Customer service leaders are spending more on AI than ever, but the real question is whether customers are seeing the benefits. According to Gartner, AI spending in customer service increased 38% this year, while overall service and support budgets grew just 2%. Meanwhile, 91% of service leaders report executive pressure to implement AI in 2026. The investment is accelerating, but customer enthusiasm isn't keeping pace.
In Gartner's survey of 3,566 customers, people were roughly three times more likely to have used ChatGPT, Gemini, or Copilot for their last service issue than the company's own chatbot. And after one bad chatbot experience, only 27% would try again.
That's a difficult contrast to ignore. Companies are investing heavily in AI-powered service, yet customers are often choosing to solve problems elsewhere. It doesn't mean the benefits of AI in customer service are overstated. The research shows meaningful gains in productivity, resolution, and efficiency. But those gains aren't distributed equally. The difference often comes down to what organizations are trying to improve: the number of conversations AI handles, or the number of customer problems the business actually solves.
When executives think about AI in customer service, the first image is often a virtual agent handling conversations without human involvement. But some of the strongest evidence for AI's value comes from what happens when it assists human agents instead.
In Generative AI at Work, researchers from Stanford and MIT studied 5,179 customer support agents during a staggered rollout of an AI assistant. The results were striking. Issues resolved per hour increased 14% on average. Among novice and lower-skilled agents, productivity jumped 34%. For the most experienced agents, the improvement was minimal.
That last finding is particularly revealing. The AI wasn't making top performers dramatically better. It was helping less-experienced agents work more like them. The researchers found evidence that the technology was spreading the practices of stronger agents across the workforce. Customer sentiment improved, and agents were more likely to stay.
For contact center leaders, this changes the business case. The persistent cost challenge isn't simply how much an experienced agent can accomplish in an hour. It's the constant cycle of attrition, recruitment, training, and ramp-up that leaves a significant share of the workforce below full proficiency.
If agent assist can help a three-month employee perform closer to a three-year veteran, the value extends well beyond faster conversations. This is where conversational AI and agent assist can work together: supporting agents in real time while automating routine interactions that don't require human judgment.
McKinsey documented a North American telecom provider that reduced total call volume by approximately 30%. Average handle time fell by more than a quarter, while first-call resolution improved by 10 to 20 percentage points. Those are the kinds of results that make a compelling executive business case, but the more interesting part of the story is how the company got there.
Before introducing generative AI, the telecom provider mapped customer journeys, eliminated unnecessary steps, and addressed misalignment between marketing and customer care. In other words, it fixed the process before automating it.
McKinsey's broader analysis suggests organizations may need to spend roughly $3 on change management for every $1 invested in model development. That's an important reminder for enterprise contact centers. AI doesn't automatically make a broken workflow better. It can just make that workflow operate faster, at greater scale.
The telecom's results came from combining process redesign with AI. Technology amplified the improvements rather than creating them from scratch. For leaders evaluating contact center AI, workflow readiness may be a better predictor of return than model capability alone.
So how much customer service can AI realistically automate?
McKinsey estimates that 50% to 60% of customer interactions remain transactional. These are often the most promising candidates for automation: routine requests, predictable workflows, and interactions with a clearly defined outcome. But that percentage represents the potential scope of automation, not a guaranteed result.
And the pressure on contact centers isn't disappearing. According to McKinsey, 57% of customer care leaders expect call volumes to increase over the next one to two years. Few organizations have achieved a sustained annual decline in interactions of 10% or more since 2022.
The opportunity, then, isn't necessarily to make half of all contacts disappear. It's to handle a meaningful portion of demand more effectively.
Gartner's customer research highlights another challenge. During their most recent service interaction, 49% of customers said they would have been willing to use a chatbot. Only 7% actually did. That's a 42-percentage-point gap between willingness and usage.
Customers aren't uniformly opposed to automation. But willingness alone doesn't translate into adoption. Gartner recommends starting with a limited set of issue types where virtual agents can achieve consistently high resolution rates, communicating those capabilities clearly, and expanding only when performance supports it.
