Reducing support costs through self-resolution

The situation

Ting built its reputation on strong human customer support. As support demand grew without a matching team, leadership proposed an AI-first chatbot to absorb volume. But moving too far toward automation risked undermining the very thing that set Ting apart.

I led the design work and found the real problem ran deeper.

What I noticed

A small number of issues drove a large share of support volume. These were predictable, repeatable problems with answers already in the help center.

The issue wasn’t information but discoverability. The Contact Us page made calling support the default, and self-serve was too easy to skip.

It’s not that people don’t want to use self-serve. It was just too easy to ignore.

The cross-functional workshop

Before moving into solutions, I brought together stakeholders across Knowledge, Customer Care, Marketing, Operations, and Commercial Onboarding to build a shared understanding of the problem.

This wasn’t just a design issue. It cut across teams with different priorities, and the solution had to reflect that.

It shifted the conversation. Not every issue should be automated, and we weren’t set up for an AI-first approach to work reliably yet.

Testing the AI bet

I collaborated with a UX writer and a software engineer to prototype a chatbot using both structured flows and a generative model.

Structured flows worked well as the outputs were predictable. The generative model covered more ground but hallucinated too often. In a support context, we couldn’t afford that risk.

We proposed a phased approach: start rules-based, introduce generative later.

Reframing the problem

Instead of asking how to answer more questions within a chatbot, I reframed it to:

This shifted the approach from AI-first to self-serve first. AI wasn’t ready to handle our highest-volume issues.

Designing around real moments

The redesign made a deliberate tradeoff: friction on the path to human support, in exchange for better self-serve resolution.

I removed the Contact Us page entirely and made the help center the primary entry point. From there, I designed around the three highest-volume call drivers.

Outages

Outages caused the largest support spikes, and were entirely deflectable.

I introduced a single source of truth for service status with real-time updates and alert subscriptions, then made it visible across every likely touchpoint: Google search, the chatbot, and the help center. For customers who still called in, the IVR detected their location and routed them directly to the status page, freeing agents for issues that actually needed them.

Troubleshooting

Agents were working through the same diagnostic steps on repeat. I designed the chatbot to introduce just enough friction before escalating:

• Have you checked service status?
• Want to try a quick fix?

For the help center, I tested several hierarchy approaches and UI iterations.

I landed on a simple layout that surfaced the service status page and top troubleshooting articles most prominently, optimising for resolution, not aesthetics.

Within the help articles, feedback loops were built in: dissatisfaction triggered an immediate agent handoff, and that signal fed back into the knowledge base and chatbot flows.

Mobile routing

Ting Mobile’s legacy customers were frequently misrouted: a high-friction problem with a simple fix. I added a phone number lookup step before access to support to identify their network and route them correctly. Uncommon pattern, but it solved the problem cleanly without adding meaningful friction.

Early signals

Post-launch results pointed in the right direction:

• a meaningful share of issues resolved without agent involvement
• strongest deflection in outage-related cases
• complex issues escalating as expected

Human support was being used better, not less.

Takeaway

The most important shift in this project wasn’t a design decision. It was reframing the problem from “automate more” to “shape behaviour.”

Friction, used deliberately, became a tool to slow down the path to support just enough to redirect people toward faster resolutions. Recognising AI’s limits early wasn’t a setback either. It prevented us from shipping something that would have eroded the trust Ting had spent years building.

Deflection rates are moving in the right direction, but numbers only tell part of the story. The more important questions are qualitative: where do conversations break down before resolution, what’s driving escalations that could have been avoided, and has satisfaction with support actually improved?

Those answers will shape where the strategy goes next.