What financial institutions need to scale AI agents

When a bank customer flags a suspicious charge, an AI agent could review their transaction history and summarise it for a human employee.

But these AI agents often improve efficiency only in specific use cases and have yet to deliver enterprise-wise productivity gains, said Jason Cao, chief executive officer of Huawei’s Digital Finance business, drawing on his work with financial institutions globally. 

The Chinese tech firm said it has served more than 7,100 financial customers globally, including 54 of the world’s top 100 banks. 

The next big test is whether financial institutions can connect AI agents securely with trusted data, business systems, and internal workflows – enabling them to move “from individual use cases to enterprise-wide productivity”, he said. 

“Scaling AI from pilot to production, and then across business processes, requires much more than model capability alone,” Cao said.

Financial institutions would need a shared, enterprise-wide digital foundation that lets them manage models, data, agents, permissions and costs together, he added. Without that, AI capabilities may stay fragmented across individual applications, making them harder to reuse, govern, and scale into production, he added.

THE BUSINESS TIMES INSIGHTS

Powering Asean's Future 2026

Find out how enterprises are navigating power constraints and sustainability targets

ACCESS THE INSIGHTS

To help financial institutions develop and deploy AI agents at scale, Huawei introduced openJiuwen, an open-source platform for multi-agent collaboration. Open source allows users to access and adapt the source code, subject to the applicable licence’s terms. 

Such a unified enterprise foundation lets firms build, run, manage and govern agents, Cao said. Because the code is open, they can adapt the platform to their evolving needs. 

This helps them to better “own their intelligence”, he said. They can choose, evaluate, and change AI models as their needs evolve, while managing their own data and knowledge and building reusable agent capabilities.

Reining token costs

Financial institutions also face a second challenge: Managing the cost and performance of artificial intelligence systems as usage grows.  

As AI use grows, the key issue is not just the unit cost of tokens, but how token consumption relates to workload, performance and business value, Cao said. 

Tokens are the unit of text processed by AI models. They are commonly used to calculate usage-based charges. 

A complex reasoning task and a customer-service interaction, for example, may each require a different model, with a different balance of capability, performance and cost. 

This calls for systematic management across the token lifecycle, from selecting the right models for the task to tracking token consumption, he added. 

“Token management should increasingly be treated as part of enterprise AI operations. The goal is to improve the value delivered by each workload, rather than simply minimising consumption,” Cao said.

This was produced in partnership with the Monetary Authority of Singapore and the Global Finance & Technology Network.

For more stories, go to bt.sg/sff26