Enterprise Tech Shift: Trust Bank cuts incident triage time to two minutes with AI agents and the Scaling Frontier

In an important development shaping the global Business space, Singapore’s Trust Bank has cut the time it takes to triage incidents from between 15 and 20 minutes to around two minutes using artificial intelligence (AI) agents built on Amazon Bedrock AgentCore. Recent observations, according to dispatches from NewsData.io Business & Tech Wire, point to structural shifts with notable ramifications for industry participants and analysts alike.

Singapore’s Trust Bank has cut the time it takes to triage incidents from between 15 and 20 minutes to around two minutes using artificial intelligence (AI) agents built on Amazon Bedrock AgentCore. When an incident is raised, the agents follow a runbook to pull logs and data from the bank’s systems, analyse them and post a likely cause, with their reasoning, in the incident ticket. That work is done before the on-call engineer joins the call. When the digital bank, which is backed by Standard Chartered and FairPrice Group, launched in September 2022, it ran about 50 microservices. It now runs more than 180 and pushes through more than 100 changes a week. “This is not an innovation exercise for us. This is a survival exercise for us,” Srinivas Patil, chief technology officer at Trust Bank, told Computer Weekly. “We might be 99.99% up, but since the app is the only channel for the customers, for even a single customer who is experiencing an issue on the app, the bank is down for them,” he said. “It’s not about saying nothing should go wrong. It’s about when things go wrong, how fast can we react and how fast can we resolve the issue,” said Patil. The bank measures itself on mean time to resolve (MTTR), he added. For now, the agents handle IT operations, with security to follow. From runbooks to a hybrid solution Before the agents were deployed, on-call engineers had to establish whether a problem lay in the infrastructure or a service, then work through a root cause analysis (RCA) before deciding how to mitigate it. The bank’s first attempt to solve the problem about two to three years ago was to use a set of automated runbooks. At the speed at which Trust Bank ships changes, the runbooks went stale faster than engineers could update them. A proof of concept that fed all the data to an AI model came next, but Patil said he “could not trust the outcome on a consistent basis, and that is even more problematic than a traditional runbook”. The bank took a hybrid approach, using runbooks to execute the prescribed triage steps and AI models to conduct the analysis and produce an RCA to guide engineers. AI has also sped up software development at Trust Bank, though Patil said the constraints now sit at either end: the quality of the requirements going in and the bank’s ability to run what comes out. To address this, the bank has built an AI backbone for its software development lifecycle (SDLC). Business teams use AI to write detailed requirements, including the metrics, runbooks and remediation steps when things go wrong, in a machine-readable format that the agents later refer to during an incident. If the runbook is incorrect – for example, we are not looking for the right patterns – it doesn’t matter how great the AI is, the answer is always going to be wrong Srinivas Patil, Trust Bank Recommendations, not actions Patil said the bank has been “very deliberate and conservative in ensuring that the AI is providing a recommendation, not taking an action”. Every analysis the agents post explains the reasoning and the data points behind it, so the engineer on the call can verify the work before acting on it or digging further. Runbooks are also run against a golden dataset to check that they pick up the right signals and follow the right workflow. “If the runbook is incorrect – for example, we are not looking for the right patterns – it doesn’t matter how great the AI is, the answer is always going to be wrong,” Patil said. About 65% of the RCAs the agents produce are now actionable. “The remaining 35% are really complex, hard-to-find edge cases where the engineer still needs to use their domain knowledge and expertise to identify the root causes,” he said. Trust Bank is now deploying change review agents, which look across every service that a new feature touches before it goes into production and trace those changes back to the original requirements. For instance, the agents check whether the operational basics defined by the site reliability engineering (SRE) team are in place, such as indexes and rate limits, and whether third parties have been informed of the launch, then alert the technology operations team to the likely impact. “It’s like a guardian angel that helps us with the pace of changes that we are making, so that we can validate above and beyond what humans can potentially validate,” Patil said. The bank is also experimenting with agents that watch a release for four weeks after it goes live, looking for the 1% of cases that testing misses, such as customers with unusual account set-ups or transactions. Why AgentCore Trust Bank was built on and runs nearly all of its services on Amazon Web Services (AWS). The hyperscaler’s AgentCore agentic platform gives the bank a production-level runtime for agents with built-in session isolation, identity, observability, OpenTelemetry and evaluation hooks . “We have a very small SRE team, and we don’t want them to spend their time and effort running a bespoke agent platform,” Patil said. There are three engineers in that team. In the bank’s DevOps model, incidents can be escalated to third-level engineers if need be, while a 24/7 cloud operations team has two engineers on shift at all times. Trust Bank also uses PagerDuty, its incident response system, and Sumo Logic, both of which have their own AI capabilities. “The value is in getting the data points from across all of those tools and utilising our engineering and domain expertise to come up with the right RCA,” said Patil. Take a 503 gateway timeout, for example. A monitoring tool can flag the error, but Patil wants to know what it means for the business: whether a timeout at a fraud check gateway could increase card fraud rates or whether customers making investment transactions will be affected. Watching the token bill Bedrock’s single application programming interface (API) lets the bank quickly test models from multiple suppliers, and each new agent goes to the architecture board with its model assessments, costs and benefits before deployment. For incident response, the bank uses Anthropic’s Claude Opus 5 model for reasoning and Claude Haiku for lower-order tasks, with the agent deciding which to call. Most of its models come from Anthropic, but it is also experimenting with OpenAI and open source models . “Where the token consumption shoots up is if you try to make everything an AI problem,” said Patil, who reckons about 80% of what gets labelled an AI problem is normal engineering, with AI adding value in the other 20%. He pointed to an anti-money laundering (AML) system launched last year, where standard engineering did most of the work of sourcing and preparing data, and AI was brought in only at the end to generate a structured report. Token consumption is a real issue in software development, where engineers are heavy users. But the bank does not want to stifle innovation by “arbitrarily putting in caps to say who can or can’t use AI”, Patil said, adding that it is building pipelines and training engineers to monitor where their token use is coming from instead. Still, engineers tend to reach for Opus when Haiku or Sonnet would do, Patil said, and that’s also why the bank is embedding model selection into the skills of AI agents. “Engineers’ time is ultimately money,” he said. Read more about AI in APAC APAC enterprises are consuming more tokens even as unit prices fall, and the advice on what to do about it runs from smarter caching and gated model access to renting their own infrastructure. Bumrungrad International Hospital is using Salesforce’s Agentforce to summarise and route patient correspondence , with automated appointment booking and voice agents on the cards. DBS has strung together as many as 80 agents to prepare credit approvals for big corporate clients but says the industry’s ability to police such systems is advancing at a fraction of the pace of the technology itself . At Boomi World Tour Sydney, Brickworks, Bakers Delight and Cochlear explain how they are shoring up the integration, data and governance foundations that AI agents depend on .

Market & Strategic Implications

Beyond immediate headlines, market participants are weighing secondary effects. The intersection of capital allocations, regulatory scrutiny, and shifting macroeconomic postures continues to elevate risk sensitivity across comparable assets and jurisdictions.

As further clarity emerges in upcoming briefings, institutional observers emphasize unit economics, policy enforcement, and counterparty exposure as primary barometers for long-term trajectory.