GoTyme’s GenAI root cause analysis cut median incident diagnosis from 34.5 minutes to 10, and restored banking services 3.5 times faster.

Inside GoTyme’s GenAI Root Cause Analysis

GoTyme is a multi-country digital banking group serving more than 22 million customers across South Africa, the Philippines, Indonesia and Hong Kong. It operates in seven countries and employs more than 2,000 people. Its product and technology development hub, GoTymeX, accounts for more than 900 engineers, designers and technology specialists based in Vietnam. Those engineers built a GenAI root cause analysis system on AWS DevOps Agent, which runs on Amazon Bedrock foundation models. The agent investigates automatically the moment it detects a problem, reaching across more than 100 Amazon Web Services accounts and 3,000 microservices.

Thirty Minutes of Log Reading Before Any Fix

Previously, engineers spent the first 30 minutes of every incident assembling context by hand. They gathered data, correlated logs, checked recent deployments and ruled out external systems. Only then could the repair itself begin. Meanwhile, alert fatigue spread across the teams, and unnecessary infrastructure costs accumulated as systems scaled without anyone understanding the underlying causes. “Our customers depend on us for their daily banking needs, and every minute of delayed diagnosis is a minute they may not be able to access their accounts,” said Hieu Ta, director of engineering at GoTymeX.

Faster Recovery for 22 Million Customers

With GenAI root cause analysis in place, median root cause identification fell from 34.5 minutes to 10 minutes. Service recovery therefore runs 3.5 times faster. The stakes behind those minutes are considerable. Amazon Web Services notes that high-impact IT outages cost financial services firms an average of $1.8 million per hour, an industry figure rather than GoTyme’s own. Other banks have put GenAI to work closer to the customer instead, such as Bank of Georgia, whose assistant answers 78% of customer queries.

Why It Matters

  • The payoff is bounded by how much of an incident is diagnosis rather than repair. Where the fix itself takes hours, shortening the hunt saves comparatively little.
  • A useful readiness signal is engineers who open every incident by assembling the same context by hand. That assembly is the part a model can absorb.
  • Diagnosis makes a sensible first agentic use case, because a wrong answer is cheap and a human still approves the fix.
  • The prerequisite is instrumentation rather than the model itself. An agent can only correlate the signals a company already collects.