The "finally paying off" narrative has arrived in customer support AI — and for a specific tier of deployments, it is accurate. Bank of America is a useful anchor: Erica, its AI financial assistant launched in 2018, handled roughly 700 million interactions with 20.6 million users in 2025 alone, bringing its lifetime interaction count to 3.2 billion, according to a PR Newswire release from BofA. That is not a pilot. That is a production system at genuine scale, in a domain with high compliance requirements and measurable outcomes, where the economics have clearly crossed a threshold.
The same pattern shows up across a narrow band of vendors and deployers. Decagon, a customer support AI startup, closed a tender offer in March 2026 at a $4.5 billion valuation — triple the $1.5 billion it announced in June 2025, according to TechCrunch. Customers including Avis Budget Group, 1-800-Flowers, and Oura Health are reporting deflection rates above 80 percent, meaning the vast majority of support interactions are resolved without a human agent ever entering the loop, according to a16z research citing Decagon customer data. Financial app Chime reported a greater than 60 percent reduction in contact-center operating costs and a doubled Net Promoter Score after deploying Decagon. ServiceNow, running AI agents across its own internal IT and HR operations, reports that 90 percent of IT support requests are now self-served, incident closure runs seven times faster, and the company has documented $355 million or more in realized value from the deployment, per its own CEG AI CoE documentation. ASAPP, which runs a multi-model orchestration stack on AWS Bedrock using five to seven LLMs in coordinated choreography, reports that its GenerativeAgent delivers a 77 percent reduction in cost per chat, a 49 percent increase in self-service engagements, and solves up to 40 percent more problems without escalating to a human agent, according to an AWS case study.
These are real numbers. They are not projections. They are the output of real deployments with named customers and, in the case of Bank of America and ServiceNow, self-reported corporate disclosures.
But the distribution of these wins is the more important story.
The pilot-to-production gap is not a footnote. Deloitte's survey of AI-adopting organizations found only about 11 percent are actively using agentic AI systems in production, even as roughly 30 percent are exploring the technology — a gap that speaks for itself. Gartner in June 2025 estimated that over 40 percent of agentic AI projects will be cancelled by the end of 2027, citing its own research. Those numbers sit in obvious tension with the headline metrics. And the companies with documented wins share a specific profile: high-volume, structured, compliance-heavy support workflows — financial services and IT service management — where the integration surface is deep, the workflows are documentable, and the outcome metrics are unambiguous.
Outside those domains, the pattern frays. The customer support space has been promised as paying off since at least 2020. The previous wave delivered rule-based chatbots that could handle a narrow FAQ tree and deflect anything outside the happy path to a frustrated user. The deflection-rate metrics that vendors cite now — 80 percent, 91 percent first-call resolution — sound categorically different. They are, in several respects. Large language models are capable enough to handle genuine variation in customer intent. Orchestration patterns — model routing, context management, multi-step tool use — have matured into deployable architectures rather than research prototypes. And the cost of inference has dropped significantly since 2022, changing the unit economics for high-volume support operations.
The deflection rate itself deserves scrutiny, though. A high deflection rate means an interaction was resolved without human handoff. It does not automatically mean the customer got what they needed, that the problem was fully solved, or that the interaction cost less when you account for LLM compute, system integration, and ongoing maintenance. There is also the cost trajectory problem: Gartner's January 2026 analysis predicts that generative AI cost per resolution will exceed the cost of offshore human agents by 2030, per Ringly's compilation of Gartner data. The efficiency argument — that AI agents are cheaper than human support at scale — is not a constant. It is a function of current inference costs that are themselves moving.
The ROI numbers are directionally consistent but methodologically suspect as a category. Companies see an average return of $3.50 for every $1 invested in AI customer service, climbing to 124 percent or higher ROI by year three, according to Ringly data cited across multiple sources. Cost per customer interaction dropped 68 percent after AI implementation, from $4.60 to $1.45, per the same data set. These figures come from early movers and vendor surveys, which means they reflect selection bias — companies that had the workflows, integration depth, and operational maturity to succeed are the ones reporting results. They are not the industry average. They are the winners' sample.
ASAPP's architecture is instructive here. Running five to seven LLMs in choreographed routing on AWS Bedrock is not a plug-and-play deployment. It is an engineering-intensive orchestration stack that requires ongoing calibration, evaluation, and operational investment. The sophistication of the architecture explains the results. It also explains why the barrier to entry is not the AI capability — it is the organizational infrastructure to deploy and maintain it.
The "finally paying off" framing holds for a specific tier of deployments and a specific category of companies. What has genuinely changed since 2020: the LLMs are capable enough, the orchestration patterns are real, and the economics work in structured, high-volume, measurable domains. The wins are not evenly distributed. They cluster in financial services and IT service management — sectors with structured workflows, deep integration ecosystems, and outcome metrics that are hard to argue with. Gartner predicts that by 2029, agentic AI will autonomously resolve 80 percent of common customer service issues without human intervention — plausible, but an extrapolation from a category of deployments that currently represents a thin slice of the global contact center landscape. Gartner estimates there are 17 million contact center agents worldwide. The agentic AI deployments with verified outcome data serve a fraction of that population.
The "finally" in the headline is earned, but narrowly. If the pattern that separates the companies in production from the majority still in exploration is what you are trying to understand, the answer is not in the AI. It is in whether the organization had the workflows, integration depth, and operational discipline to build the infrastructure layer that makes the AI useful. The agentic layer crossed a threshold. The average enterprise has not.
