Process review | telecom contact centre | June 2026
process reviewtelecom refund requesttelecom contact centreJune 2026
Telecom refund request
Bottlenecks, ranked recommendations, implementation roadmap, and expected impact.
~5 minfaster per call, from 12 minutes to 7
~70%less after-call work, with a cleaner paper trail
~15,000 hrsof agent time saved across the centre each year
executive summary
Handling a refund request spans the agent, verification, the billing systems, and retention. The current call runs about 12 minutes of handle time, plus wrap-up. Verification, investigating the cause of the charge, above-limit escalation, and after-call notes account for most of the cost.
Four steps drive the time: verifying the caller, analysing the root cause of the charge, escalating refunds above the agent limit, and logging account notes. None is the judgement itself. Each can be improved without removing the human refund decision or the agent limit.
The roadmap targets about 7 minutes of handle time, most after-call work removed, and higher first-contact resolution. Low-effort assistance comes first. Root-cause analysis and auto notes follow.
scope and approach
This review covers the refund call end to end, across high call volume. Interviews with the team captured owners, systems, timing, controls, and exceptions. Each step is rated from L0 manual to L3 agent.
The refund decision and any movement of money retain human approval and a hard agent limit. Assistance prepares and routes the work. People decide the refund.
at a glance
~12 min
handle time today
4
functions involved
4
bottlenecks
~40%
handle-time reduction target
process map
Four functions, with the eligibility decision and the goodwill exception. Bottlenecks use severity colors. Each step shows its automation level.
bottleneck analysis
Four steps account for most of the handle time and the wrap-up.
Verification, medium severity
Manual identity checks and account lookup add friction to the start of every call and repeat whenever a call is transferred.
Root-cause investigation, high severity
Working out why the charge happened, across billing, network status, and whether a warning SMS was sent, is the longest part of the call and the most inconsistent between agents.
Above-limit escalation, high severity
Refunds over the agent limit leave the call, open a ticket, and force a callback, turning a minutes-long resolution into days and a second contact.
After-call notes, medium severity
Writing up the account note adds wrap-up to every call, and the quality varies, which causes repeat calls and confusion later.
recommendations
Each step is set at the level that fits its risk, so the workflow spans the full range. Autonomous agents run the non-monetary investigation and drafting, L3. Rules run the deterministic gates and the payout, L2. A person owns every refund and goodwill decision, L0. Autonomy never touches the money.
1. Agent assist, screen pop and verification L2 ruleslow effort
opportunity
Verifying the caller and searching billing screens adds friction to every call.
approach
Authenticate on connect and auto-open the account with recent charges and open issues surfaced.
why L2
Identity and account retrieval must be deterministic and auditable under CPNI and GDPR, so rules run the gate, not a model.
impact
Removes verification friction and the manual lookup on every call.
2. AI root-cause analyzer L3 agentmedium effort
opportunity
Investigating the cause across billing, network status, and warning history is the longest part of the call.
approach
An agent reasons across billing, outages, and the warning SMS end to end, then proposes the cause and eligibility with the evidence cited.
why L3
The investigation runs without a person and is bounded to analysis. It proposes only, the refund stays a human L0 decision, so autonomy here is high-leverage and safe.
impact
Cuts investigation from minutes to seconds and makes the call consistent.
3. Eligibility and refund calculator L2 ruleslow effort
opportunity
Working out the exact eligible amount and prorating by hand is slow and error prone.
approach
A policy engine returns the exact amount, the recommended method, the reason code, and the timeline.
why L2
The amount must be exact and explainable, so a policy engine computes it rather than a model.
impact
Removes calculation errors and the manual prorate.
4. Smart escalation routing L2 ruleslow effort
opportunity
Above-limit refunds leave the agent, open a ticket, and force a callback.
approach
Auto-route the request with context to the right approver and give the customer an instant reference and an SLA callback.
why L2
Routing follows defined rules, while the approval above the limit stays a human decision.
impact
Shortens above-limit resolution and removes the manual ticket.
