Your AI is handling the volume. Customers are quietly losing faith. TELUS Digital found two-thirds of enterprises can't measure whether their CX AI is actually working. Forrester warned this would damage customer trust in 2026. The data says it already has - silently, without anyone looking.
One account I worked with had been building their automated customer service response system for nearly two years. The logic was sensible: handle the predictable questions automatically, free the human agents for harder conversations. Most companies were doing something similar.
But every time they reviewed the system, they found gaps. A customer asking about a delivery delay tied to a specific carrier change - not covered. A billing question that only appeared after someone had upgraded and then downgraded in the same billing cycle - not covered. So they added it. Then the next one. Then the one after that.
I asked their team lead how many scenarios they were managing. He paused longer than I expected. "Somewhere north of four hundred," he said. He wasn't entirely certain of the number.
The bot was answering. What nobody was tracking - systematically, at scale - was whether any of those answers were actually resolving anything. Adding scenario four hundred and one was operationally easier than auditing whether scenarios one through fifty still worked.
That account wasn't unusual. That's the problem.
The Scale of the Blind Spot
New research from TELUS Digital, published this month, surveyed 815 enterprise decision-makers across 12 countries on AI in customer experience. The finding that stopped me: 61% of organizations are spending over $10 million annually on CX AI delivery. Two-thirds do not have the automated quality assurance infrastructure to know whether that investment is working. Only 32% have deployed the monitoring systems needed to evaluate AI interactions at scale.
As Sean Nolan reported in "TELUS Digital Exposes the Enterprise AI Performance Gap" on CX Today - one analyst put it plainly: "Adoption of AI-powered solutions in CX has moved fast but enterprises haven't caught up to optimizing it quite yet."
Tens of millions deployed. No real view of what's happening at the interaction level.
That's the same situation that team lead was in, just at a different scale. They'd built something they could no longer see inside - and had no system to tell them when it broke.
Forrester Said This Would Happen
In October 2025, Forrester Research released their 2026 B2C CX Predictions. One finding: one-third of companies would damage customer trust by deploying premature genAI self-service in 2026, driven by cost pressure rather than customer readiness. As reported in "AI Self-Service Backlash Coming in 2026, Warns Forrester" by Patrecia Meliana on ContentGrip - their Chief Research Officer was direct: "In 2026, trust and value will be the guiding beacons. Superficial efforts won't cut it anymore."
That prediction is no longer about the future. We're in it.
The Forrester warning and the TELUS measurement gap describe the same problem from two different vantage points. Companies deployed AI because the pressure to cut costs was real and the business case was clear. Most didn't build the infrastructure to know what that deployment was doing to their customers month by month. Cost savings appeared on the dashboard. Trust erosion didn't have a field.
Why Nobody Is Looking
This isn't a story about careless organizations. Most companies in that TELUS survey have competent, well-intentioned people running their CX operations. The problem isn't attitudinal. It's structural.
The person who approved the AI deployment is measured on cost reduction. Costs came down - at least the visible ones. The person who hears real customer frustration is in the contact center, logging tickets into a system that doesn't connect to the AI team. The person managing the renewal is working from a relationship picture that may be six months out of date.
Each person is doing their job. The accountability architecture just doesn't create a line of sight between what the AI is doing in an interaction and what that does to the customer relationship downstream.
Nobody built that line - because it wasn't in the scope document, wasn't in the budget, and wasn't in anyone's quarterly targets.
So the bot keeps answering. The cost metrics stay green. And somewhere downstream, a customer who received the wrong information twice, gave up on self-service, and never escalated - because they didn't think it was worth the effort - is now in a renewal conversation that's harder than the account team expected. No one flagged it. No one built the system to flag it.
What the Breakdown Actually Looks Like
Most commentary on CX AI failure reaches for dramatic churn statistics. I'll skip that, because this failure mode is subtler - and the subtlety is exactly what makes it dangerous.
When a customer has a bad AI interaction, most of them don't submit a complaint. They quietly lower their expectations. They stay on the platform for now, take the renewal call, go through the motions. But the relationship that made the renewal easy - the one built on confidence that their needs would actually be met - has already degraded.
That's the part nobody shows in the case study.
The account from the opening had strong headline metrics for almost eighteen months. CSAT was flat, not declining. It looked like stability. What was actually happening: the customers most frustrated with the bot had stopped using it entirely, which meant they'd stopped rating it. Flat scores were masking a hollowing-out that no one was measuring, because no one had built the measurement layer.
By the time the renewal conversations started getting harder, the cause had been invisible for months. The AI had been "answering" the whole time. Nobody knew what it had cost.
The Question Nobody Asked Before Launch
I'm not arguing against AI in customer service - that argument is effectively settled. The question now is execution quality, not direction.
But deploying and operating are two genuinely different decisions, and most organizations are treating them as one.
The TELUS data shows a promising shift in stated priorities: 47% of enterprises cited CSAT and NPS improvement as a top priority, while average handle time reduction ranked at just 19%. The instinct is right. The gap is between what companies say they're optimizing for and whether they've built the infrastructure to actually measure it.
Before the next AI launch - or the next scenario added to an existing system - the question worth sitting with isn't "will this handle more volume?" It's: how will we know, at the interaction level, when this fails in a way we didn't anticipate? Who owns that answer? And what happens in the gap between the failure and the moment anyone finds out?
I don't have a clean universal answer to that third one. Failure modes depend too much on context, on the specific distance between what the AI was trained to handle and what customers actually bring to it. What I've consistently seen is that organizations which ask those questions before launch are the ones that still have visibility six months later. The ones that don't are adding scenario four hundred and fifty while the problems compound silently underneath.
What Comes After the Trust Event
The TELUS survey found that organizations with the most observability into AI performance are shifting fastest toward quality outcomes - away from pure efficiency metrics, toward something that actually measures whether the customer got what they needed. That's not a coincidence. Measurement is what connects intent to result.
The companies that hold customer relationships in the AI era won't be the ones with the most automation. They'll be the ones that can see inside what they've built - which turns out to be harder than building the AI itself.
Whether most enterprises figure this out before their first significant trust event or after it - I'm genuinely not sure. For two-thirds of them, based on what the TELUS data shows, the clock is already running.
