Voice AI agents are no longer a pilot project. Banks, insurers, lenders and collections teams are putting automated agents on live calls, and the pace is accelerating. What is not keeping up is the layer that checks what those agents actually say, and how the customer on the other end is responding.
The proliferation we expected
When we started building Smartbot, we expected voice to follow the same path as chat: first a few cautious deployments, then an explosion of agents handling everything from balance enquiries to arrears conversations. That is now happening. Cheaper speech models, better language models and ready-made voice agent platforms mean a contact centre can stand up an automated agent in weeks, not years.
The result is a growing number of conversations with customers that no human is listening to in real time.
The guardrail gap
Most voice AI agents are tested before launch and reviewed after the fact, if at all. In between, there is very little live oversight. That matters in financial services for three reasons:
- Vulnerability is often audible before it is stated. A customer who is distressed, confused or under pressure may never say so in words. Their tone says it for them.
- Agents can sound fine and still produce a poor outcome. A response can be fluent and polite while pushing a customer towards a decision they do not understand.
- Post-call review does not scale. Sampling a small share of calls cannot evidence good outcomes across thousands of automated conversations.
Why this is a Consumer Duty question
The FCA has said it expects firms to apply existing frameworks, including the Consumer Duty and the Senior Managers Regime, to AI rather than waiting for AI-specific rules. Its Mills Review is looking at how AI, and agentic AI in particular, could reshape retail financial services through to 2030 and beyond.
Under the Consumer Duty, a firm has to be able to show that customers, including those in vulnerable circumstances, are getting good outcomes. That obligation does not change because the voice on the call is automated. If anything, it raises the bar: the firm needs evidence, not assumptions, that its agents are behaving well in live conversations.
What a real-time guardrail looks like
A guardrail for voice has to work while the call is happening. In practice that means:
- Analysing tone and text together, so a calm voice saying worrying words, or the reverse, is not missed.
- Producing a signal the agent platform can act on, such as escalating to a human, changing approach or ending an automated flow.
- Keeping a record of what was detected and when, so the firm can evidence the outcome later.
This is what Smartbot already does for human-handled calls through real-time tone and text sentiment analysis. The same approach applies to automated agents, and arguably matters more there.
An API anyone can subscribe to?
This is the idea we are now exploring. Smartbot is available today on AWS Marketplace. We are looking at whether to also offer a subscription API that any voice AI agent, whether built in-house or bought from a vendor, could call to get a live compliance signal on every conversation.
The thinking is simple. Firms should not have to rebuild their agent stack to get oversight, and agent vendors should not each have to invent their own safeguards. A shared, independent layer that sits alongside the agent would give compliance teams one consistent view across all of them.
We have not committed to a design or timeline, and we would like to hear what would make it useful. If you run a contact centre, build voice agents, or sit in compliance at a regulated firm, we would value a conversation.
Talk to us
If you would like to discuss guardrails for voice AI agents, or want to try Smartbot today, get in touch through our contact page or try it free on AWS Marketplace.
