Agents weren’t ignoring the churn nudge. They were disqualifying it, in real time.
A U.S. telecom company added an AI churn-risk nudge to the tool its support agents use on live calls, and believed it was working. I spent three days on site with 12 agents and found it was flagging the wrong customers, at the wrong moment, with the wrong information. The team paused further investment and refocused on making it useful.
SituationA new AI nudge the team believed was working
In March 2025 the company launched a real-time churn nudge inside the tool its customer support agents use on every call. When the model flagged a customer as likely to leave, a card appeared prompting the agent to show appreciation and review retention offers.
The team believed the nudge was doing its job and was preparing to invest more in it. Three months in, they wanted to know how agents were actually using it.
What agents saw on screen (redrawn)
The four prototype screens I walked agents through. Real screens are confidential, so the card is redrawn.
TaskEvaluate the nudge in the moment it is used
I was brought in as the UX research consultant and ran the study on my own, from plan and interview guide to fieldwork, analysis and the stakeholder workshop. The brief was to assess five things: whether agents could find the nudge, understand it, use it easily, whether it met their expectations, and whether it actually helped.
Is the feature being used?
What happens in the moment the nudge appears, mid-call?
ActionGo to the desks, watch the calls, then reframe together
01Three days at the agents’ desks
Two screens, a live queue on the wall and the rhythm between calls were part of the finding. A video call would have missed all of it.
Split-second decisions under call pressure aren’t remembered accurately. Talking through the prototype screens surfaced their reasoning as it happened. Retrospective interviews would have been quicker, but they give the tidied-up version.
Agents are measured on call metrics, so being observed can feel like an audit. I opened every session by saying I was there to understand the tool, not to evaluate them.
02Found why agents set it aside
Where the nudge landed in a typical call
Switching to retention mid-diagnosis meant a full context reset, so most agents left it until the end, if they had time.
Every agent knew the nudge and saw it several times a day. Most still treated it as a last resort, after their own troubleshooting and knowledge base.
“To be honest, ‘review offers’ isn’t always the answer. I gotta fix that first. I’ll look at the nudge after, if I have time.”
“Likely to disconnect” often contradicted what agents could hear: customers calling about a simple issue, satisfied and friendly.
“It’s always like 70 percent of the times, it gives this wrong info. But I remember the customer was jovial and happy.”
It appeared before agents understood the problem. Jumping to offers then felt like not listening.
“If I immediately jumped to the nudge, it might feel like I’m not really listening to the customer.”
“Review offers” led to the same list agents already had in the offers section, not offers for this customer.
“It will give the same information that ‘more offers’ provide. I have seen it.”
Agents wanted help with the problem in front of them: troubleshooting steps, the right knowledge-base link, a bill explained in plain English.
“If the nudge could just tell me, ‘Hey, check this for that problem,’ that would make things a lot easier.”
After seeing it so often, agents couldn’t tell the risk levels apart, and another score on the same screen sometimes said the opposite.
“Though the information presented is definitely different, it looks the same.”
03Learned what agents already rely on
The thread I hadn’t planned turned out to matter most. Experienced agents had their own churn signals, and the nudge was competing with that judgement instead of adding to it.
How agents spot a customer who may leave
| What customers say | What agents pick up on |
|---|---|
| “I’m moving, so I want to leave.” | Calling back several times in a short span |
| “Another carrier’s offers are better.” | Frustration about outages that keeps coming back, even after a fix |
| “If you can’t fix my problem, I’m leaving.” | Long pauses, a loud or rude tone |
| “I saw an offer yesterday and now you say I can’t have it.” | Turning down offers and discounts |
| “Lots of network issues, so I want to move.” | Asking whether any bills are still pending |
The nudge against the five things we set out to assess
The nudge wasn’t being ignored. It was being disqualified, in real time: wrong customer, wrong moment, wrong information, so agents trusted their own judgement instead.
04Turned findings into decisions with the team
Unfavourable findings can stall a team. I framed them as seven opportunity areas, each with “how might we” questions and ideas, then ran a workshop with product, design and engineering to prioritise them by impact, feasibility and fit with the roadmap. Ideas were sorted into immediate fixes, medium-term design work and long-term exploration.
Seven opportunity areas from the research
ResultThe team stopped building on the wrong assumption
Is the feature being used?
How do we make it genuinely useful on a live call?
- The findings led the team to reconsider the data behind the prediction and the timing of the nudge, moving towards a more agent-centred approach.
- My recommendations included: say why a customer is flagged, show it after the issue is known, add step-by-step help and plain-English bill breakdowns, and make feedback on the nudge easy to give.
- On attribution: the decision to pause and the design changes sat with the product team. My part was the evidence, the framing and running the workshop that turned it into a plan.
ReflectionWhat I’d do differently
- Pilot the prototype first. A few screen errors (agent names that didn’t match, a score that contradicted the nudge) confused agents mid-session. A dry run would have caught them.
- Watch for test anxiety. Some agents hesitated, as if their knowledge was being checked. I’d spend more time on rapport before showing any screens.
- Pair it with usage data. Analytics would have shown low engagement, but not the moment-by-moment decision behind it. Together they would have made the case faster.
- Low adoption is a research brief, not a verdict. Agents weren’t wrong to skip it. Experienced agents know things the model doesn’t.
Client details are anonymised and the agent tool is redrawn, as real screens are confidential. The sketch at the top is an illustration. The research, quotes and outcomes are real.