How the role of the contact centre is changing in the era of AI

Our leadership has worked in contact centre technology since 1984, and the pattern is familiar. Automatic call distributors turned a room of telephones into a managed queue, and made the contact centre measurable for the first time. Interactive voice response and, later, speech recognition let callers serve themselves for simple requests such as balances and payments, and moved routine work away from people. Operator services showed how much could be automated at very large scale, and how quickly callers notice when automation is poorly designed. Computer telephony integration put customer data on the agent's screen. Multichannel routing brought email, chat and messaging into the same queues. Cloud platforms removed the hardware and made new capabilities available every few weeks rather than every few years.

Each wave did two things at once. It took some work away from people, and it changed the nature of the work that remained. When IVR handled balance enquiries, agents spent more of their day on problems that menus could not solve. AI follows the same pattern, but with a difference that matters. Earlier automation followed scripts that someone had written in advance. A large language model can understand loosely worded requests, hold a natural conversation and, when it is connected to the right tools, carry out the request itself. That makes it more capable than any previous wave, and it also makes design, governance and testing more important than they have ever been.

For much of its history the contact centre was treated as a cost to be contained. Success meant answering quickly, keeping calls short and moving work offshore or online where possible. That view made sense when most contacts were simple and interchangeable. It makes less sense now. Customers increasingly sort out simple things themselves, through apps, websites and automated assistants. What reaches the contact centre is more likely to be the request that went wrong elsewhere: the delivery that did not arrive, the bill that does not add up, the bereavement, the complaint, the customer about to leave.

Those are the moments where a customer decides whether to trust an organisation again. A contact centre that resolves them well protects revenue and reputation; one that handles them badly undoes the work of marketing and product teams. AI sharpens this shift. As automation takes more of the routine volume, the human part of the contact centre becomes concentrated on the conversations with the most at stake, and the automated part becomes, for many customers, the first and sometimes only experience of the brand. Both halves now carry the relationship, which is why the contact centre deserves board-level attention rather than only a cost line.

Most AI in contact centres so far has been AI that answers. A virtual agent or chat assistant grounded in approved knowledge can tell a customer the returns policy or the opening hours, and that has real value. But the requests that matter most usually need something to be done: an address changed, a delivery rearranged, a subscription moved, a payment plan set up. Answering is not enough when the customer needs an outcome.

Agentic AI is AI that acts. It works out what the customer needs, decides which steps to take and carries them out through tools that reach your systems of record, such as CRM, billing and ordering. In Genesys Cloud CX, agentic virtual agents are created in Genesys Cloud AI Studio, and their tools are data actions: governed, permissioned integrations to web services and back-end systems. Genesys has also announced support for open protocols, including the Model Context Protocol (MCP), for connecting AI agents to external tools and data. Whatever the route, the principle is the same. The AI can only do what its tools allow, so each tool needs a defined purpose, validated inputs, least-privilege access, an audit trail and clear rules on when to confirm with the customer or hand over to a person.

For customers, the most visible change should be self-service that finishes the job. A customer who explains a house move once, in their own words, and has their services moved on the right date has had a better experience than one who navigated a menu to reach a queue. The second change is continuity. Context from an automated conversation should travel with the customer when they switch from messaging to voice, or from a virtual agent to a person, so nobody has to repeat themselves. Customers also need to know when they are dealing with AI, and to be able to reach a person when the situation calls for it.

For agents, the change is more mixed, and it is worth being plain about it. As AI resolves more simple requests, there will be fewer of them for people to handle, and some organisations will need fewer people for that work over time. The contacts that remain will be harder: complex, emotional, unusual or high-value. AI assistance helps here. Genesys Agent Copilot, for example, determines customer intent and suggests relevant next best actions to the agent, and can produce conversation summaries, which reduces searching and after-call work. Agents' skills shift towards judgement, empathy, negotiation and spotting when something is wrong, including fraud. New roles also appear around the AI itself: conversation design, knowledge curation, testing and reviewing what automated agents did and why.

Supervisors move from watching queues towards coaching people through harder conversations and overseeing automated ones. Genesys describes Genesys Cloud Copilot as also supporting supervisors, for example by summarising interactions so they can spend more time on coaching. Supervisors increasingly need to review what a virtual agent did, not only what an agent said, and to recognise patterns such as rising hand-offs or repeated tool errors.

Workforce planning becomes harder rather than easier. When automation absorbs simple contacts, the remaining work has longer and more variable handle times, and forecasts built on history need revisiting. Planners need to model automated and human demand together, account for the effect of a virtual agent change on the queues behind it, and plan time for training as skills shift.

Quality and compliance gain a new subject: the AI itself. The UK Information Commissioner's Office publishes guidance on AI and data protection covering accountability, transparency, lawfulness, fairness, security and data minimisation, and individual rights; it notes that the guidance is under review following the Data (Use and Access) Act. In practice that means telling customers when they are talking to AI, offering a route to a person, limiting what personal data AI can see and retain, explaining automated decisions that affect people, and keeping records that show who or what took each action.

Average handle time served the cost-centre era well, but on its own it now gives misleading signals. When AI removes the quick contacts, average handle time for people rises even though the operation is working better. A virtual agent that keeps customers in conversation without resolving anything can look like success on a containment measure while it frustrates them. Leaders need measures that describe outcomes, not just activity, and that cover the automated and human parts of the contact centre together.

Three risks recur. The first is poor data. AI that acts on inaccurate customer records or out-of-date knowledge will make mistakes faster and at greater scale than people would, so data quality and knowledge governance come before automation. The second is unclear accountability. For every action an AI agent can take, someone must decide on whose authority it is taken, who owns the tool, and who answers when it goes wrong. The third is over-automation: pushing AI into conversations that need human judgement, or making it hard to reach a person, which saves effort in the short term and costs trust in the long term.

We approach AI the way we approach any contact centre change, with a governed method and quality gates. Discovery uses contact-reason, volume and handle-time analysis to find requests AI can resolve end to end. Each tool gets an interface control document defining inputs, outputs, errors, authority and audit. The dialogue design specification sets confirmation steps and hand-offs. Negative and failure-path testing covers timeouts, missing data, ambiguous customers and attempted misuse before real users accept it. Authority is widened only as evidence supports it.

1. Choose a small number of high-volume, well-understood requests with clear success criteria.
2. Fix the data and knowledge those requests depend on before automating them.
3. Agree an authority model: what AI may do, on whose authority, and where a person must approve.
4. Build narrow, auditable tools rather than broad access to systems.
5. Design the hand-off to people as carefully as the automation itself.
6. Test failure paths and misuse, then pilot with monitoring in place.
7. Measure outcomes, review audit records and widen scope step by step.