A virtual agent that reliably resolves ten common issues can create more value than one that attempts a hundred and leaves customers uncertain about what happens next. The goal is to earn trust through performance, not maximize automation coverage on a dashboard.
Gartner identifies three recurring frustrations: the chatbot misunderstands the issue, provides generic information, or makes it harder to reach a human.
Each one creates a familiar experience. The customer explains the problem, receives information that doesn't solve it, and eventually starts over with an agent. The interaction may count as an AI engagement. From the customer's perspective, it was another step between them and a solution.
Expectations are also changing. Among customers who use generative AI, 58% have already used it to complete a task on their behalf. In B2B settings, that figure rises to 74%. People increasingly expect AI to do things: book appointments, place orders, submit documents, update accounts, and manage subscriptions.
That raises the bar for enterprise conversational AI. A chatbot that can explain how to change a billing address is useful. One that can securely update the address in the CRM and confirm the change has actually resolved the request.
The difference isn't simply a more sophisticated language model. It's integration with the systems where customer service work gets done. One Primero's Cognigy and NiCE CX integration for Ferrellgas, for example, enabled automated payment handling and more seamless agent handoffs, helping reduce call abandonment from 90% to 30%.
The strongest experience isn't necessarily AI-only. It's one where AI resolves what it can, recognizes what it can't, and transfers the conversation with context intact.
For years, a central promise of customer service AI was straightforward: automate more interactions and reduce the number of agents required. The evidence is pointing toward a more nuanced outcome. In a Gartner survey of 321 service leaders, only 20% reported reducing agent headcount because of AI. Nearly 80% plan to move agents into new roles, while 84% are adding skills to frontline positions.
Gartner also predicts that by 2027, half of organizations planning significant AI-driven workforce reductions will abandon those plans. By 2028, it expects more than half of service organizations to double their technology spending without corresponding staff reductions.
That's not necessarily a failure of AI economics. It suggests the value may be showing up somewhere different from where the original business case expected it. Agents can handle more complex work. New hires can reach proficiency faster. Organizations can absorb additional demand without increasing headcount at the same rate.
One finding is particularly telling: 58% of leaders plan to upskill agents into knowledge management specialists. It makes sense. Every virtual agent, AI assistant, and automated workflow depends on accurate, accessible information. If the underlying knowledge is outdated or inconsistent, even a capable AI system will produce disappointing results.
Knowledge management is becoming part of the operating model for AI, not simply a documentation task. That also suggests a different way to measure return: cost per resolved contact, capacity gained, ramp time reduced, and customer outcomes improved. Those measures may tell a more complete story than the number of agent seats eliminated.
Organizations seeing meaningful results don't necessarily automate the most. They make more deliberate choices about where AI belongs.
Six practices stand out:
Successful enterprise AI increasingly looks less like a standalone technology initiative and more like disciplined service operations with better tools.
There's no universal answer, and that's worth acknowledging before committing to another platform or pilot. The starting point depends on your interaction mix and where cost concentrates today. If attrition and ramp time drive cost, conversational AI and agent assist inside the agent desktop is often the faster return. If a small number of transactional intents drive most volume, a narrowly scoped virtual agent with deep back-end integration is the better first move.
For an organization where a handful of transactional requests account for much of the volume, a narrowly scoped virtual agent with strong back-end integration may make more sense. Both approaches can deliver measurable value. The right starting point depends on where the friction and cost sit today.
Before choosing a use case, establish a baseline for handle time, resolution rates, repeat contacts, and agent proficiency. Identify the systems each workflow needs to access. Decide who will own performance, knowledge quality, and improvements after go-live.
The benefits of AI in customer service are no longer purely theoretical. But neither are they automatic. The organizations that capture them will be the ones that treat AI as a way to improve service outcomes, not simply increase the percentage of interactions handled by a machine.

