Vietnam-based Techcombank serves more than 13 million retail e-banking customers and saw assets grow by 21% and net profit grow by 18% in 2025, while its cost-to-income ratio declined from 32.7% to 30.8%. The bank has invested around $1 billion in recent years in digital transformation, data capabilities, technology platforms and talent.

Techcombank's AI approach is centred on centralised governance, secure deployment and reusable architecture. Tran Hoang Quan, chief technology officer of Techcombank, said the bank is building a modular AI agent platform and redesigning parts of its software development lifecycle (SDLC) to reduce business analysis timelines, improve productivity and support broader AI adoption across the organisation.

Centralising AI governance and platform development

Different teams initially attempted to deploy AI independently, creating risks of duplication and fragmented investment. To address this, the bank established a central governance framework led by the chief executive officer (CEO), under which AI initiatives require structured approval before proceeding. AI initiatives currently span engineering, retail and corporate banking, operations, human resources (HR) and security, with a central team providing platform and governance guardrails while business units identify use cases.

Tran described security as "foundational". "All AI models, training data and outputs are hosted internally within Techcombank's environment. Nothing goes outside Techcombank," he said. This approach influenced the decision to adopt Amazon Q over alternatives that did not meet internal hosting and data protection standards.

The AI agent platform: modular, cloud-enabled and reusable

In 2025, the bank launched an initiative to develop an agentic platform to support "Virtual Squad" development teams across the SDLC. The modular AI agent platform is built using open-source frameworks including LangChain and LangFlow, combined with retrieval-augmented generation (RAG) architecture. Foundation models are hosted through Amazon Bedrock, a platform for building generative AI applications. The platform is deployed within Techcombank's secure cloud environment and integrated with internal systems.

"Key drivers include improving software development productivity by reducing design, coding and testing time through AI agents supporting roles like business analyst, quality engineer, enterprise architect, solution architect, IT project manager, and developer," Tran said.

The platform is designed for reuse. Within the business analyst capability, multiple sub-agents handle requirement analysis, impact assessment and use case generation, which can be used across other roles as well.

"Previously, building an agent with full functionality required approximately eight months. With the new platform, we can get an agent going in less than a week," Tran said. "The most important part is to get the data right."

Data architecture and real-time integration

The agent platform is designed to integrate with Techcombank's broader data ecosystem, including the enterprise data lake, Jira (for development workflow tracking), operational systems and compliance systems.

"Rather than allowing retail, marketing and business teams to maintain separate knowledge silos, product data is structured centrally so that all divisions reference a single source of truth," Tran said.

The bank is exploring real-time AI use cases that analyse system data streams to predict potential outages. While still in development, this reflects the bank's move towards streaming analytics and proactive infrastructure monitoring.

Structuring data to reduce hallucination

Tran said early hallucination issues were linked not to the model itself but to how documentation was ingested. "When you just dump data in there, you get a lot of hallucinations," he said. He cited login features across retail and corporate banking as an example: stored together in documentation, the model could not distinguish context when queried.

The team restructured documentation at the product level, clearly segmenting domains and applications. Images were converted to text. Complex nested tables were reformatted before ingestion into the RAG system.

"These structural improvements initially increased accuracy from 50% to 60%, and eventually to 70–80%. A human-in-the-loop mechanism was introduced. If users identify incorrect output, they can inject corrections directly into the knowledge base," Tran said.

This AI rollout also forced broader documentation discipline, surfacing outdated material and improving data hygiene. Model upgrades within Amazon Bedrock improved Vietnamese-English consistency, output quality and response speed.

Impact on software development and efficiency

The first major production deployment focused on business analysts within retail banking. "The pilot began in mid-June 2025 with about 20 business analysts." The agent replicates the business analyst workflow, analysing enhancement requests, conducting impact analysis, generating user stories and preparing documentation.

"Requirements documentation time for business analysts was reduced from hours to minutes. For quality engineering, test case creation and automation effort is targeted to reduce by 20%. Initial accuracy was approximately 50% but through data restructuring and iteration, this improved to approximately 70%, with a conservative range of 65–80%," Tran said.

"Business analysts now take roughly three days versus 10 days. It is still about 20–30% human involvement but that will reduce further. It provides additional requirements that the analyst may not have considered."

On adoption, the business analyst agent is deployed across retail and corporate banking, with around 30 active users. "Techcombank reports a conservative productivity uplift of 15–20%, with higher internal targets," Tran said.

"Imagine you used to have 100 people who deliver 50 user stories. Those 100 people can now deliver 100 user stories. The initiative has resulted in approximately 1,000 hours saved and an estimated 5.6 virtual full-time equivalent ," Tran said.

The objective was not to reduce the workforce, but to expand organisational productivity. The business analyst agent represents the first layer in a broader agentic model, with outputs feeding into solution architecture, development and quality engineering.

"It's not just about the business analyst. The analyst can move faster and create more. But how can you get the whole squad to move faster?" Tran said.

The solution architecture, developer and quality engineering agents are in development. "A quality engineering agent is currently in pilot, and a security agent has been implemented. Four to five additional agents are planned for 2026."," he said.

Scaling AI across the organisation

The bank has about five dedicated AI engineers and around ten data specialists. "A two-month upskilling programme has been launched to expand internal AI capability. Engineers with Java or Python backgrounds can transition into AI roles within one to two months," Tran said.

Beyond the SDLC transformation, Techcombank has deployed AI across credit analytics, personalisation and frontline enablement.

"Rather than isolated pilots, we are building a scalable, reusable AI foundation that integrates with core systems and data pipelines," he explained the strategy. The goal is to increase productivity to double or triple what we can deliver," Tran said.