Executives Are Killing AI Projects They Can't Measure. CX Is Where Measurement Is Winnable.

Across the enterprise, executives are pulling the plug on AI initiatives they can't tie to a business outcome. The CX function could easily fall into that graveyard, except that it is one of the few places where the return on AI is genuinely measurable, if the measurement is done honestly. That takes three things: an ROI model the customer actually owns and can defend, honesty about which targets are real and which are vanity, and a measurement foundation solid enough to prove the numbers rather than assert them. Get those right and CX becomes defensible exactly where other AI projects get cancelled.

Antony Passemard is VP of Customer Strategy at Cresta, where his team runs ROI modeling and value advising for enterprises deploying AI in their service operations. A former Google Cloud executive with prior experience at AWS and Salesforce Service Cloud, Passemard has sat through enough ROI conversations to know exactly how they go wrong. His take is that measurability, not capability, is what now separates the AI projects that last from the ones that get cancelled.

"The good thing about the customer experience team is that they have a lot of KPIs. They track everything. The ability to create ROI models on every one of those KPIs is very good," he says. With no shortage of numbers, the core problem becomes how the ROI exercise is run and who ends up able to stand behind it.

The vendor-built spreadsheet collapses on contact

Passemard is candid that the traditional way vendors demonstrate ROI is self-defeating. Building the model for the customer produces a number the customer can't stand behind. "A customer says, 'Show me the ROI of deploying your solution.' We used to do that. We'd ask about the length of their calls, the cost per call, and as much as we could gather and built a whole spreadsheet," he shares.

The trouble comes the moment the customer reviews it. An assumption the vendor had to estimate, like the cost per call, gets questioned immediately, and the honest answer is that the vendor filled it in with best estimates because the customer never supplied the real figure. Interestingly, each customer tends to calculate their cost per call differently and sometimes use a figure that does not reflect the reality of the true cost. The conversation starts to focus on a single metric and loses the bigger picture. The result is a model riddled with guesswork that unravels under the first real test, which is fatal precisely at the moment the customer needs it to hold: in front of leadership. "If I am the one creating the spreadsheet, it's going to be extremely difficult for them to understand it deeply, to own it. If their executive team asks questions about the model, they're not going to be confident in their answer. It's just going to delay things and they're not going to look good for them," Passemard asserts.

Hand the customer the levers

The fix he describes is to invert ownership of the model. Rather than delivering a finished ROI number, Passemard's team provides the framework, the levers, and the ranges observed across other deployments, then has the customer populate it with their own figures. "We have a business value assessment framework," he says. "We sit down with you and you're going to build the spreadsheet. I'm going to provide you some ranges for similar customers, I'm going to tell you the levers of value, and you plug in your numbers to create a few scenarios."

The benefits compound. The customer sees that the vendor understands the domain, gains a model they can actually manipulate, and, most importantly, walks away able to explain the projections to the people who control the budget. "They can go defend that to their executive, to their board, when they pitch their AI projects. That's really important. In addition, it removed any vendor bias."

Beware the vanity target

Passemard reserves his sharpest skepticism for the headline numbers some vendors love to promise, using containment rate as the example. A large figure like 80% containment sounds impressive, but it's manufactured in ways that destroy the customer experience. "You're unlikely to get that number without significant impact on your customer experience. For example, I can just hang up on you and if you don't call back, that can be counted as a containment. Other metrics like resolution rate may be better to look at. Some use cases will have more than 80% containment with a great customer experience. Others will not, and that's ok. You may not want it. Targets should be looked at in context of the overall performance of the contact center," Passemard points out.

His point is that a metric detached from the customer outcome it's supposed to represent can be gamed into meaninglessness, which is why an honest ROI model can't rest on surface numbers at all. It has to rest on knowing what actually happened inside each conversation. This requires understanding the conversation at a depth most systems never reach.

Tie behaviors to outcomes on every call

The owned model and the honest targets hold up only if the underlying measurement is real. That's the part most systems get wrong, because they describe conversations at the surface without explaining what drove the result. "Most tools will tell you what the call is about: the intent, maybe a couple sub-intents, the sentiment throughout the call, whether it concluded as a success or failure. That's it," Passemard says.

The gap he points to is the difference between observing that a call went well and knowing which agent behaviors produced that outcome: greeting the customer, asking the right question at the right time, following the right steps. Establishing that link makes ROI provable rather than asserted, and doing it credibly means proving it across every conversation rather than the handful a supervisor has time to review. "Supervisors typically listen to 1 to 3% of calls and score agents on those. When you can listen to 100% of the calls, and apply a deep understanding of behaviors within the calls, the scorecards are without biases, holistic and particularly insightful on every single call. You can link those behaviors to churn, to revenue, to whatever outcome you want to track."

That full-coverage scoring is what makes the ROI model verifiable rather than aspirational. With every call measured, a deployment can be A/B tested against the exact figures the customer built into their own model. "It's easy to track over time if the deployment is actually efficient and yields the results you planned through your ROI exercise," Passemard explains.

In an environment where executives are increasingly unwilling to fund AI they can't measure, the winning move is radical measurability: a model the customer owns and can defend, honesty about which targets are real, and a measurement foundation deep enough to tie specific behaviors to specific business outcomes on every call. CX doesn't have to be a place where AI spending disappears into an unprovable promise. Done right, it's one of the few places where the return can be shown, defended, and tracked, which is exactly why it survives the budget scrutiny that kills less measurable projects elsewhere.