AI Talk: Re-engineering banking workflows with AI-enabled decision-making 

 Views on news  
AI firms such as Anthropic and OpenAI are increasingly shifting to token-based pricing that charges customers based on usage instead of a subscription-based service. As a result, engineers at Deutsche Bank are allocated quotas for tokens ​and can request additional ​capacity but have ⁠to demonstrate value, with learnings then shared across the organisation. We don’t want to slow people ​down and want them to keep going, but we also want ​to get ⁠a return," said Denis Roux, CIO, investment ​bank at Deutsche Bank. One approach to saving costs is what’s referred to as enterprise memory – a cache of answers to questions that engineers and bankers ask repeatedly.   

 

AI powering decision making and value creation 

The AI capability that brings the most value in banking is the rapid assembly of information and its application across various systems. Conversations now go beyond efficiency and extend to agentic AI, which excels at connecting different data sources in real time and understanding semantic relevance.

 

Some tasks such as those related to compliance, also benefit from the fact that there is less variance between agentic than human output. As the technology develops, agentic AI will increasingly automate end-to-end workflows. However, the line must be drawn between use cases where agents can take action and those where they are only leveraged for recommendations.  

 

Agentic AI is well suited to automating the early stages of decision making, where an event or a major change can trigger an agentic process. Glean, for example, can connect many different data sources and by combining the event-driven change signals with user context, can take the user to a starting point which feels personalised and actionable. With automated workflows, trust and the traceability of actions to data sources are key and this is where governance must come into play too.

 

To build trust, businesses need an enterprise-grade framework for agentic AI that understands permissions, people, different data sets and which can operate in real time. The main challenge is to maintain this framework and refresh it when new data becomes available. As users keep checking AI outcomes manually and find that they are valid, their confidence in using these systems will increase as well.  

 

The success of agentic AI should be measured through metrics such as human impact, organisational readiness, business value, RoI and scalability. Real life demonstrations of how agentic AI can improve the business user experience can go a long way in building momentum behind deployments. You must also strike a balance between embedding AI into existing workflows that employees are already familiar with and creating new, AI-native ones.

 

However, you must make sure you don’t reinvent the wheel, building your specific customer experiences and new products on top of the infrastructure that companies specialising in data management, connectors or semantic layers can provide. Areas where AI-native workflows have brought benefits include pitch creation, model creation, generating memos and signal monitoring, chaining into idea creation and action. AI technology is now mature enough to be used in GRC too and forinternal audits, as well as for population control testing, where technology can remove risk from internal processes.  

 

The panel’s advice 

• AI is moving from an information to a decision supporting tool.   
• Don’t trust the model but the evidence.  
• Don’t make technology decisions without involving business teams.  
• Start by decomposing legacy workflows and find the friction points that can be automated away.  
• Use the data in your systems to personalise AI outputs. 
• Don’t just adopt AI but redesign workflows around it. Always work cross-functionally in AI projects.