Taking pressure off a contact centre by moving routine requests out of the phone queue

The challenge

Inbound call volume was growing, and a large share of it was routine: the same handful of questions, each requiring an agent to look something up and read it back.

That work occupied the queue that customers with genuinely complicated problems were waiting in. Hiring more agents would have absorbed the volume without addressing why the volume existed.

The context

The information customers were calling to get already existed in internal systems. There was simply no way for a customer to reach it themselves, so every request became a phone call by default.

Meanwhile, the customers concerned were already using a messaging platform constantly for everything else in their lives, and were not using the phone by preference.

The approach

The first step was categorising the calls rather than automating them. Once the routine requests were separated from the ones that genuinely needed judgement, it was clear the routine ones did not need intelligence at all — they needed access.

So this was not built as an AI project. It was a channel and an integration problem, and treating it as one made it faster to build, cheaper to run and far more predictable in production.

The solution

A messaging channel was integrated as a first-class customer utility, backed by automated flows that handle the common requests: identify the customer, fetch the answer from the systems that already hold it, and reply.

Anything outside those flows is handed to a person with the context already gathered, so the agent starts the conversation informed rather than starting it by asking. The flows and the internal systems share one source of truth, so a customer and an agent never see different answers.

AI & automation

Very little, deliberately — and that is the point of including this one.

The requests being handled here had known shapes and deterministic answers. A model in that path would have added cost, latency and a category of failure the business did not need, in exchange for nothing a rule could not do more reliably. The judgement that mattered was in deciding where the boundary sat between what could be answered deterministically and what still needed a person.

That boundary moved later. Speech and free-text questions genuinely do need language understanding, and that is a different piece of work — see the voice agent case study for what the same problem looks like once a model earns its place in the path.

Engineering

Integrating a third-party messaging platform into an established operational stack meant handling delivery guarantees, rate limits and identity matching between the platform and the internal customer record.

The automated flows had to fail visibly rather than silently: an unanswered customer is worse than an obviously escalated one, so anything that could not be resolved was routed to a human rather than left. Existing internal systems and CRMs were extended rather than replaced, because the operation had to keep running throughout.

Outcome

Inbound call-centre traffic came down and service response improved. Customers got routine answers immediately instead of queuing for them, and agents spent more of their time on the cases that actually needed an agent.