5. Auto call summary and notes L3 agentlow effort
opportunity
After-call note taking adds wrap-up to every call and quality varies by agent.
approach
An agent generates the CRM note and the reason code from the transcript and files them, flagging low-confidence cases for review.
why L3
A note is not a money action, so the agent files it autonomously and people review by exception rather than confirming every one.
impact
Removes most after-call work and gives a clean paper trail.
implementation roadmap
wk 1wk 4wk 8wk 12
Screen pop and verification
Eligibility calculator
AI root-cause analyzer
Auto call summary and notes
Smart escalation routing
now next later
expected impact
Average handle time
now
12 min
target
7 min
Root-cause investigation
now
4 min
target
1.5 min
After-call work
now
3 min
target
1 min
risk and change management
medium Trust in automated verification
Risk. Customers may distrust an automated identity check. Mitigation. Keep a clear fallback to agent verification and explain the step plainly.
high Automating the refund decision
Risk. An automated approval could refund against policy or miss fraud. Mitigation. Keep a human approval gate on every refund and a hard agent limit. The AI recommends, a person decides.
high Data protection and compliance
Risk. Identity and billing data fall under CPNI and GDPR. Mitigation. Enforce least-privilege access, log every access, and keep verification compliant by design.
next steps
Confirm the handle-time and after-call-work baseline and instrument the four bottleneck stages.
Deploy screen pop with verification and the eligibility calculator in the first weeks.
Pilot the AI root-cause analyzer with a group of agents before rolling it out.
Review impact at the end of the quarter against the targets in this report.
Implement the roadmap
Quotien can build the recommended automations, connect the CRM and billing systems, and support rollout. Human approval remains mandatory for every refund and any movement of money.
implementation prdrefund handling, AI involvement by stepprepared by Quotienv1.0 / June 2026
Refund handling automations
What AI does at each step of the telecom refund call, the automation level chosen for each step and why, and the controls that keep the refund decision with a person.
context and problem
A refund request runs about 12 minutes of handle time plus wrap-up, spanning the agent, billing, and retention. Across the contact centre that is roughly 15,000 agent hours a year. Four steps drive the cost: verifying the caller, investigating the root cause of the charge, escalating refunds above the agent limit, and writing the account note. None of the four is the refund decision itself.
The opportunity is to remove the preparation and investigation toil while leaving the judgement, whether to refund and how much, with a person inside a hard limit. This brief specifies that split step by step, and the level of AI involvement each step earns.
design principle, AI involvement by level
Each step is assigned the level that fits its risk, so the workflow spans the full range. The dividing line is the money, autonomy does the non-monetary work, a person owns every refund and goodwill decision.
L0 manual
The refund and goodwill decision. A person decides and acts within a hard agent limit.
L1 assisted
Request intake and call wrap. AI drafts and a person stays in control.
Root-cause investigation and note generation. The agent works autonomously within guardrails and proposes, never deciding the refund.
goals and non goals
goal
Reduce average handle time from about 12 minutes to about 7, a 40 percent reduction.
goal
Cut after-call work by about 70 percent and raise note completeness.
goal
Cut root-cause investigation from about 4 minutes to under 2.
goal
Turn above-limit resolution from days and a callback into minutes inside an SLA.
non-goal
Automating the refund or goodwill decision. A person approves every refund within the agent limit.
non-goal
Removing the human identity gate. Verification stays deterministic and compliant.
non-goal
Replacing the CRM, billing, or ticketing systems of record. Quotien orchestrates above them.
current and proposed flow
Current. Greet and verify by hand. Pull billing. Find the charge. Investigate the cause across systems. Decide eligibility. Approve or escalate. Process. Write the note. Close.
Proposed. Verify and screen-pop on connect. Surface the charge in context. The analyzer proposes the cause and eligibility with evidence while verification runs. A person decides. The calculator returns the exact amount. Routing handles above-limit cases. The note is drafted for the agent to confirm. Every movement of money keeps human approval.
AI involvement by step
Verification and account context L2 rules
behavior
Authenticate on connect through IVR and OTP with knowledge checks, then auto-open the account with the plan, recent charges, and open tickets surfaced.
why L2
Identity is a compliance gate under CPNI and GDPR. It must be deterministic, repeatable, and auditable, not probabilistic.
fallback
On failed or low-assurance authentication, hand to agent verification with the step explained to the customer.
