Recovery protects trust and operations
Deterministic recovery cuts disputed debits at the source, and the cost of not having it is measured in customers who give up rather than complain. Kenyan evidence on redress is sobering. Innovations for Poverty Action's Financial Consumer Protection Survey found that only 43 percent of people who lost money in the past 12 months sought redress at all, and 9 percent of those cases were never resolved.
Apply that ratio to duplicate debits on your channel and the complaint volume in your dashboard is a fraction of the actual incidents. The rest become quiet attrition and a channel reputation you can't see in a ticket queue.
Clean recovery also changes your operations. Support stops reconstructing events from fragments when every instruction carries one reference with a full state history, and finance stops hunting matched pairs in end-of-day files.
Tests must reproduce interrupted journeys
Recovery rules that haven't been tested under real interruption are assumptions. The evidence you need is a set of controlled scenarios that show suppressed duplicates and the trace a support agent can pull afterward.
Regulated institutions are already formalizing this. Amazon Web Services describes how Nationwide Building Society integrated controlled fault injection experiments into its cloud infrastructure to demonstrate resilience to regulators.
The part worth borrowing is the framing. These teams treat failure testing as evidence they can show a supervisor, which is precisely the posture a Kenyan bank needs when the Central Bank asks why a customer was debited twice. Write your timeout scenarios as a documented suite with expected outcomes, and keep the results.
Redial tests verify duplicate suppression
Redial testing proves that a second dial cannot create a second debit. Test the customer who redials three seconds after the drop and the one who redials twenty minutes later, during pending processing and after completion.
Concurrency is the case teams skip. High-concurrency environments allow identical retries to hit the processing layer in the same millisecond, which bypasses check-then-insert logic before the first write commits.
So fire simultaneous requests with the same reference and assert that exactly one ledger entry exists afterward. A test that sends them sequentially will pass while your production code is still broken.
Delay tests expose uncertain outcomes
Slow responses create the pending states that duplicate payments feed on, so inject them deliberately. Delay the core banking response and the callback, one at a time, and confirm that the pending instruction stays blocked from resubmission.
Tooling for this is routine now. Fault injection using Toxiproxy in a pipeline can add 5 seconds of latency between a payment orchestrator and a mock gateway to validate behavior.
Latency is the cheapest fault to simulate and the closest to what Kenyan customers actually experience on a congested evening. If your retry and timeout settings have never been exercised against a slow dependency, they're configuration.
Disconnect tests preserve transaction evidence
Cut each link and check what survives. Break the connection between the USSD gateway and your application and between the mobile money platform and your callback endpoint, then verify restart behavior and the records left behind.
Safaricom built for this case. When Daraja cannot reach your ResultURL, it calls your TimeoutURL instead, which signals a problem with your server's availability.
That endpoint is frequently stubbed and forgotten. Test it properly, because it's the only signal you'll get that an instruction resolved somewhere you couldn't see, and your reconciliation window depends on catching it.
Operator timing must be verified directly
Get current production timing constraints in writing from each operator and aggregator, because the numbers differ by network. Africa's Talking documents that a Safaricom session runs 180 seconds but times out after 30 seconds of customer inactivity, while Airtel applies a 15 second application response timeout and a 60 second customer response window.
Those gaps matter directly. A recovery design tuned to a 30 second customer window behaves differently on a network that allows 60, and an orchestration call that comfortably returns in 12 seconds will be cut off by Airtel's application response limit.
Confirm session duration and response deadlines for Safaricom and Airtel as they stand today, since published figures age and aggregator settings vary by account. Assume nothing is shared across networks until someone at the operator confirms it.
EGS can review the recovery design
If you're investigating duplicate debits now or preparing a USSD launch, an independent review of the recovery design is the fastest way to find where your channel converts a timeout into a second debit. EGS assesses existing and new channels across recovery rules and the evidence your support teams can retrieve after an incident.
That review covers the specifics this article has worked through: whether your payment reference is generated independently of the session and whether retries carrying that reference are genuinely idempotent under concurrency.
Talk to EGS about USSD payment recovery and the engineering changes needed to address duplicate-payment risk.