Why one of banking’s biggest AI tests comes after launch

For financial institutions, trust is won or lost in every customer interaction. As AI becomes more embedded in customer communications, this technology has the potential to strengthen trust by making consumer requests more efficient and secure, especially high-stakes moments when customers try to update payment settings or navigate online banking alerts.

However, AI also comes with risks such as hallucinations and exposed personal data that threaten customer trust and put their finances in jeopardy. A bank transfer gone wrong, incorrect financial information shared or the failure to provide the right fraud guidance can quickly weaken the customer experience.

Banks are right to invest heavily in guardrails, but governance alone will not safeguard secure AI-powered customer interactions. That’s something most institutions find that out after launch.

What makes an agent successful post-pilot

Financial institutions already recognise the importance of protecting customer interactions. Recent research from Sinch found that 78% of financial institutions rank trust, security and compliance as a leading investment priority for their AI communications programmes. However, the 69% of financial institutions have rolled back an AI communications agent, showing that commitment and investment do not automatically translate into reliable performance or a return on that investment.

Rollbacks are often a necessary response once an institution identifies that a live agent is behaving incorrectly. While the decision to withdraw an agent can demonstrate that monitoring and governance processes are working, identifying a problem after deployment is different from preventing it from occurring.

As such, organisations must have robust governance and monitoring frameworks in place before rolling out an agent. If not, engineering teams risk treating the symptoms of an issue, rather than assessing the root cause of why a model is behaving incorrectly. It takes time to trace whether a customer is experiencing difficulties due to a fault with the model or communications infrastructure itself. Often, the organisation only discovers it once a customer has reported it.

Guardrails prevent issues from happening in the first place and the right observability mechanisms quickly identify those the safety controls couldn’t prevent, ultimately reducing the impact on the customer.

Managing AI communications agents

Many agents can be used as shortcuts to more mundane tasks such as accessing transaction information, asking questions about terms and conditions of a loan request or asking about ways to manage your personal finance portfolio. Once an AI customer support agent is live, however, it is expected to perform during the moments when customers need immediate and dependable support. For a bank, this could be during an outage affecting mobile payments.

AI failures during these high-stakes periods can spread quickly across the organisation as customers move towards contact centres and increase the workload for human customer service agents precisely when demand is already high. Engineering teams are then diverted towards diagnosing and rebuilding the agent while inaccurate, delayed or insensitive interactions create wider reputational and regulatory risks. While the impact on the support queue is easier to measure, engineering burden and reputational damage are costs of failure that often go unseen. Those costs need to be worked into wider conversations about the return of AI, alongside the time or money an agent saves when it works well.

Leaders should not treat deployment as the measure of AI success. A successful deployment is one that continues to perform reliably in production with monitoring in place to identify when an agent goes wrong before the problem reaches the customer. Meeting this standard requires both effective oversight and communications infrastructure that gives institutions clear visibility into how agents are performing, allowing them to identify and address problems before customers are affected.

Building trust into communications infrastructure

When organisations think about AI deployment, a great deal of thought goes into establishing guardrails at the model level. But the infrastructure supporting an AI agent is also fundamental to maintaining trust once it begins communicating with customers. When a bank’s AI agent discloses credit card information when it shouldn’t, for example, this is the AI infrastructure’s data masking guardrail failure in production.

Sinch’s research found that satisfaction with communications infrastructure was a stronger predictor of AI deployment confidence than investment levels, guardrail maturity or the length of time an organisation had been operating AI in production. Yet, most enterprises in the research reported shortcomings with their current providers. In fact, 42% of organisations reported insufficient reliability for AI at scale from their current provider, and 37% identified limited multichannel capabilities.

These shortcomings have specific implications for financial services, where maintaining context across a customer journey is essential. A customer discussing a suspicious payment through chat should not need to begin again when transferred to a fraud specialist by phone, while someone disclosing a vulnerable personal or financial situation should not repeatedly have to explain it because different systems cannot share the appropriate context securely.

More than half of the organisations surveyed are custom-building the ability to preserve context when customers move between channels because their platform does not provide it natively. As financial institutions introduce more agents across messaging, email and voice, continually recreating these capabilities can consume engineering resources and make governance more complex with every deployment.

The correct platforms should provide a reliable foundation of native security, monitoring and orchestration capabilities, while financial institutions add safeguards reflecting their products, customers, risk profiles and regulatory obligations.

Scaling communications that prioritise trust

When someone’s personal finance is at stake, the introduction of AI can make customer communications speedier and more accessible when they are needed most. However, delivering that promise takes more than launching a communications agent. Banks need oversight of the entire journey including the information the agent draws on, the messages it sends and the point at which a human takes over.

As AI reaches more customers and handles more complex requests, banks must keep those interactions accurate, secure and joined up. In financial services, trust is shaped by every exchange, so the technology around a model matters just as much as its capabilities.

Alejandro Murcia, Director of Specialized Sales, Sinch