What Customer Service AI Actually Delivers: Five Companies, Real Numbers

Will Drewes

Founder, Fern Strategy · 4 min read

The results are in. Read them carefully.

Customer service is where AI deployments are furthest along and where the published outcomes are most concrete. Five companies have put real numbers on the board. They're worth examining closely, because the pattern matters as much as any single figure.

Bank of America: volume at scale

Bank of America's virtual assistant has handled 2 billion customer interactions and resolves 98% of queries within 44 seconds. That's not a pilot number. That's a production system absorbing the kind of volume that would require thousands of additional agents to staff manually.

The relevant detail for anyone in financial services: this is a regulated environment with strict requirements around what can be said, how it's logged, and what gets escalated. The fact that it operates at this scale means the governance layer had to be built to hold. You can't run 2 billion interactions in banking on a black box.

NIB Health Insurance: cost structure, not just efficiency

NIB, an Australian health insurer, reported $22 million in savings, a 60% reduction in customer service costs, and a 15% decrease in calls reaching live agents. Health insurance customer service sits in a quasi-regulated space: member communications, claims questions, coverage explanations. Getting those wrong has compliance consequences, not just satisfaction scores.

The 60% cost reduction is the number most headlines would lead with. The 15% call deflection is the more operationally meaningful one. It tells you what actually changed in the workflow: a meaningful share of contacts that previously required a human now don't.

Banco Itaú: three metrics that point at the same problem

Banco Itaú published a 22% reduction in cost-per-call, a 50% reduction in abandoned calls, and a 10%+ reduction in average handle time. Read those together. Abandoned calls drop when wait times drop. Handle time drops when agents have better information faster. Cost-per-call drops when both improve.

These aren't independent wins. They're the same operational bottleneck, addressed at multiple points. That's what a well-scoped deployment looks like versus a point solution.

Liberty: routing and triage, not generation

Liberty (the insurance group) used AI to classify and route support tickets by topic, urgency, and sentiment. The results: 73% decrease in first reply time, 11% faster resolution, and 9% increase in customer satisfaction.

This is worth pausing on. The 73% drop in first reply time came from routing and classification, not from AI writing responses. The agent still handled the conversation. The AI just made sure the right ticket got to the right person with the right context, faster. That's a narrower use case than most AI deployments get pitched as, and it produced the largest single-metric improvement in this set.

Fairmoney: smaller scale, same pattern

Fairmoney, a Nigerian digital lender, reported 20% faster response times and a 15% improvement in customer satisfaction after deploying AI support workflows. The numbers are more modest than the others here, but the directional consistency matters. Faster response correlates with higher satisfaction across every case in this set.

What this means for leaders in regulated industries

A few things stand out across these five.

First, the highest-impact results came from well-scoped workflows: routing, triage, deflection of routine queries. Not open-ended generation. The Liberty result is the clearest example, but NIB's call deflection number tells the same story.

Second, every company in this set operates in a regulated or quasi-regulated environment. Banking, health insurance, lending. That's not a coincidence. These industries have the highest volume of repetitive, rules-bound customer interactions, which is exactly where AI earns its keep. Stable rules, messy inputs.

Third, none of these outcomes are auditable from the outside. You can see the headline numbers, but you can't see the escalation logic, the error rates, the edge cases that got flagged, or how the system behaves when a customer asks something it wasn't trained on. That's the part that keeps compliance and operations teams up at night, and it's the part that has to be designed before deployment, not retrofitted after.

If you're evaluating a customer service AI deployment, the question to press on is: what does the audit trail look like, and who owns it when something goes wrong?