Estimated reading time: 5 minutes · Last updated:
Cambridge's 2026 study indicated that 81% of the surveyed financial-services firms had adopted AI to some degree, yet only 14% described that adoption as transformational for strategy or competitive advantage. That gap highlights a practical problem: institutions can roll out models and agents quickly, but maintaining technical control, preserving business rules and demonstrating value remain difficult. Huawei unveiled Fintelligent at HUAWEI CONNECT on 18 September 2026 as a three-part platform — Agent Factory, Token Factory and Data-Knowledge Factory — intended to help banks govern agents, control token costs and retain institutional knowledge. The regulatory backdrop includes the Monetary Authority of Singapore’s AI guidelines, published on 7 October 2026, which set phased compliance milestones for 2027 and 2028.
Every financial institution should own its AI and keep the choice and evolution of AI in its own hands.
Jason Cao, CEO of Huawei Digital Finance BU
Key takeaways
- Cambridge's 2026 research found that 81% of the financial-services firms it surveyed had adopted AI in some form.
- Only 14% of firms in Cambridge’s survey said AI was transformational for their organisation’s strategy and competitive advantage.
- Huawei launched Fintelligent at HUAWEI CONNECT on 18 September 2026; its components include Agent Factory, Token Factory and Data-Knowledge Factory.
- The Monetary Authority of Singapore issued AI risk guidelines on 7 October 2026, with Sections 3–4 effective from 7 October 2027 and Sections 5–6 from 7 October 2028.
Table of contents
- Key takeaways
- What the Cambridge numbers show for banks
- How Huawei’s Fintelligent frames agent governance
- Managing cost and measuring value with tokens
- Operational tests: legacy rules, MAS guidance and real-world tools
- How adoption could play out
- What to be careful about
- Frequently asked questions
What the Cambridge numbers show for banks
Cambridge's headline figures illustrate a common adoption pattern: many institutions have added AI tools, yet few treat the change as strategic. In the survey, 81% of financial-services firms had integrated AI in some way, while only 14% regarded that integration as transformational; those paired numbers point to pilots and isolated uses rather than broad process or business-model redesigns.
That distribution fits responses from risk, operations and product teams: departments trial models for document processing, chat and analytics, but final decision-making and governance often remain manual. The result is fragmentation—multiple teams run different models, and the bank struggles to show a consolidated return on investment or to prevent divergent policy versions from persisting in production.
How Huawei’s Fintelligent frames agent governance
Huawei presented Fintelligent as a platform to keep AI development and evolution under a bank’s control. At the product core are three coordinated capabilities: an Agent Factory for building and operating agents, a Token Factory for planning and optimisation of compute use, and a Data-Knowledge Factory to supply agents with up-to-date institutional material.
Huawei says openJiuwen, the open-source agent platform that underpins the Agent Factory, enables a bank-wide marketplace where departments can publish, discover and reuse agents and skills. That reuse is the explicit response to the duplication and versioning problems Cambridge identified: by sharing assets centrally, a bank can reduce duplicated effort and ensure agents draw on the same policy and knowledge sources.
Managing cost and measuring value with tokens
Cost and model selection are practical limits on when AI moves beyond pilots. Cambridge's 2026 survey reported that 76% of large financial institutions in the sample had difficulty quantifying the returns of AI deployments; that aligns with routine engineering trade-offs — a document-heavy due diligence workflow requires larger models and far more tokens, while a single-field lookup during a call demands much less capacity.
Huawei’s Token Factory, labelled TokeNexus in its materials, is positioned to plan, produce and schedule tokens so banks can match model class and compute to the business task. Tokens are units of content processed by models; choosing the right token budget for a use case is the foundational step for cost control, and for making a clear numerator/denominator comparison when teams try to estimate time saved versus operating cost.
Operational tests: legacy rules, MAS guidance and real-world tools
Bringing agents into production forces institutions to reconcile three elements: policies embedded in legacy systems, human review workflows, and regulator expectations. DBS’s credit-memo tool illustrates this: DBS says the application serves roughly 1,500 employees worldwide and automates over 70 tasks associated with preparing those documents, while relationship managers and credit risk managers continue to approve the final output.
Regulation sharpens the obligation to govern. The Monetary Authority of Singapore issued AI risk guidelines on 7 October 2026 that require institutions to inventory AI use and apply proportionate controls, with Sections 3–4 expected from 7 October 2027 and Sections 5–6 by 7 October 2028. Those milestones turn internal governance projects into time-bound compliance workstreams for firms operating in or from Singapore.
| Capability | Purpose | Example name |
|---|---|---|
| Agent management | Build, publish and reuse agents across departments | Agent Factory / openJiuwen |
| Cost & capacity | Plan and optimise tokens, match model class to task | Token Factory / TokeNexus |
| Data & rules | Supply agents with institutional documents, rules and pipelines | Data-Knowledge Factory / Financial Agentic Data Solution |
How adoption could play out
The case for
- Centralised agent marketplaces and reusable knowledge could reduce duplicated engineering work and accelerate scale, allowing more units to move from pilot to production.
- Token planning tools that match model class to task could make cost measurement concrete and raise the share of uses that managers count as valuable.
The case against
- If governance and inventory requirements become a heavy operational lift, banks may slow rollouts and keep AI in isolated pockets to avoid compliance risk.
- Persistent fragmentation of models and policies could lock in manual review steps, keeping the fraction that sees AI as transformational low.
What to be careful about
- Fragmented agent deployment that preserves divergent business rules in production.
- Underestimated operating costs when teams choose larger models for convenience without token planning.
- Regulatory non-compliance for institutions subject to MAS expectations after the 2027 and 2028 deadlines.
The bottom line
The Cambridge numbers and Huawei’s response together show where enterprise AI work now sits: widespread technical adoption, limited strategic rework, and a growing set of governance and cost problems. Platforms that enforce reuse, expose token costs and pipeline institutional knowledge address real operational gaps, but they do not remove the need for clear ownership, human oversight and regulator-aligned inventories. For banks, the practical test is whether investments in agent marketplaces, token planning and knowledge pipelines convert scattered pilots into measurable, governed capabilities before compliance deadlines tighten.
What to watch
- Watch whether firms complete the inventories and controls required by MAS Sections 3–4 from 7 October 2027.
- Watch whether firms meet MAS Sections 5–6 by 7 October 2028.
Frequently asked questions
Why do so few firms call AI transformational despite broad use?
Cambridge’s 2026 research shows 81% of firms use AI but only 14% call it transformational; that gap reflects piecemeal deployments, limited reuse of assets and weak cost measurement rather than lack of technical capability.
What does Huawei’s Token Factory aim to solve?
Huawei pitches its Token Factory (TokeNexus) to plan, produce and optimise token usage so banks can match model size and compute to the job; this is proposed as a way to address the measurement problem Cambridge found, where 76% of large institutions struggled to show AI value.
What do MAS guidelines require and when do they take effect?
The Monetary Authority of Singapore issued AI risk guidelines on 7 October 2026 that require institutions to maintain an inventory and apply proportionate controls; expectations in Sections 3–4 take effect from 7 October 2027 and Sections 5–6 from 7 October 2028.
Related reading
