DBS holds off on letting AI agents run on their own as controls lag capability
The bank 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
DBS Bank has pushed artificial intelligence (AI) agents deep into its corporate lending workflow and given staff across the bank the tools to build their own, but it is refusing to let those agents act without a human checking their work.
The reason, said DBS chief data and transformation officer Nimish Panchmatia, is that the technology for supervising agents is nowhere near as mature as the technology for building them.
“The speed at which innovation is happening in terms of capability is, for argument’s sake, say 5x. The speed of innovation of governance and control and management of these agents is at 1x,” he said in an interview with Computer Weekly. “So, there’s a gap and we need to close this gap before we allow autonomy.”
His comments mark one of the more explicit statements yet from a major bank on the limits of agentic AI. Agentic systems differ from the generative AI (GenAI) chatbots of the past three years in that they can plan a sequence of steps and carry them out, rather than simply producing text on request.
DBS, Southeast Asia’s largest bank by assets, said in its 2025 annual report that data analytics and AI initiatives unlocked about S$1bn in economic value during the year, drawn from more than 2,000 models across over 430 use cases.
The bank’s most advanced agentic AI deployment sits inside its corporate banking business, where a chain of roughly 70 to 80 agents assembles the credit memos used to approve lending to large corporate customers.
The task has traditionally consumed days of a relationship manager’s time, with Panchmatia describing the slog of working through a 120-page annual report, a sustainability report of similar heft, and the client’s financials on top of industry analysis.
The agents now pull that material together, including external news, internal notes on the client and competitive analysis, before delivering it in a pre-formatted structure. The output is not approval-ready, he said, and the relationship manager still interrogates it through a chat interface, asking the system to justify its conclusions and produce supporting evidence.
What has surprised the bank is what the agents notice. “The recommendations that are coming out sometimes are actually really interesting – things that may not have been intuitive straight away to a human,” he said, citing cases where the system flagged what a client’s competitors in other markets were doing.
You’ve got to know what’s going on. And until you know that, you need to be careful about what you do and be more deterministic – not completely but be more deterministic in your agent build than autonomous
Nimish Panchmatia, DBS
In several instances, he added, clients responded warmly to ideas their relationship manager would not otherwise have raised.
On the customer-facing side, DBS has deployed its Joy chatbot, which provides detailed, conversational responses to a range of corporate banking queries. In early trials of the technology, which began in February 2025, the chatbot has handled over 120,000 unique chats from around 4,000 corporate clients each month, the majority of whom are small and medium-sized enterprises.
When agents take the easy way
Building sprawling agents that handle dozens of tasks in sequence, Panchmatia said, makes failures harder to trace: “Your problem – unexplainability, the blast zone – becomes bigger and bigger.” That’s also why DBS caps the number of steps any single agent chain performs, with checks applied between stages.
The deeper concern is behavioural. Agents optimise for efficiency, not compliance. “The technology is designed in such a way that it will find the easiest path. The easiest path may not be the right path,” he said.
He described a scenario in which an agent permitted to query two databases is instructed to use only one for a given task, but later discovers that using both produces a faster answer. Such cases have been documented elsewhere in the industry, he said, though not at DBS. As such, every deviation from unexpected behaviour must be flagged and the process halted for review.
“You’ve got to know what’s going on. And until you know that, you need to be careful about what you do and be more deterministic – not completely but be more deterministic in your agent build than autonomous,” Panchmatia said.
That burden falls heavier on banks than on other sectors. “Our requirements are far more stringent than, for example, the travel industry or the retail industry, because they don’t have that many regulations,” he said, adding that the industry’s tooling for governing agents remains nascent as a result.
Three layers and a control plane
DBS’s AI infrastructure comprises three layers. At the base sits what the bank calls its system of knowledge – structured and unstructured internal data plus external feeds. Above it is a “factory” or foundry where agents and reusable skills are built and run. The top layer is consumption, covering customer channels and internal workflows.
Wrapped around all three is a control plane handling identity, observability, traceability, policy enforcement, evaluation and security, complete with kill switches and governance actions to ensure agents behave as intended.
DBS also draws a hard line between employee and enterprise agents. Personal and team agents, such as Microsoft Copilot, can be assembled by staff themselves and are walled off from production systems. Enterprise agents, which touch production, are all built in-house and governed under the bank’s standard software development lifecycle controls.
To manage token costs, DBS runs its AI workloads on its private cloud, where capacity is elastic, so the constraint is consumption rather than hardware. The bank minimises token usage by holding memory at the application level and passing only incremental context to the model.
Capabilities are also rationed by role to keep consumption in check. Handing video generation to all 40,000 of the bank’s employees would send token usage soaring, Panchmatia said, when in practice only the marketing department needs it – so the capability is confined to that team, which manages it within a budget much as it would its spending on advertising agencies.
Panchmatia was dismissive of the industry’s newest vogue term, tokenomics, as just a buzzword, noting that tokens are just an additional cost item to be managed on top of electricity, water and rent in the bank’s day-to-day business. “It’s a bit overhyped,” he said.
The bank’s architecture has been model-agnostic from the outset, a decision Panchmatia said was made on the expectation that large language models would be commoditised. That has since played out in the industry, he suggested, pointing to investments that financial institutions had poured into explainability shortly before reasoning capabilities became a free commodity sparked by the emergence of players such as DeepSeek.
“We foresaw that models are going to become a utility. What we do above the model is going to become the critical part for us,” he said, referring to the harness or software scaffolding around a model to increase the chances of an agent getting things right.
Amid the rapid pace of AI developments, Panchmatia noted that the old IT investment cycle of between18 and 36 months, with paybacks expected in similar time frames, no longer applies. “You’ve got to be super nimble, super flexible and ready to pivot as soon as something new comes about,” he said.
Asked whether artificial general intelligence (AGI) was drawing closer, Panchmatia said he remained open-minded but unconvinced that humans would end up on the sidelines: “AGI or whatever people want to call it – if your door lock breaks, AI is not going to fix it; you’re going to need a carpenter.”
