At Singapore FinTech Festival, Huawei set out a decade-long view for AI in finance, and the near-term hurdles: latency, accuracy, integration and governance.
Banks are now asking a harder question of AI: how to move from neat proofs of concept to services that hold up under real traffic, audit and legacy integration. At the recent Singapore FinTech Festival, the debate centred on making assistants usable in production: accuracy that doesn’t dip, responses that feel instant and governance that stands up to scrutiny.
SFF marked its 10th year with impressive scale: more than 70,000 participants from 142 countries, around 300 sessions and over 900 speakers under the theme ‘Technology Blueprint for the Next Decade of Finance’. AI dominated the agenda, of course, alongside tokenisation and quantum, with policymakers homing in on standards for safe deployment – governance, explainability, data lineage – while bank teams compared notes on integration and operating conditions.
Huawei used SFF 2025 to spell out how it thinks banks get from pilot to product. On stage, Huawei Digital Finance chief Jason Cao explained: “In the next 10 years we do believe everyone will have an AI super agent or super assistant,” an interface that “can understand our intention.”
The point isn’t another shiny front end; it’s moving the start of the journey to an assistant that brokers requests across specialist agents inside the bank.
Moving from app-first
Cao set that idea against the last decade’s app-first approach. If the mobile app was the front door for digital, Huawei now puts an assistant at the front of the stack. The framing also changes from tool to teammate.
“We don’t call it AI assistant, we think it’s AI colleague,” he said, adding that decisioning is moving away from fixed rule sets and raw data toward models that capture institutional knowledge and let agents act on it. It’s a people-and-systems story, not just a new button on a screen.
Cao’s Singapore vision focused on the nuts and bolts. He linked the assistant concept to two production yardsticks. That intent recognition for consumer interactions, he said, “should be at least 90 per cent… this is what today we already achieved.”
Latency needs to feel natural in the channel: “Customer-facing services – if the latency is more than two or three seconds the people will not use it… today we can achieve 1.2 seconds.”
The benchmark comes from a mobile wealth management journey built with a master–worker agent set-up and long-term memory so context carries between sessions.
Embracing AI
On adoption, Huawei says it’s seeing two tracks. “There are two approaches… The big banks… build a big AI farm… [smaller banks] focus on the high-value scenarios first and start from that and then go to the end-to-end process,” Cao said, while revealing 500-plus AI use cases across office, operations, marketing, risk and service as evidence that some banks have moved beyond proof-of-concept territory.
Much of the conversation centred on the engineering layer. Huawei showcased FinAgent Booster (FAB) as a way to close the gap between a tidy demo and a service that actually meets channel agreements. In practice, it’s a catalogue of workflows, connectors and micro-component plug-ins (MCPs), plus patterns for intent routing, multi-agent orchestration and persistent memory: designed to slot into AI-native and legacy environments rather than force a rebuild.
The company nests this in a four-layer view – compute, platforms, data/knowledge, scenario agent – so teams can reuse the same plumbing across use cases.
Two examples did the explaining. One was mobile wealth advice, where Huawei says stack tuning and hierarchical agents cut end-to-end interaction times from ~10 seconds to ~1.2 seconds while keeping context, so the assistant feels like an ongoing conversation rather than a reset.
‘No one pretended this is easy’
“AI implementation is not the easy thing,” Cao said, pointing to two snags that often stall programmes: an engineering gap (keeping accuracy and latency when traffic spikes, with proper fallbacks) and an organisational gap (reworking processes and roles so people and agents actually operate together under audit). Underneath both sit data readiness and knowledge engineering – turning tacit expert judgement into machine-usable knowledge so agents can work inside risk decisions.
On tech strategy, Huawei argued for evolution over tear-downs. Prior spend on cloud-native cores and multi-active data centres is the reliable backbone; the job now is adding an interaction/knowledge layer so assistants can reason and orchestrate across what banks already run. In the slides, that separation – compute to platform to data/knowledge to agents – is there to keep teams out of one-off builds and move them towards reusable patterns.
Ecosystem and partners
SFF also doubled as an ecosystem update. Huawei’s RongHai programme, its partner network for co-developing and delivering financial AI solutions, is a year old now, and the company says partner-led deployments are live in more than 20 countries. Around the show it flagged new cooperation agreements, including Atmaal in Saudi Arabia, with Neuxnet, Speakly AI and TrustDecision as named partners; CMA, Instadesk and MagicEngine also joined the network.
The aim is to build a “eight capability” cluster spanning model development, agent engineering, industry knowledge bases and scenario applications, so banks can reuse proven components rather than start from scratch.
Delivery-wise, Huawei leans on co-innovation: banks bring scenarios, domain knowledge and data; vendors bring compute, platforms and toolchains; the result should be reusable blueprints rather than one-offs. Cao also put some timings on when that will show up in the market: “Investors will clearly see this divergence in the next 24 to 36 months in valuations and market share,” he said, with boards advised to watch for strategy, use-case penetration, value indicators and organisation, not just pilot counts.
And the test from here? Huawei’s own markers – more than 90 per cent intent accuracy, about 1.2 seconds end-to-end latency, and portability of the engineering patterns across different policy and legacy environments – give banks something concrete to check through 2026.
If those numbers hold in production, with traceability and safe recoveries, the assistant-first idea moves from slideware to service. If they don’t, the problem isn’t vision; it’s the plumbing.