Case study · Training & enablement
Training built from what QA kept finding.
For a wireless-carrier support team, I turned ten months of QA evaluations into a training deck, a call-flow job aid, and a tool agents use on live calls. Then I measured what changed.
- QA evaluations
- 300 across 46 agents
- Average QA score
- ~42% → ~73%
- Team NPS
- 55 → 70
01 / Operating context
A support team handling everything a carrier customer can call about.
At OneSupport, my team supports a regional wireless carrier’s customers: billing and payments, plans, device and network troubleshooting, and international travel. I own coaching and quality for the team, and I scored every QA evaluation in this case myself: 300 calls across 46 agents, December 2025 through September 2026.
Everything on this page is de-identified. The carrier isn’t named, internal system names are replaced with generic ones, and the sample bill in the deck is invented.
02 / Diagnosis
The gap was connection, not process.
In the first three months, 91% of evaluations missed the rapport question and 85% missed an empathy statement. Agents usually knew the procedure. What they skipped were the moments that make a customer feel heard.
That created a paradox I built the training around: an agent can read the script, satisfy most of the checklist, and still leave the customer feeling dismissed. The goal was language that meets the audit requirement and lands with the customer.
03 / Training
A diagnostic playbook for getting to the root.
The training deck teaches agents to look past the surface complaint (“my bill is too high”) to the cause underneath: billing math, network realities, or silence the customer reads as indifference. Six core slides are below, rebuilt for the web from my original deck. The customer bill in the original is replaced with a sample.
04 / On-call tools
Help in front of agents while the customer is on the line.
Training fades if nothing supports it during the call. I built two pieces of performance support: a job aid for the structure of the call, and a probing tool for how deep to dig.
Job aid · 5-page PDF
Five-phase customer call flow
One page per phase. Each gives the call sequence, an order of operations, example language for two common scenarios, and the coaching watchouts QA flagged most.
- 1Greet & respond
Hear the reason for the call, then offer help and acknowledge the concern in the same response.
- 2Verify & explain
Protect the account while explaining why each verification step is needed.
- 3Build rapport
Use the customer’s history as a brief bridge into discovery.
- 4Resolve & retain
Find the underlying cause, offer accurate options, and document during the call.
- 5Recap & close
Summarize, check satisfaction explicitly, and close with appreciation.
Live-call tool · interactive
Probing & troubleshooting cockpit
358 probing questions across 50 call situations, each paired with its troubleshooting steps. Agents search it mid-call, copy the exact wording, open an empathy and de-escalation reference, and build the root-cause confirmation recap as the customer talks. It works fully below; try searching “hotspot” or “autopay”.
05 / Measured result
~42% → ~73%
average QA score, first three months vs. the two most recent
Show as table
| Month | Average QA score | Evaluations |
|---|---|---|
| December 2025 | 41.5% | 44 |
| January 2026 | 48.0% | 66 |
| February 2026 | 46.1% | 66 |
| March 2026 | — | 0 |
| April 2026 | 45.0% | 33 |
| May 2026 | 43.3% | 21 |
| June 2026 | 52.3% | 22 |
| July 2026 | 52.6% | 17 |
| August 2026 | 72.7% | 15 |
| September 2026 (through Sep 25) | 72.8% | 16 |
The behaviors the training targeted moved the most. Rapport and empathy misses fell by more than half. Dead air barely moved (29% → 26%), so it is the next thing to work on.
- Dec 2025 – Feb 2026 (n=176)
- Aug – Sep 2026 (n=31)
Show as table
| Behavior missed | Dec 2025 – Feb 2026 | Aug – Sep 2026 |
|---|---|---|
| Rapport question | 91% | 32% |
| Empathy statement | 85% | 35% |
| Willingness-to-help statement | 41% | 10% |
| Dead air / long silence | 29% | 26% |
| Branded close | 29% | 16% |
55 → 70
Team NPS, most recent 30 days compared with the period before.
What these numbers don’t prove. I scored every evaluation myself, and the two most recent months have 15–16 evaluations each. March has no data. Other things changed over the same period, and 30 days is a short window for NPS. The movement lines up with the training and tools, but I don’t claim they caused all of it.
06 / Why it matters
The same loop I use for AI enablement.
Find the constraint in the data, design the intervention, put it into the workflow where people actually work, and measure the result honestly. Here, the intervention was training and performance support rather than a model. The discipline is the same one behind my QA automation work.