The largest-known AI agent deployment in banking

Turning years of knowledge into day-one capability

The sheer variety of inbound support requests, which spanned thousands of query types across dozens of products and subscription tiers, meant that getting the AI agent to the right answer required a knowledge base as broad and precise as the product range itself. To automate any customer support cases, the bank first needed to share its internal knowledge with Gradient Labs.

Gradient Labs ingested the bank's extensive knowledge base, containing over 1,200 articles detailing everything from the nuances of each subscription tier to international card use, fraud procedures, payment delays, and account management. Additionally, the bank’s team added around 50 internal Notes, sources of information not publicly available in help articles, but important to provide additional context or specify action items for time-sensitive offers or campaigns.

But a bank’s knowledge base alone isn’t enough to deliver high-quality, automated customer support, so Gradient Labs also ingested thousands of historical support conversations held by the human agent teams. By reviewing these conversations, the agent automatically pulled out nuanced, context-specific guidance that previously lived only in the heads of experienced customer operations staff. These insights, called Facts in the platform, generated another 700 reference points for the agent to draw on in live conversations. 

The result of this work was an agent that could handle the breadth of the bank’s product portfolio from day one, without months of training or a lengthy ramp-up period. This set the stage for the bank’s first successful use cases and rapid scale-up.

Two high bars to clear: compliance and quality

Before going live, the bank wanted to guarantee two things: first, that the AI agent would meet the strict compliance requirements expected of a regulated financial institution, and second, that the AI agent would deliver the same quality standards as the human agent teams.

To address the compliance requirements, the bank enabled the full suite of Gradient Labs’ finance-specific guardrails. These are purpose-built for the compliance and safety requirements of the industry and cover everything from prompt injection detection and vulnerability identification to financial advice detection, sensitive information handling, and complaints flagging. 

Every single conversation is screened against these guardrails in real time, before and after the agent responds, ensuring that no reply reaches a customer unless it meets the bank's compliance standards. 

Worth noting: once the bank deployed, these guardrails ran over 9 million times, providing a critical and scalable layer of security in every conversation handled by the AI agent. 

Finally, before going live, the bank ran its own independent assessment of the agent’s outputs against its internal quality criteria. The bank holds its human agent teams to a QA score of 95%. The agent needed to match this score.

Gradient Labs delivered at 98% QA score out of the gate, and has consistently maintained a higher QA score than the human agent teams at the bank. 

Getting beyond FAQs 

In initial deployment, the bank began with low-hanging fruit, routing FAQ-related inquiries to the AI agent via chat. But these use cases aren’t the ones that tax the bandwidth of human support teams. To be truly impactful, the AI agent needs to take on more complex, nuanced inquiries without simple answers. 

So, the bank quickly expanded the types of use cases handled by the AI agent. The most common queries were related to payment troubleshooting, such as investigating a delayed payment, handling transaction disputes, recovering payments, or helping a customer understand a pending transaction. Other popular use cases covered safety and security, such as cancelling a fraud report, authorising a blocked payment, or blocking a lost card. Customers also used the agent to open a new personal or joint account, or to find out more information about loans and loan applications. 

In 2025, the bank’s most commonly handled queries related to its subscription tiers — a nuanced topic encompassing everything from upgrade and downgrade requests, billing confusion, third-party perk access, and cancellation concerns. Because subscription tiers touch so many different parts of the customer experience, there's rarely a one-size-fits-all answer. The right response instead hinges on a customer's plan history, payment status, or what they're actually trying to achieve. It's the kind of topic where customers frequently need patient, step-by-step guidance to feel confident they've landed in the right place.

The AI agent resolved all these different query types with 98% quality assurance — higher than the best human agent at the bank — and resolved over 280,000 conversations for half a million individuals. 

Reaching the next tier of complexity 

Bolstered by the results of automating more complex queries, the bank turned its focus to Procedures. In Gradient Labs, Procedures are natural-language instructions that guide the AI agent into taking real actions on a customer’s account, just as a human agent would. The true operational cost of customer support sits in complex, multi-step cases, so the bank knew it was critical to automate more of these workflows. It prioritised deployment based on its most common queries. 

One of the bank’s highest-volume inquiries was for ordering a replacement card, so the team chose this for its first automated procedure. This use case requires the AI agent to navigate account verification, check for active fraud flags, confirm delivery details, and handle edge cases, all while maintaining the seamless, digital-first experience the bank is known for.

Once live, the AI agent ran this procedure 10,000 times in under a year, fully resolving card replacement inquiries without human touch and freeing up operations teams to focus on cases where human judgment is required.