How Sahel Transfer recovered 18 points of KYC completion in 11 weeks
CRO
A worked example, not a client result. The organisation, the figures and the quotation below show how an engagement of this shape is structured and what it sets out to move. They are not outcomes achieved for a named client.
Kora PaymentsAI & Automation
Kora Payments cut support contacts by 58% with voice and WhatsApp agents
Three repeat questions dominated the queue. Absorbing them before a human saw them freed the team for the cases that actually needed judgement.
−58%support contacts
<40saverage resolution
−58%SUPPORT CONTACTS
<40sAVERAGE RESOLUTION
+31%KYC RECOVERY
9 wksTO LIVE
Measured over 9 weeks against a matched prior period. Attribution and confidence notes available on request.
The situation
Kora Payments ran forty agent counters and a growing app, and the support team had not grown with either. Peak periods produced queues measured in hours, and the escalation path from an agent counter to head office was a phone call to someone's mobile.
The team's own view was that support volume was the cost of doing business in a corridor where senders are anxious and transfers are occasionally held. That is half right. Anxiety is inherent. Answering the same three questions manually is not.
They had looked at chatbots once, decided the technology was not ready for a regulated financial context, and shelved it. The concern was legitimate. The conclusion was two years out of date.
What we found
We classified twelve weeks of contact reasons before proposing anything. Three questions accounted for sixty-one percent of all inbound volume: where is my money, why was my transfer held, and what is my rate today.
None of the three requires judgement. All three require access to a system the sender cannot see. That is an automation problem, not a service problem.
The fourth finding was more interesting. A meaningful share of contacts came from senders stuck mid-verification who had given up on the app and phoned instead. Those were being handled as support tickets rather than as recoverable revenue.
Funnel and diagnostic table
| Stage or area | Position | Cumulative or detail | Read |
|---|---|---|---|
| Where is my money | 34% | No | Transfer status lookup |
| Why was my transfer held | 16% | Partial | Status plus escalation rule |
| What is my rate today | 11% | No | Live rate lookup |
| KYC and verification help | 14% | Partial | Guided, with human fallback |
| Complaints and disputes | 9% | Never | Human only, always |
| Everything else | 16% | Case by case | Classification and routing |
What changed
1. Classified every contact reason before building anything, so automation targeted volume rather than what was easiest to build.
2. Built an inbound voice agent handling transfer status, rate enquiries and hold explanations, with authentication before any account data is spoken.
3. Added WhatsApp Business API for the same three questions, because a large share of the corridor's senders reply there and nowhere else.
4. Wrote strict escalation rules: anything involving a complaint, a dispute, a compliance hold or a distressed sender routes to a human immediately with full context attached.
5. Routed stuck verifications out of the support queue and into a recovery flow, which turned a cost line into a revenue one.
The escalation rules took longer to agree than the automation took to build, and that was the right allocation of effort. In a category where a held transfer can mean a family does not eat, the boundary between what a machine handles and what a person handles is not a technical decision.
Nobody lost their job. The support team's contact volume more than halved and their average handling time on the remaining cases went up, because those were the cases that deserved the time.
The team stopped answering where is my money forty times a day and started handling the cases that actually needed a person.
— Kwame Mensah, Operations Lead, Kora Payments
How the work ran
Contact reason classification
Twelve weeks of tickets categorised before a single automation was specified.
Voice and WhatsApp agent flow
Three questions handled end to end, with authentication and a documented escalation path.
Verification recovery routing
Stuck senders moved out of the support queue and into a recovery sequence.
What moved
| Metric | Before | After | Change |
|---|---|---|---|
| Total support contacts | Baseline | −58% | Improved |
| Average resolution time, automated | Not applicable | Under 40 seconds | New |
| Average handling time, human cases | Baseline | +26% | Deliberate |
| Verification recoveries from support | Near zero | +31% | New revenue |
| Escalation accuracy | Not measured | 96% | Audited weekly |
| Support cost per completed transfer | Baseline | −47% | Improved |
Mandatory line beneath: “Measured over 9 weeks against a matched prior period. Attribution and confidence notes available on request.”
The engagement in brief
Scope and shape
- MTO with agent network
- UK → Ghana
- 9 weeks
- AI & Automation
We were worried it would feel robotic to people who are already anxious. The escalation rules are what made it work.
— Kwame Mensah, Operations Lead, Kora Payments
Same funnel, different corridor?
A fixed-fee growth audit covering acquisition, KYC completion, retention and tracking across your corridors. You keep the roadmap whether or not we work together.