Techcombank engineers reusable AI agent platform to speed up software program growth

Vietnam-based Techcombank serves greater than 13 million retail e-banking clients and noticed property develop by 21% and web revenue develop by 18% in 2025, whereas its cost-to-income ratio declined from 32.7% to 30.8%. The financial institution has invested round $1 billion lately in digital transformation, information capabilities, expertise platforms and expertise.

Techcombank’s AI strategy is centred on centralised governance, safe deployment and reusable structure. Tran Hoang Quan, chief expertise officer of Techcombank, stated the financial institution is constructing a modular AI agent platform and redesigning elements of its software program growth lifecycle (SDLC) to scale back enterprise evaluation timelines, enhance productiveness and assist broader AI adoption throughout the organisation.

Centralising AI governance and platform growth

Completely different groups initially tried to deploy AI independently, creating dangers of duplication and fragmented funding. To handle this, the financial institution established a central governance framework led by the chief govt officer (CEO), underneath which AI initiatives require structured approval earlier than continuing. AI initiatives presently span engineering, retail and company banking, operations, human assets (HR) and safety, with a central staff offering platform and governance guardrails whereas enterprise items establish use instances.

Tran described safety as “foundational”. “All AI fashions, coaching information and outputs are hosted internally inside Techcombank’s surroundings. Nothing goes exterior Techcombank,” he stated. This strategy influenced the choice to undertake Amazon Q over alternate options that didn’t meet inner internet hosting and information safety requirements.

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

In 2025, the financial institution launched an initiative to develop an agentic platform to assist “Digital Squad” growth groups throughout the SDLC. The modular AI agent platform is constructed utilizing open-source frameworks together with LangChain and LangFlow, mixed with retrieval-augmented technology (RAG) structure. Basis fashions are hosted via Amazon Bedrock, a platform for constructing generative AI purposes. The platform is deployed inside Techcombank’s safe cloud surroundings and built-in with inner methods.

“Key drivers embody bettering software program growth productiveness by lowering design, coding and testing time via AI brokers supporting roles like enterprise analyst, high quality engineer, enterprise architect, answer architect, IT challenge supervisor, and developer,” Tran stated.

The platform is designed for reuse. Throughout the enterprise analyst functionality, a number of sub-agents deal with requirement evaluation, influence evaluation and use case technology, which can be utilized throughout different roles as nicely.

“Beforehand, constructing an agent with full performance required roughly eight months. With the brand new platform, we are able to get an agent moving into lower than per week,” Tran stated. “Crucial half is to get the info proper.”

Information structure and real-time integration

The agent platform is designed to combine with Techcombank’s broader information ecosystem, together with the enterprise information lake, Jira (for growth workflow monitoring), operational methods and compliance methods.

“Quite than permitting retail, advertising and marketing and enterprise groups to take care of separate data silos, product information is structured centrally so that every one divisions reference a single supply of fact,” Tran stated.

The financial institution is exploring real-time AI use instances that analyse system information streams to foretell potential outages. Whereas nonetheless in growth, this displays the financial institution’s transfer in direction of streaming analytics and proactive infrastructure monitoring.

Structuring information to scale back hallucination

Tran stated early hallucination points had been linked to not the mannequin itself however to how documentation was ingested. “While you simply dump information in there, you get a number of hallucinations,” he stated. He cited login options throughout retail and company banking for instance: saved collectively in documentation, the mannequin couldn’t distinguish context when queried.

The staff restructured documentation on the product stage, clearly segmenting domains and purposes. Pictures had been transformed to textual content. Advanced nested tables had been reformatted earlier than ingestion into the RAG system.

“These structural enhancements initially elevated accuracy from 50% to 60%, and finally to 70–80%. A human-in-the-loop mechanism was launched. If customers establish incorrect output, they will inject corrections immediately into the data base,” Tran stated.

This AI rollout additionally compelled broader documentation self-discipline, surfacing outdated materials and bettering information hygiene. Mannequin upgrades inside Amazon Bedrock improved Vietnamese-English consistency, output high quality and response velocity.

Impression on software program growth and effectivity

The primary main manufacturing deployment centered on enterprise analysts inside retail banking. “The pilot started in mid-June 2025 with about 20 enterprise analysts.” The agent replicates the enterprise analyst workflow, analysing enhancement requests, conducting influence evaluation, producing person tales and making ready documentation.

“Necessities documentation time for enterprise analysts was diminished from hours to minutes. For high quality engineering, take a look at case creation and automation effort is focused to scale back by 20%. Preliminary accuracy was roughly 50% however via information restructuring and iteration, this improved to roughly 70%, with a conservative vary of 65–80%,” Tran stated.

“Enterprise analysts now take roughly three days versus 10 days. It’s nonetheless about 20–30% human involvement however that can cut back additional. It supplies further necessities that the analyst might not have thought-about.”

On adoption, the enterprise analyst agent is deployed throughout retail and company banking, with round 30 energetic customers. “Techcombank studies a conservative productiveness uplift of 15–20%, with increased inner targets,” Tran stated.

“Think about you used to have 100 individuals who ship 50 person tales. These 100 individuals can now ship 100 person tales. The initiative has resulted in roughly 1,000 hours saved and an estimated 5.6 digital full-time equal ,” Tran stated.

The target was to not cut back the workforce, however to develop organisational productiveness. The enterprise analyst agent represents the primary layer in a broader agentic mannequin, with outputs feeding into answer structure, growth and high quality engineering.

“It is not simply concerning the enterprise analyst. The analyst can transfer sooner and create extra. However how are you going to get the entire squad to maneuver sooner?” Tran stated.

The answer structure, developer and high quality engineering brokers are in growth. “A high quality engineering agent is presently in pilot, and a safety agent has been applied. 4 to 5 further brokers are deliberate for 2026.”,” he stated.

Scaling AI throughout the organisation

The financial institution has about 5 devoted AI engineers and round ten information specialists. “A two-month upskilling programme has been launched to develop inner AI functionality. Engineers with Java or Python backgrounds can transition into AI roles inside one to 2 months,” Tran stated.

Past the SDLC transformation, Techcombank has deployed AI throughout credit score analytics, personalisation and frontline enablement.

“Quite than remoted pilots, we’re constructing a scalable, reusable AI basis that integrates with core methods and information pipelines,” he defined the technique. The aim is to extend productiveness to double or triple what we are able to ship,” Tran stated.