Real-Time AI Agent Assist and Post Call Analytics

Discovery and concepts

I began by watching live recordings of customer service agents’ screens while they were on calls, and mapped their common tools and processes.

My biggest takeaway was that agents didn’t need more information - they already had 4+ windows open at once. They needed the right answer at the right moment, somewhere they could glance without losing the customer.

Supervisors had the opposite problem. They weren’t on the calls, so they were watching a dozen conversations they couldn’t hear. They didn’t need more detail either - they needed triage: which call is going wrong, why, and whether it’s worth stepping into.

Stakeholder alignment

Interactive prototypes of realistic scenarios, user journey walkthroughs and outlined competitor analysis allowed our small team to get multiple stakeholders, VP and executive alignment quickly.

Designing for AI 

The AI Agent Assist model could listen to a call, dig through everything the company knew and hand back an answer to an agent in seconds. My job was ensuring that agent could trust the output enough to actually use it, mid-sentence, with a customer waiting.

• Trust: every answer shows its source.
• Wrong answers: thumbs up and down on every suggestion, plus manual ask and search as a fallback.
• Human in the loop: suggestions inform; the agent decides what to say.

Launch and go to market

RingCentral’s own support agents tested AI Assist on real calls before launch.
Two findings changed the design:

• In the beginning, AI answers came back very lengthy, with a lot of detail. We needed to organize the information in a way that was still helpful, but not overwhelming.
• Too many AI suggestions at once became noise.

So we restructured the answer itself, leading with the short version and keeping the full detail and its source one tap away. Then we adjusted the timing and added a tiered hierarchy, so the right suggestion appeared at the right point in the call.

Post-call analytics, with a larger team

Everything AI Assist heard on a live call fed straight into ACE (formerly RingSense), RingCentral’s interaction analytics tool. 

Agent Assist was a lean team moving fast. ACE was a more established product with a larger UX team already in place. My job shifted from owning every pixel to being the bridge: carrying what we learned on live calls into their roadmap, and keeping the live and post-call experiences reading as one product.