There is a growing narrative that AI in customer service is being trained to “say no,” prioritizing containment and deflection to avoid passing customers over to human agents rather than solving problems to their satisfaction.
Enterprises have long measured the success of automation in their customer service operations through “containment” or “deflection” rates that focus on how many customer interactions were handled by an AI chatbot or agent rather than escalated to a person.
But, as Steve Blood, VP for Market Intelligence at Five9, told CX Today, that mindset defines AI success through outdated metrics, creating distorted incentives and fragmented customer experiences.
“Ultimately, it’s because it’s the easiest thing to measure. If I can say I can deflect or contain a hundred thousand inquiries a week, and each one costs five euros, that adds up to half a million a week, 26 million a year. It is simple mathematics and people buy it, unfortunately.”
Related articles
- Five9 Bets on a Unified CX Platform as AI Reshapes Contact Centers
- Five9 Positions AI at the Core of CX Transformation Strategy
Why “70% Deflection” Is a Failed Metric
Vendors and enterprises routinely boast that their AI-driven service operations are achieving 70-80 percent containment, which means that most customer inquiries never reach a human agent. The savings are easy to calculate, which makes containment attractive to executives trying to justify their investments in automation.
“The main issue is that there are still too many discrete siloed departments touching the customer,” Blood said.
Each department focuses on protecting its project and uses the containment number to justify its existence. “It’s just a departmental metric, not a view of the customer. That’s how they build their business case, so they don’t even think about what it means for the customer,” Blood added.
Those numbers often reveal little about whether the customer’s problems were solved.
Instead, Five9 argues that automation can significantly improve the customer experience when it is designed around outcomes such as first-contact resolution, customer satisfaction, and customer effort, rather than raw containment rates and operational savings.
“Metrics and core outcomes are actually related,” Blood said. “The outcome is a real-world result, something that’s achievable, and then the metric is going to be the business’s way of measuring whether the outcome was achieved.”
“You have to start with, why are you doing this? You want to save money? No, you’re doing this to improve customer experience. If you get it right, you will save money.”
An example of a good customer outcome would be if 70 percent of interactions were resolved to the customer’s satisfaction without repeat contact or escalation, Blood said. “That’s very different from 70% of calls were deflected.”
From a customer’s perspective, the problem is solved, which the business can determine from the resolution rate metric. To claim real containment, the business would need to collect enough data to understand whether it has truly resolved the customer’s problem, Blood explained.
That could be picked up by a self-service agent asking the customer whether their problem has been solved to their satisfaction. First contact resolution also comes into play here as a more relevant metric than containment, Blood said.
“But again, we need to know, has the customer switched channels? Did they start on the website and find they couldn’t deal with the issue and then moved to the phone channel? If you’re measuring FCR across these things separately, then you’re never going to know if you truly achieve first contact resolution.”
Without full visibility into the customer journey, departments end up measuring success in isolation, whereas in reality customers frequently switch channels after failed interactions.
A chatbot may technically “contain” an inquiry, but the customer goes on to phone support, email the company, or post publicly on social media. So while the website team that created the bot might boast of 80 percent containment, the contact center team gets the follow-up call from the disgruntled customer, and the connection between the two interactions is not captured.
If the customer at this stage is not satisfied with the service response, they are likely to turn to social media platforms to vent their frustration, which can result in reputational damage to the brand.
Customer Satisfaction Score (CSAT) is a useful metric in determining how the customer feels about the interaction. While CX leaders often use CSAT to identify employee soft skills, they could extend it to AI agents to track whether a change to a prompt for a large language model (LLM) affects customer satisfaction positively or negatively. However, businesses should avoid overreliance on CSAT, Blood said.
“The other desirable customer outcome is ‘that was easy.’ That is the customer effort score. And that doesn’t get used enough. We use CSAT and NPS way too much.”
Gartner found that 96 percent of customers who have high effort experiences are more likely to become disloyal, which has a direct effect on revenue from repeat repurchases, Blood noted.
Reducing customer frustration matters more than trying to create exceptional or “delightful” moments. “It’s actually not necessary and neither is it profitable,” Blood said.
“Harvard Business Review found that… the extra benefits of doing that were very minimal. The outcome of their research was that loyalty is driven more by preventing frustration than trying to exceed a customer’s expectations.”
Blood advised focusing more on customer effort: “forget this idea of trying to extend or exceed customers’ expectations. Just do what they want and do it easily. Those are the metrics we should be using to achieve outcomes on behalf of the customer.”
“Obviously there are business outcomes as well and some of those are more internally centric. But as I always say, if you start with the customer outcome and you achieve that, then some of those business outcomes are going to follow through, especially when they’re cost reduction focused.”
Where AI-driven Containment Can Be Effective
AI can deliver major value when deployed appropriately. Some businesses may be able to achieve 80 percent containment because those inquiries can be self-served.
The key is identifying where self-service genuinely aligns with customer intent. Customers typically welcome automation when the issue is simple and the process reduces effort.
“It’s analyzing the inquiries to understand how you can help them, making sure that we can actually resolve it,” Blood said.
Organizations should assess several factors before automating an interaction, starting from whether there is a known and feasible resolution path, the complexity, urgency and emotional sensitivity of the issue, compliance and regulatory requirements, and overall customer effort.
Certain types of inquiries, particularly those that carry emotional weight such as fraud disputes or billing issues, require human empathy and reassurance that AI systems struggle to provide.
When AI systems are properly designed, they can create faster and easier experiences that customers actively prefer. But AI-driven containment goes too far when customer effort increases or resolution quality declines, Blood said. Repeat interactions with AI systems about the same unresolved issue damage trust.
“Customers are very quickly going to get frustrated if it’s obvious to them that they’re being contained [and] not allowed to escalate to a human.”
The issue becomes even more severe when the company itself caused the problem. In those situations, rigid automation can amplify customer anger rather than resolve it, Blood pointed out.
“I’m contacting this company because they’ve screwed up, they’ve made a mistake, and they’re containing me in this AI agent so I can’t actually get what I want done.”
As long as those realities exist, human judgment and intervention will remain necessary.
What Human-in-the-Loop Should Mean
Many organizations still think of human escalation as a last resort when automation fails. Blood argued that human-in-the-loop practices should extend far beyond escalation handling.
Human oversight plays a central role in risk management, especially around financial, legal, regulatory, or reputational decisions. Governance is becoming increasingly important as new AI regulations emerge globally, particularly in Europe, where the EU AI Act is leading with hefty fines for non-compliance.
Human involvement is also key for continuous learning. Service representatives can review failed cases and provide feedback to identify gaps in automation coverage, improve workflows, and help refine prompts and policies.
Blood also sees human agents acting as “experience orchestrators,” stepping in when interactions with customers require the ability to deal with ambiguity and emotional complexity. Emotionally-aware AI can signal the need during a self-service interaction to connect the customer to a human.
“We design it for lowest customer effort, which includes escalation path to a human so that they don’t feel contained. If the customer has confidence that the AI is going to support them, and if it can’t, they’ve got an escalation path to a human, then that is going to build trust.”
If a company starts by defining customer outcomes for self-service, the focus shifts to designing AI to meet customers’ needs rather than internal efficiency targets.
Success will increasingly be defined by whether customers can resolve issues quickly, confidently and without needing to repeat themselves across channels, which requires organizations to design systems around delivering consistent customer outcomes rather than simply automating the most interactions.
