Agentic AI in the contact centre, explained: how it works on Genesys Cloud

Most contact centre automation to date has been scripted. A bot flow follows a path someone designed, recognising intents and filling slots, and when the customer steps off the path it hands over. Generative AI made these bots better at understanding and answering in natural language, but they still mostly answer. An agentic AI agent is given a goal, such as rebooking a delivery or resolving a billing query, and works towards it: it decides which questions to ask, which systems to consult, which actions to take, and whether the outcome is good enough or a person should take over.

That shift matters because the value in most contact reasons lies in doing something, not in saying something. A customer who asks where their order is wants it found, and possibly redirected. An agent that can look it up, check the options against policy, make the change and confirm it has resolved the contact. One that can only explain the policy has moved the effort back to the customer. Agentic AI is therefore as much an integration and governance subject as an AI one: it is only as useful as the actions it is allowed to take, and only as safe as the limits placed on them.

Genesys Cloud CX delivers agentic AI through agentic virtual agents, configured in AI Studio and placed in Architect voice or digital bot flows. Each agent has a profile that sets its role and the guidelines it works to, the tools it may use, the knowledge it may draw on, and how a conversation starts and ends. Tools are usually Genesys Cloud data actions, the same integration mechanism your flows already use, so an agent can only reach the systems and operations you have deliberately exposed to it. Guardrails let you set what it must not do or discuss, including rules that detect attempts to override its instructions.

Genesys states that its agentic virtual agents are powered by Large Action Model technology from Scaled Cognition, positioned as action-oriented and less prone to invention than general-purpose language models. The underlying agent model is versioned: agent model v2 is the current default, with v1 deprecated from 1 October 2026, and Genesys offers an opt-in period to test a new model before it becomes the default. An agentic virtual agent can also delegate specialist tasks to external AI agents over agent-to-agent connections while keeping ownership of the customer conversation, although the external server integration this depends on has been marked as coming soon.

• Role and guidelines: who the agent is, what it is for and how it should behave
• Tools: data actions and other integrations it may call, each with defined inputs and outputs
• Knowledge: the governed content it may use to answer
• Start and exit: how it greets, when it finishes and how it hands to a person or another flow
• Guardrails: topics and actions it must refuse, and protection against attempts to override its instructions

Agentic agents earn their place in conversations that are varied, need several pieces of information or several systems, and still end in a clear outcome: changing an appointment, resolving a delivery problem, updating account details after the right checks, or diagnosing a clearly described fault. These are journeys where a scripted flow either becomes unmanageably complex or hands over too often, and where the customer is better served by a conversation than a menu.

They are a poor fit for short, predictable transactions that a scripted flow already resolves, such as taking a meter reading or confirming a balance; for decisions that need human judgement, discretion or empathy, such as bereavement, vulnerability, complaints and disputes; and for anything the organisation is not yet able to expose safely through a tool. Agentic AI is also not a substitute for missing processes: if agents today resolve a contact by improvising around a broken system, an AI agent will not do better. In our experience, most contact centres end up with a mix of scripted flows, agentic agents and people, chosen journey by journey.

An agentic agent is a new kind of colleague with access to your systems, and it should be onboarded like one. Every tool needs an interface contract: what it does, what it accepts, what it returns, how it fails and how long it may take. Guardrails need to be written with the people who own the risk, not only the people who want the capability. And the agent needs to be tested the way customers will use it, with real customer language, real accents and real background noise, on calls placed from outside the platform, including the awkward cases: the customer who changes their mind, the one who gives the wrong account number, and the one who asks for a person straight away.

After go-live, the work continues. Conversations should be reviewed and scored like any agent's, failures traced to the tool, the knowledge or the instruction that caused them, and changes released through the same flow publishing and change control as the rest of the contact centre. Callers should be told they are talking to AI, and should always have a clear route to a person, with the context of what has been done passed to the agent who takes over.

1. Define each tool's contract, including errors and timeouts
2. Agree guardrails, refusals and hand-off rules with compliance and operations owners
3. Test with real calls and real customer language before launch
4. Tell customers they are talking to AI and keep a route to a person
5. Review, score and improve conversations after go-live

Genesys meters agentic virtual agents in AI Experience tokens, per interaction rather than per minute, and charges an interaction as soon as it enters a flow that contains an agentic virtual agent block, even if the call never reaches that block. Flow design therefore affects cost as well as experience, and an agentic block belongs only in the flows that need it. Our AI Experience tokens guide explains the rates and allowances, and our voice AI cost calculator compares agentic agents with scripted flows, Genesys Virtual Agent and third-party routes per resolved call, using published list prices.

Governance matters as much as cost. An agent that acts on customers' accounts is part of your control environment: who approved its tools, who can change its instructions, how its decisions are logged and how a customer can challenge one. Regulation is catching up: the EU AI Act's transparency duties for AI that talks to people now apply, and the UK's new rules on automated decision-making apply where a decision has a legal or similarly significant effect. Our guide to the EU AI Act and UK GDPR for contact centres sets out what applies and when.

Every major contact centre platform and many AI specialists now describe their products as agentic, so the word alone tells a buyer little. Two developments are worth watching because they change what is possible rather than what it is called. The first is open standards for connecting agents to tools and to each other: Genesys documents agent-to-agent connections and the Model Context Protocol alongside its own data actions, which points to agents that can work across platforms rather than inside one. The second is more choice of voice and model: at Xperience 2026 Genesys announced a Deepgram integration as available and an ElevenLabs integration as expected, alongside its own agents.

QVCCS is a Genesys referral partner, not a reseller, and our view is a practical one. Start with contact reasons and outcomes, not with the technology; choose the simplest automation that resolves each journey; expose only the tools you can govern; and prove each agent with real calls before it meets real customers. Agentic AI is a significant step for contact centres, and the organisations that benefit most will be the ones that treat it as a change to how work is done, designed and tested with the same care as any other.