The agents followed the script perfectly, and customers still left

The problem

A 300-agent UK contact centre for a consumer subscription business had a number that would not behave. Customer satisfaction slid from the low-to-mid 80s into the mid-70s over roughly two quarters, and escalations to management crept up over the same stretch. Nobody could point to a cause, because on paper the team was doing everything asked of it. Agents were following the mandated complaint-handling script at a high rate. Compliant call after compliant call, and satisfaction still fell.

The answer was not in the satisfaction figures. It was in the recordings behind them. I listened to calls where agents kept reading the next scripted line to a customer who had already explained, twice, why that line did not apply to their situation. The quality-review process was marking every one of those calls "compliant," which meant it was rewarding the exact behaviour driving satisfaction down. The training had built compliance to a script. It had never built judgement about when to leave one.

The learner and the constraints

This is a high-turnover frontline team spanning a wide range of tenure and confidence, with no consistent supervisor coaching between calls. A three-week agent and a three-year agent were handed the same rigid script and the same instruction to follow it, with no way for either of them to recognise, in the moment, that the instruction had stopped working for the person on the line. The script was the only support they had for a real-time decision, and it was withholding permission to make one.

The question I set myself

The approach, and why

I did not write more compliance content. I built training around the decision agents were already quietly making and had nowhere to act on.

• Go to the calls, not the dashboard. I coded the QA call-review transcripts against a compliance-versus-judgement framework instead of the existing pass-or-fail adherence rubric. The coding showed agents could name the problem afterwards, in debrief, but had no in-call mechanism to act on what they already knew.
• Name the actual gap. Agents already had the underlying communication skill. What they lacked was structured, in-the-moment permission to depart from the script. That is a decision-support gap, not a knowledge gap, and it changes what you build next.
• Rehearse the decision, not the definition. The scenario module is a branching, voice-only call, so I built it around the 5 Moments of Need "apply" and "new/different" moments, the exact point where an agent has to decide, live, whether to keep reading or address the issue directly. Because it is voice-only, every scene ships with a synchronised transcript, so the practice holds up to WCAG 2.2 AA rather than assuming everyone can rely on audio alone.
• Let the wrong call play out. Each branch runs its consequence forward rather than marking a choice right or wrong. Staying on script shows the caller's frustration building; departing from it shows the call settling. The feedback explains the mechanism each time, not just the outcome.
• Replace the script with checkpoints, not a longer script. The decision-framework job aid gives agents three to five questions they can silently answer in two or three seconds without breaking call flow, mapped to the same checkpoints the scenario trains, so the module and the job aid read as one system rather than two.
• Change what QA rewards. The evaluation plan retires script-adherence as the Level 3 measure and replaces it with a new QA coding category: whether an agent exercised judgement correctly, in either direction, rather than defaulting reflexively. I named, in one line, that QA scoring itself has to move from adherence-only to judgement-inclusive coding for that measure to mean anything. I did not build the QA-assessor retraining that follows from it; that is a real next step, and naming it honestly matters more than pretending it is out of scope.

Artefacts

The outcome

The only figure with any real weight is the one the business already had: CSAT sliding from the low-to-mid 80s into the mid-70s over roughly two quarters, with escalations to management rising over the same window. I am not putting a number on the escalations. I never had one, only the direction, and I am not going to invent one to make a slide look tidier. Everything past that point, the QA judgement-coding proportion and the CSAT recovery, is a design target, not a claimed result, because no cohort has run this module yet.

What I took from it

The training was never the weak link. The quality process was. You can teach someone to read a script perfectly and still leave them with no way to help the person who has already told them, twice, that the script has stopped working. Fixing that meant designing the moment agents were allowed to think, not just the words they were allowed to say, and it meant admitting that the measurement system had to change alongside the training, or the training would be graded on the exact habit it was trying to break.

Further reading: measuring behaviour, not completion, microlearning in the flow of work.