AI in Service Operations: Where It Earns Its Keep — & Where It Doesn’t
Nearly nine in ten companies now use AI somewhere in their business, according to McKinsey’s 2026 State of AI survey, and service operations is one of the functions where cost reductions are most commonly reported. That same survey found the share of companies seeing any AI impact on earnings has been stuck at 37 percent for two years running. Both things can be true at once, and anyone running a contact center, a field service organization, or a shared services team has probably felt the contradiction firsthand. The tools work, but the business case is harder to find than the demo might have suggested.
The honest version of the AI-in-service story lives in the specifics: which tasks, which customers, which parts of the workflow, and what has to be true underneath for any of it to hold.
Where AI earns its keep in service operations.
The most durable evidence comes from the least glamorous use case. In the largest field study to date, researchers from Stanford and MIT tracked roughly 5,000 support agents at a Fortune 500 software company as a generative AI assistant was rolled out to help them respond to customers. Issues-resolved-per-hour rose about 14 percent overall. For newer and lower-skilled agents, the gain was 35 percent, while top performers saw little change either way. Attrition among agents with the tool ran nearly 9 percent lower. Nobody was replaced; the AI made the people already doing the work faster and, apparently, less likely to quit.
That pattern repeats across the use cases that are currently holding up in production. Case summarization and after-call documentation, which used to consume minutes of every interaction, now take seconds. Knowledge retrieval—the act of finding the right answer inside a sprawling and often outdated knowledge base—is where language models are strongest and where the payoff is immediate. Triage and routing, that is, deciding what a request is and who should handle it, is a classification problem that AI does well and humans find tedious. In field service, the same logic applies to scheduling and dispatch, where the constraints are many and the math is unforgiving, and to predictive maintenance, where the value comes from acting on sensor data before the truck rolls.
Self-service belongs on this list too, with a caveat. AI handles high-volume, well-defined requests, the password resets and order-status checks and policy questions, more consistently than the IVR trees and FAQ pages it replaced. Klarna, the most-cited example, reported its assistant handling two-thirds of customer chats within a month of launch, with resolution times under two minutes. What happened next is the more instructive half of the story.
Where the hype outruns the result.
Klarna’s leadership spent much of 2024 telling the market that AI could do the work of 700 service representatives. By mid-2025 the CEO was recruiting human agents again, telling reporters the company had let cost become “too predominant” and had ended up with “lower quality.” The AI still handles most first-contact volume. What changed was the assumption that it could handle the rest.
That assumption is where most disappointment in service AI originates. Gartner’s 2024 survey of nearly 6,000 customers found that 64 percent would prefer companies not use AI for customer service at all, and 53 percent would consider switching to a competitor over it. Their top concern was the fear of being unable to reach a person. What customers object to is being trapped inside automation with no way out. An AI layer that resolves the easy 70 percent and then makes the hard 30 percent harder to escalate has moved cost from the contact center to the churn line.
The second failure mode is buying something that was never really an agent. Gartner estimates that of the thousands of vendors marketing “agentic AI,” only about 130 offer real agentic capability; the rest are chatbots, assistants, and RPA tools with new labels. The same analysts predict more than 40 percent of agentic AI projects will be cancelled by the end of 2027 over escalating costs, unclear value, or inadequate risk controls. In service operations, the practical symptom is an “agent” that can answer questions but cannot actually issue the refund, reschedule the technician, or update the case, all because nobody built the integrations that would enable it to do so.
The third is the one we see most often. AI gets laid on top of a workflow that was never redesigned, a knowledge base nobody maintains, and customer data that lives in multiple systems that disagree. Deloitte’s 2026 enterprise survey found 37 percent of organizations using AI at a surface level with little or no change to the underlying process, and only 21 percent with a mature governance model for agents despite three-quarters planning to deploy them within two years. McKinsey’s data points in the same direction: among the small group of companies reporting meaningful earnings impact from AI, nearly three-quarters fundamentally redesigned the workflow first. Among everyone else, one in four did.
The questions worth asking.
Before the pilot, before the vendor shortlist, the enterprises getting real value from service AI tend to have answered a handful of questions that the ones getting disappointed skipped:
What work, specifically, are we automating, and how will we measure whether it got done? “Deflection” is not an outcome. Resolution, repeat-contact rate, time to answer, and cost per resolved case are.
When the AI is wrong, how does the customer reach a person, and how quickly? If the answer involves the customer typing “agent” four times, the design is not finished.
Is our knowledge base and customer data good enough for a machine to reason over? A language model will answer confidently from a policy document that was superseded in 2022. The data foundation is the AI project, whether or not it appears in the budget.
Are we redesigning the workflow or bolting a tool onto the old one? The productivity gains in the research came from changing how agents worked, with AI in the loop, rather than from the tool alone.
Who owns the governance? Not the vendor, and not the innovation team. Someone with authority over the service operation has to be accountable for what the system is allowed to do, what it must escalate, and how its decisions are audited.
What this means for the platforms you already run.
Both Oracle and Salesforce are building agents directly into their service applications, and the pace of that investment is real. It also means the questions above are no longer hypothetical for anyone running Fusion Service or Service Cloud. The capability will arrive with the next release whether the organization is ready for it or not.
Apex IT’s view, after years inside these platforms, is that the AI itself is rarely the hard part. The challenge lies in the quality of the data, the integrations that enable the AI to act, and a commitment to change management that can encourage a service team to trust it. That has been true of every platform capability we have implemented since 1997, and it is true of this one. If you are deciding where AI belongs in your service operation, we would welcome the opportunity to talk to you about your business.
Sources:
McKinsey, The State of AI: Global Survey 2026 (August 2026)
Brynjolfsson, Li, and Raymond, “Generative AI at Work,” NBER Working Paper 31161
Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027” (June 2025)
Gartner, “64% of Customers Would Prefer That Companies Didn’t Use AI for Customer Service” (July 2024)
Deloitte, State of AI in the Enterprise 2026
CX Dive, “Klarna changes its AI tune and again recruits humans for customer service” (May 2025)