose-built LAM where actions are first-order training objects rather than LLM-plus-tools — is architecturally distinct from what competitors are deploying. Scaled Cognition was founded by Dan Roth, formerly Microsoft's corporate Vice President for conversational AI, and Dan Klein, an AI professor at UC Berkeley; the company received a strategic investment from Genesys in October 2025. Roth framed the reliability case directly in the announcement:
In the enterprise, 80% accurate is 100% useless for automation. LLMs are mainly designed to generate text, not execute tasks — and in the real world, that gap leads to hallucinations and policy drift. The foundation of trustworthy automation is super-reliability, not super-intelligence.
The early adopter data will tell that story — and Genesys appears to be building the kind of foundation that gives it a reasonable claim to answer it well.
My take
There's a measured quality to how Genesys is handling this launch that is worth noting when you consider how saturated the market is with overclaims. The governance framework Szilagyi describes is more operationally specific than most vendor announcements in this space, and the company is deliberately not overstating where it is in the journey. The Jevons' Paradox data — 100%+ conversation growth against six percent agent growth — is the most concrete signal in the whole briefing, and it deserves more attention than it typically gets in the autonomous AI conversation. If that pattern holds at scale and across verticals, it reframes the workforce displacement narrative entirely — that AI is serving the customers nobody was reaching before, not replacing human agents.
The standards commitment and the measured rollout with early adopters both read as signs of a company playing a longer game than the press release cycle typically rewards. The honest answer to whether to act now still depends on where you sit — existing Genesys customer with a contained use case, or enterprise selecting a platform architecture for the next five years. Those remain genuinely different decisions.
What I keep returning to is the trust data surfaced by Genesys's research: 70% of consumers distrust AI agents beyond basic information exchange. The enterprise case for autonomous resolution is compelling on operational and economic grounds. Whether customers will accept an AI acting autonomously on their behalf in a billing dispute, a healthcare query, or a financial account — is something else. Governance solves the enterprise's audit problem, but does not automatically solve the customer's trust problem.