Root-cause analyzer L3 agent
behavior
An agent reasons end to end over the disputed line item, plan, usage, outage feed, and notification logs, classifies the probable cause, checks whether the warning SMS was sent, and proposes the eligibility decision with the evidence cited.
why L3
The investigation runs autonomously within guardrails and is bounded to analysis. It proposes only and never approves a refund, so the eligibility decision stays a human L0 step. Autonomy on the work, not the money.
confidence
Each proposal carries a confidence score. Below threshold, the case routes to an agent with the evidence and no recommendation shown as an answer.
fallback
On missing data or low confidence, degrade to manual investigation with the gathered evidence attached.
Eligibility and refund calculator L2 rules
behavior
Once a person confirms the cause, a policy engine returns the exact eligible amount, the prorating, the method, the reason code, and the refund timeline.
why L2
Policy and regulatory correctness must be exact and explainable. Amounts are computed by rules, not generated by a model.
Above-limit escalation routing L2 rules
behavior
Routes the request with a full context package to the right approver, gives the customer an instant reference and an SLA callback, and removes the manual ticket.
why L2
Routing is rule driven. The approval above the limit stays a human decision.
Auto call notes and disposition L3 agent
behavior
An agent generates the structured CRM note, the reason code, and the follow-ups from the transcript and files them, flagging low-confidence cases for review.
why L3
A note is not a money action, so the agent files it autonomously and people review by exception rather than confirming every one. The time saved outweighs the cost of an occasional correction.
model and data approach
reasoning
Vendor-neutral LLM with retrieval over billing, usage, outage, and notification data. The analyzer reasons and cites, it does not compute amounts.
determinism
Amounts, prorating, reason codes, and routing run on a policy engine, so money outputs are exact and reproducible.
privacy
PII is redacted before model calls. Customer data is never used for training. Access is least-privilege and logged.
latency
Analysis runs during verification, so the recommendation is ready when the agent reaches eligibility, with no added hold time.
integrations and dependencies
crm
Account, notes, and disposition. Read and write.
billing
Charges, plan, and the refund amount and reason code.
identity
IVR and OTP for verification under CPNI and GDPR.
network and tickets
Outage feed, notification logs, and ticketing for escalations.
success metrics
primary
Average handle time 12 minutes to 7. After-call work down about 70 percent. Root-cause investigation 4 minutes to under 2. Above-limit resolution from days to inside the SLA.
quality
First-contact resolution and repeat-call rate. Cause-classification precision and recall. Agent override rate held in a healthy band, high enough to show judgement, low enough to show value.
guardrail
Refund error and reversal rate, policy-exception rate, verification fraud rate, and customer satisfaction must not regress against baseline.
acceptance criteria
Given an inbound call, when verification passes, then the account, plan, and recent charges are on screen with no manual search.
Given a disputed charge, when analysis completes above the confidence threshold, then a cause and an eligibility recommendation with cited evidence are presented for a person to confirm.
Given a low-confidence case, when analysis completes, then the call routes to an agent with the evidence and no recommendation is shown as a decision.
Given any refund, when it is applied, then a person approved it within the agent limit and the decision, evidence, and amount are logged.
Given an above-limit request, when it is escalated, then it reaches the right approver with context and the customer receives a reference and an SLA callback.
Given a completed call, when the note is generated, then it is filed to the CRM automatically, and low-confidence notes are flagged for agent review.
guardrails and edge cases
The refund and goodwill decision is L0 by design. AI never approves a refund, and a hard agent limit and a human gate apply to every movement of money. Verification stays compliant by design under CPNI and GDPR, with least-privilege access and full access logging. Any component failure degrades to an agent handoff with the gathered context, never a blocked call.
Suspected fraud, repeat disputes, vulnerable customers, partial or credit-only outcomes, and outage-driven refund spikes route to human review. Approval rates are monitored across customer segments to detect drift or bias.
rollout
phase 0
Shadow mode. The analyzer runs and its predictions are logged but not shown, to measure precision and recall before exposure.
phase 1
Pilot one team with the deterministic wins first, verification, screen-pop, and the eligibility calculator, against a defined call set.
phase 2
Enable analyzer proposals and auto notes for the pilot. Track override rate, handle time, and the guardrail metrics.
phase 3
Expand team by team behind go or no-go gates on the guardrail metrics, with a one-step rollback for each capability.
