Deploying agentic AI to automate customer support

For a large telecom operator, customer support is both a major cost centre and a primary driver of loyalty. Volumes are enormous and seasonal, the questions range from trivial to highly technical, and customers move fluidly between app, web, phone and messaging — expecting continuity that channel-based operations struggle to provide.

Our client's support organisation had grown through acquisition and expansion, and its knowledge had grown with it: product details, troubleshooting steps and policy lived across multiple systems, internal wikis and, critically, the heads of experienced agents. New staff took months to become productive, answers varied between agents, and average handle times reflected the effort of assembling information rather than the difficulty of the problem.

Previous attempts at automation had hardened the scepticism that often surrounds AI in the enterprise. Rule-based chatbots handled only the narrowest scripted questions and frequently frustrated customers into demanding a human. Any new approach therefore had to clear a high bar on three fronts at once: it had to be genuinely useful across real queries, it had to be safe and governable enough for a regulated operator, and it had to earn the trust of a support workforce understandably wary of being replaced.

We started by separating ambition from hype. In discovery we analysed real contact data to identify where autonomous resolution was both feasible and valuable, and — equally important — where it was not. That produced a clear-eyed scope: a set of high-volume, well-bounded intents to automate first, with everything else routed cleanly to people.

The architectural principle was that an agent is only as trustworthy as its grounding and its guardrails. Rather than relying on a model's general knowledge, we grounded the agents in the operator's own content using retrieval-augmented generation, so every answer was drawn from authoritative, current sources. Agents were given controlled tools — scoped, audited integrations into billing, account and provisioning systems — that let them act on a customer's behalf within strict permissions, not just talk.

Governance was designed in from the outset. We defined where agents could act autonomously, where they required confirmation, and where they had to hand off to a human, with full conversation logging and evaluation throughout. We also designed the operating model deliberately around the workforce: agents were positioned as a capability that removes repetitive work and assists human experts on complex cases, and frontline staff were involved in shaping and reviewing the system. That framing turned potential resistance into ownership.

At the core of the solution is a fleet of AI agents built on leading foundation models (Anthropic Claude) and orchestrated with an agent framework (LangChain). A unified knowledge layer ingests the operator's documentation and policy into a retrieval system, so agents answer from a single, governed source of truth rather than fragmented copies. When information changes, it changes once — and every channel reflects it immediately.

The agents are channel-agnostic by design, serving web, app and messaging through a consistent backend so a customer receives the same quality of help wherever they start. Controlled tool integrations allow agents to perform real tasks — checking a bill, diagnosing a connection, updating a plan — each call authenticated, permission-scoped and logged. Human-in-the-loop escalation is a first-class path: when confidence is low, policy requires it, or the customer asks, the agent hands over to a person with the full context already assembled, so the customer never repeats themselves.

Observability and evaluation underpin the whole platform. Every interaction is captured and measured against resolution, accuracy and satisfaction, feeding a continuous improvement loop and giving operations and risk teams the transparency a regulated environment demands. Analytics on this data (warehoused in Snowflake) also surface emerging issues — a spike in a particular fault, say — earlier than channel-based reporting ever could.

The agents now resolve 65% of incoming queries autonomously, end to end, without human involvement — a figure that rule-based automation had never approached. For the contacts that still reach a person, average handle time dropped by 45%, because agents handle triage and context-gathering and hand over complex cases already framed for an expert.

Customers feel the difference as consistency and availability: the same accurate answer regardless of channel or time of day, and immediate help at 2 a.m. as readily as at 2 p.m. Support is now genuinely always-on, and the variance that came from differing agent experience has largely disappeared.

The economic and organisational impact is structural rather than marginal. The operator can absorb volume growth and seasonal spikes without scaling headcount linearly, and its people have shifted toward the complex, high-value interactions where human judgement matters most. Onboarding is faster because the knowledge no longer lives only in experienced heads, and leaders have richer, real-time insight into what customers are actually contacting them about.

The decisive factor in moving agentic AI from demo to dependable production was discipline, not model choice. Grounding answers in the organisation's own knowledge, scoping tools tightly, and designing explicit guardrails and escalation paths are what made the agents safe enough for a regulated operator to trust with real customers.

Adoption is as much an organisational design problem as a technical one. By positioning agents as a capability that removes drudgery and assists experts — and by involving the workforce in shaping it — the operator avoided the resistance that derails many AI initiatives and instead built internal advocates.

Most strategically, the operator now owns a reusable AI platform: governed agents, a unified knowledge layer and an evaluation framework that extend naturally to sales, retention and internal operations. The first deployment paid for itself; the lasting value is the capability to apply governed, autonomous AI wherever the business needs it next.