The press release version of AI in customer service is about deflection rates and cost reduction. The honest version is more complicated. Some things got faster. Some things got worse. And the teams that are doing it well made choices that the vendor demos do not show you. Here is what is actually happening.
A customer service leader I know spent most of last year rolling out an AI-powered support system. The pilot numbers were strong. The board presentation went well.
The vendor was credible.
Six months into production, her CSAT scores had dropped four points and her team's escalation volume had increased by 22%.
The AI was handling more tickets. The customers were less satisfied. And the connection between those two things was not immediately obvious to the organization until they started pulling conversation transcripts. (For more on the gap between vendor claims and reality, read my thoughts on AI workflow automation for business: what is real, what is hype).
The AI was resolving issues. It was not resolving the customer's experience of having an issue. And in customer service, those are often not the same thing.
This is the part of the AI for customer service story that does not make it into vendor demos. Not because the vendors are dishonest, but because controlled demonstrations do not surface the edge cases, the escalation dynamics, and the customer emotional experience that production environments do.
Here is the honest version of what is happening with AI in customer service in 2026 - what is working, what is not, and what the teams doing it well have figured out.
What AI for Customer Service Actually Does Well
Let me start with what is genuinely working, because the picture is not uniformly negative.
Tier 1 resolution at scale. Simple, repeatable, clearly defined issues - password resets, order status checks, basic troubleshooting for known issues - are being handled by AI at significantly lower cost and often faster than human agents. For high-volume support operations, this is real and meaningful.
Response time at all hours. AI does not have a shift schedule. For customers who need a response at 2am and whose issue is genuinely simple, AI is better than the alternative of waiting until business hours.
First-contact information gathering. AI is effective at collecting the structured information needed before a human agent engages - account details, issue description, previous troubleshooting steps. When this is done well, the human agent arrives at the conversation with context rather than starting from zero.
Knowledge base retrieval and summarization. AI is strong at quickly surfacing relevant internal knowledge base content. Agents who use AI-assisted tools during live conversations consistently resolve issues faster than agents working without them.
Per Salesforce's State of Service report from 2025, 83% of service organizations using AI report productivity improvements. But the same report notes that customer satisfaction improvements are reported by only 52% - a significant gap that reflects the experience-versus-resolution distinction.
Where AI in Customer Service Consistently Falls Short
The failures are as consistent as the successes. And they tend to cluster in the same places.
Complex or emotionally charged situations. An AI can resolve a billing discrepancy. It cannot hold a conversation with a customer who has been frustrated for three interactions and is considering churning.
The cues that signal emotional state - tone, escalation pattern, history - are real but not structured enough for current AI to handle without losing the customer.
Ambiguous issues that require judgment. Many real support issues do not have a clear resolution path. The customer's problem is somewhere between two categories, or it requires a non-standard solution, or it involves a policy question that the written policy does not cleanly address.
AI resolves toward the nearest defined path. That is not always where the customer's issue lives.
Situations where the wrong resolution causes a downstream problem. In customer success contexts, particularly in B2B, the support interaction is one moment in a longer account relationship. An AI that resolves a ticket correctly but in a way that leaves the customer feeling dismissed can damage a relationship that a human agent would have preserved. (This dynamic is exactly why we need to rethink The Retention Math Most CS Teams Get Wrong).
The ticket is closed. The account relationship has moved.
At Zendesk, we had visibility into how AI-assisted resolution was affecting renewal conversations downstream. There were cases where technically correct AI responses had left a relationship impression that human agents were dealing with months later. The support metric was fine.
The relationship metric was not.
What the Press Release Version Gets Wrong
The dominant narrative about AI in customer service focuses on efficiency metrics: deflection rate, average handle time, cost per ticket. These are real metrics. They are also incomplete.
Customer service is not a cost center that happens to interact with customers. It is a relationship function that happens to handle issues. When you optimize only for the efficiency metrics, you optimize for issue resolution and risk optimizing against relationship quality.
The press release version shows AI deflecting 40% of tickets and reducing average handle time by 30%. It does not show you the 15% increase in ticket re-open rate, or the decline in one-touch resolution quality, or the customers who reached a human agent after failing with the AI and arrived in that conversation already frustrated.
I am not suggesting AI does not work in customer service. It does. But the organizations seeing the best results are measuring both the efficiency layer and the relationship layer simultaneously - and using the relationship signals to constrain where AI is applied.
How the Best Teams Are Using AI in Customer Service
The teams that are getting good results have made a different set of choices than the ones that are struggling.
They applied AI where the cost of errors is low, not where the volume is highest. Volume and cost-of-error are not the same thing. A high-volume query that is also emotionally sensitive - a billing dispute, an account termination question - is a poor candidate for AI resolution even though the volume would make it look attractive on the deflection rate metric.
They designed AI as a speed layer for agents, not as an agent replacement. The most effective implementations I have seen use AI to help human agents work faster and more consistently - surfacing relevant knowledge, suggesting responses, flagging escalation signals - rather than replacing the human in the conversation.
This approach does not produce the same headline deflection numbers. It produces better CSAT and lower re-open rates.
They built explicit escalation logic before they launched. Not as an afterthought, not as "we will tune the escalation threshold over time." They mapped the trigger conditions for escalation before the first customer interaction went live, and they tracked escalation patterns from day one as a quality signal.
The escalation rate is not a failure metric. It is a calibration metric. A 15% escalation rate that sends the right tickets to humans is better than a 5% rate that lets the wrong ones through.
They gave agents transparency into what the AI had already done. The customer who reaches a human agent after a failed AI interaction is already frustrated. The agent who can see the AI interaction transcript and understand what was tried can acknowledge that immediately and take a different path.
The agent who has no context treats the customer as if they are starting fresh. The customer experience of those two paths is completely different.
In my work with clients on AI customer service implementations, the consistent finding across successful deployments is that the design of the human-AI handoff is at least as important as the AI itself. How the transition happens, what information transfers, and how the agent is briefed determines the customer's experience more than the AI's accuracy rate.
The Metrics That Actually Matter
Customer satisfaction score, net promoter score, and customer effort score are better outcome metrics for AI customer service quality than deflection rate or handle time.
These are relationship signals. They tell you how the customer feels about the interaction, not just whether the issue was technically resolved.
Track them segmented: AI-handled vs human-handled vs escalated. The comparison across segments tells you whether your AI is performing to the standard you need, and where the gaps are.
Also track: re-open rate (the customer had to come back because the first resolution did not hold), time-to-resolution including any re-opens (not just the first ticket close), and customer retention rate for segments that had AI-handled versus human-handled interactions.
These metrics require more analytical work. They also tell you something real about what is happening with your relationships, not just your queue volume.
What Business Leaders Should Decide Before Deploying AI in Customer Service
The decisions that determine success are made before the AI is turned on.
Decide which issues AI is allowed to fully resolve, which issues AI assists on, and which issues require human handling from the start. This is a strategic decision, not a tuning parameter. It should be made by people who understand your customer relationships, not just your support operations efficiency.
Decide what escalation looks like. Who does the customer reach? How fast?
With what context? The escalation experience is part of the customer experience of your AI implementation.
Decide what your success metrics are before you deploy. If you set them after, they will be defined by what the system happened to produce, not by what actually matters to your customers.
Decide how you will measure relationship quality, not just issue resolution quality. If you can only measure the ticket, you will optimize only the ticket.
