This is a sponsored blog post by AI software company ElevenLabs.
The first generation of automation in financial services was a cost decision, and customers could tell.
Decision tree chatbots and IVR menus reduced volume by deflecting it, and the experience taught a generation of policyholders and account holders to say “agent” and wait for the queue.
Institutions are adopting AI agents for a different reason:a better category of customer experience: every contact resolved on the spot, at any hour, in any channel, or handed to a person who already knows the customer and the situation – with the savings reinvested in the human service that matters most.
Admiral, one of Europe’s largest insurance groups, frames its own ambition as building the most trusted customer experience in insurance, and is explicit that this is a customer experience transformation.
That framing matters for the question every institution asks – build in-house or buy a platform – because the two goals set very different bars.
Automation built to cut cost only has to be cheaper than the queue. An agent built to carry the customer relationship has to be as good as the business’ best people, or better.
Once that is the bar, the useful question is no longer build versus buy. It is which parts of the experience only the institution can supply, and which parts are already solved.
What makes a good customer experience
What makes an agent feel trustworthy is largely invisible: turn-taking that holds up when a caller interrupts, latency low enough that pauses feel natural, background noise handling so that what a customer says is captured accurately without repeating themselves.
This orchestration layer is what delivers a natural conversation, meaningfully impacting whether customers engage with the AI agent and allow it to resolve their question or ask to be routed to a human.
Admiral did not build this layer. By starting from a platform with audio orchestration built in, the team went straight to production work on the use case itself rather than the infrastructure underneath it. The results followed: a loan settlement request that took around five minutes in the old journey now completes in roughly half the time, with customers rating the calls 4/5 or 5/5.
What Admiral built instead
Admiral concentrated its engineering on ensuring agents had the right knowledge, were programmed to follow the right deterministic workflows, and stayed compliant with insurance industry policies. That is where the work belongs – the knowledge, workflows and guardrails are Admiral’s expertise, and encoding them well is what turns a natural conversation into a resolved request.
It set the production bar at matching or beating its best human agents. It routes vulnerable callers and customers in arrears straight to a person while the team learns the edge cases.
The team describes the principle as raising the validation bar without lowering the compliance bar.
Most financial regulation is outcome-based, so this is also where the compliance case is won: what risk teams need are the right outcomes, evidenced and controlled, and the institution’s own standards are what define them.
Reaching customers, then improving in front of them
Experience is only transformed once agents are live, and in a regulated institution the constraint is rarely writing code – it is getting through review and earning customer trust.
Admiral pairs an engineer who knows the architecture with a business owner who knows the local market, and treats that pair as the deployment unit, so the agent reflects how customers in each market actually speak and the people who own the risk shaped the build.
From there, changes ship through staged rollouts measured in hours: a gap customers hit on Monday is fixed by Tuesday.
The pattern across institutions reaching production is consistent. They set the bar as a great customer experience, owned the standards and knowledge that make the experience theirs, and bought the layers where being different is impossible and being excellent is table stakes.
Admiral’s team recently walked through its production agents, testing approach, and rollout process in a live session. Watch the full session here.
For a full framework on implementation approaches and the tradeoffs between them, see our guide to building enterprise-grade AI agents.
