Applied AI in the contact centre: start where you can measure, not where it's exciting

Start narrow, with a baseline you already trust

The programmes that stick begin with a use case small enough to measure in a single quarter, against a number the operation already reports. Two consistently deliver that: agent assist and automated quality assurance.

Agent assist, surfacing the right answer, article, or next step to a human agent while they're on the contact, works because it improves an existing metric (handle time, first-contact resolution) without putting AI directly in front of the customer. If it's wrong, an agent catches it. The blast radius is small; the upside is measurable within weeks.

Automated QA works for a different reason: most operations today score two or three percent of interactions manually and call it quality management. Moving that to near-total automated coverage isn't a marginal gain, it changes what you can see. You stop sampling and start knowing. And the baseline is unambiguous: you were scoring 3%, now you're scoring 100%, and here's what the other 97% was hiding.

Both integrate with the platforms you already run, both build the operational habit of using AI daily, and both earn the credibility to attempt the harder use cases later.

The failure mode is organizational, not technical

The most common way these programmes die is being run as a standalone "AI initiative", a separate team, a separate dashboard, a separate steering meeting. The model works; the operation never changes its behaviour around it. Agents don't trust the suggestions, supervisors don't coach against the new QA data, and within two quarters the tool is shelfware with a good demo.

Sustainable results come from embedding, not installing. That means the QA insights feed real coaching conversations, agent-assist suggestions are tuned by the people using them, and someone owns the workflow change — not just the model. AI in the contact centre is 20% model and 80% operations. Budgets are usually allocated the other way around, which is exactly why so many pilots never graduate.

A regional note

In this market there's an added trap: language and channel. A model tuned on English or on formal Bahasa Indonesia often stumbles on the mix of Indonesian, English and local expressions real customers use — and on messaging channels like WhatsApp, where the "conversation" is fragmented, informal, and full of voice notes and images. A pilot that quietly runs only on clean voice calls will look great and then disappoint the moment it meets the channel your customers actually use. Test on your messiest channel, not your cleanest one. If it survives WhatsApp, it will survive anything.

The practical sequence

Pick one contained use case tied to a metric you already report. Write down the baseline before you start — the number of arguments a programme loses later is directly proportional to how vague the baseline was at the beginning. Run it for a quarter. Insist on embedding it into daily coaching and workflow, not parking it in a side project. Then, and only then, use the credibility you've earned to reach for the harder, customer-facing use cases.

The goal isn't the most advanced AI in the room. It's a result the operation believes, delivered fast enough that the next investment is easy to approve.