May 15, 2026
PolyAI CEO: Deflection Rates Are the Wrong Metric for Contact Centre AI
Contact centre AI is often judged by how many customers it keeps away from a human agent.
Deflection rate has become the metric that leaders use to justify automation investments, while vendors lean on it to demonstrate value, even as a growing number of organisations find that high deflection does not always translate into satisfied customers.
Nikola Mrkšić, CEO and co-founder of PolyAI, argues the metric is doing more harm than good. He said that the industry’s fixation on deflection is obscuring a far more consequential measure: resolution. When deflection climbs but resolution does not follow, customer trust is the first thing to break.
Contained but Not Resolved
Mrkšić draws a distinction between two types of resolution: absolute resolution, which measures whether a customer’s issue was actually solved, and relative resolution, which measures whether it was solved given that the customer stayed within the automated system. PolyAI’s data shows a correlation between relative resolution rates approaching 100% and customer satisfaction scores of five out of five. The principle behind that correlation, Mrkšić said, is that automation should only handle queries where it can deliver a genuine outcome, and hand off to a human whenever it cannot.
The problem is that many contact centres deploy automation in the opposite configuration. They push customers toward digital channels, contain them regardless of whether the issue gets resolved, and count the interaction as a success when the customer drops off. Mrkšić calls this “digital deflection,” a practice where customers get redirected to an online channel, confirmed as redirected, and then disconnected, regardless of whether the issue was actually resolved. The metric registers a successful deflection, but the customer registers abandonment.
The frustration does not stay contained but translated into repeat contacts, channel switching, and churn as downstream consequences that most organisations track in isolation from their automation strategy. Mrkšić offered a personal example: “I churned as a customer. I’ve never looked back, because I was stuck in chat channels where I felt really helpless, and at some point, it was too much.”
Voice Is the Final Escalation Point
One of Mrkšić’s most pointed observations concerned channel switching, which he said only moves in one direction. When digital channels fail, customers migrate to voice. Yet, the reverse almost never happens.
Chat has historically been treated as a lower-stakes environment where companies run lean staffing and accept higher error rates. Voice carries a different weight.
“The kind of stuff you’re allowed to do and get away with in chat, you don’t get to get away with in voice,” Mrkšić said. Senior CX leaders who have spent years building the brand’s phone experience tend to guard the channel far more aggressively, and customers who reach it have already exhausted their patience.
The pandemic should have tilted the balance toward chat, because agents could spin up chat operations from home far more easily than they could handle voice calls with compliance constraints and background noise. Yet voice did not decline. Both channels grew. Mrkšić said:
“If voice was going to disappear, the pandemic was the moment it would have started happening. People were working from home, they couldn’t easily do voice because of compliance and remote setups, and yet voice grew.”
Humans Move Sideways
Asked whether AI will eliminate the need for human agents, Mrkšić offered a forecast that diverges from the most aggressive automation narratives. He expects headcount reductions of roughly 50%, but said at least half of the remaining workforce will transition into fundamentally different roles, operating as product managers for AI systems and refining the processes that automation depends on.
He was also candid about what he has not observed: “I’ve never seen people fire in-house agents.” The contraction lands on outsourced labour. BPO contracts shrink because those workers handle overflow and carry less institutional knowledge. He added: “The classical BPO is probably going away. The BPO turned IT services company is very much here to stay.”
The Contact Centre as Revenue Engine
Mrkšić’s most ambitious argument was that the contact centre should function as a revenue engine. Large enterprises already conduct tens of thousands of customer conversations every day, a data stream he considers more valuable than periodic voice-of-customer surveys that cost millions to commission.
When AI handles routine volume, it also frees capacity for commercial interactions that human agents rarely had time to pursue. Mrkšić described cases where clients generate millions in additional annual revenue by using AI to offer customers relevant products or upgrades during service calls. “Even if 2% of people accept, it’s extra revenue,” he said. “And the other 98% still give you a five-star rating.”
