Financial firms struggle to scale AI agents as orchestration lags

Financial institutions are moving quickly from single chatbots to networks of autonomous software agents, but most struggle to coordinate them across their operations. A study from the Capgemini Research Institute shows that while 59 per cent of banks and insurers are piloting or deploying multi-agent artificial intelligence, only 10 per cent have achieved full orchestration across their organisation.

The global survey gathered responses from 1,100 senior executives across 14 markets. It reveals that the shift to multiple cooperating agents is creating practical headaches around costs, legacy systems and governance. Roughly half of the surveyed companies cannot reliably demonstrate a financial return on their agent initiatives, and 51 per cent report having no clear structure to measure whether these tools deliver real commercial value.

Cost controls and older systems slow progress

Unlike standalone chatbots that answer customer questions, multi-agent systems assign distinct jobs to different software entities. One agent might assess loan documents, another runs identity verification, and a third prepares approval paperwork. Making these agents talk to each other without error requires modern data systems that many established businesses do not have.

Eight in ten executives identified existing technical debt as their primary obstacle to scaling these networks. In response to unpredictable expenses, 90 per cent of firms are adopting or planning intelligent model routing. This practice automatically directs simpler queries to cheap, fast models while reserving larger, expensive models for complex analysis.

The rationale behind funding these projects is also changing. By 2028, the proportion of executives tying AI budgets directly to concrete business outcomes will grow by 141 per cent. Over the same period, funding granted purely for broad efficiency gains will drop by 42 per cent.

What this means for Australian businesses

Australian small and medium business owners are being pitched automated agent tools for customer support, accounting and lead management. The findings offer several practical lessons.

• Resist deploying chained agents until single tasks work reliably. If one simple automated task breaks down, combining multiple automated steps multiplies errors and makes troubleshooting difficult.
• Sort your data before buying new automation tools. Multi-agent workflows depend on clean customer records and accessible files, meaning tidy internal databases matter more than advanced software features.
• Track actual dollars saved instead of estimated hours. Focus on measurable targets such as faster invoice collections, fewer customer support escalations or lower software bills.
• Use model routing to cap subscription costs. Small teams should choose software that lets them run everyday processing on smaller, low-cost models rather than default frontier tools for basic administrative duties.