Customer experience has become the single largest target for enterprise AI investment. New Gartner data reveals CX teams are deploying an average of 4.5 AI use cases and committing 12.7% of their functional budgets to AI initiatives.
“There’s an obvious belief within organizations, and it’s one that’s grounded in reality, that CX is one of the first or best, almost obvious candidates for AI,” says Tanguy Tallon, Senior Director, Agentic Service Line at Foundever.
“Today, AI is part of almost every client discussion,” he continues. “Until recently, only the brands we support in the tech vertical would proactively bring up AI and its CX applications. But now, it’s a topic for all clients and it’s a conversation they instigate.”
Long before breakthroughs in generative and agentic AI pushed the technology into yet another hype cycle, AI’s successful application within CX over recent decades, be it for automation or speech and text analysis, was well proven and well documented. With so much evidence available, in many cases the only remaining barrier to adoption was cost.
And yet even as costs have gone down and accessibility has gone up, the results tell a different story. Across all customer service AI use cases evaluated, just 24% are delivering positive returns. A further 24% are delivering negative returns. And perhaps most concerning, 42% sit in a no-man’s land where leaders simply cannot tell whether the technology is working or not.
This isn’t unique to CX. Across the enterprise, where 92% of large organizations have made significant AI investments over the past 12 months, 78% report that AI projects either fail outright or remain stuck in pilot. But it is within customer experience that the gap between expectation and outcome is most striking.
The gap between insights and execution
Sufficient AI maturity within the market means that any large enterprise can work with a platform provider to build a successful pilot. “The capability is there,” says Tallon. “But scaling a pilot is not a test of the technology. It is a test of the organization. And that’s where things get much more complex.”
Over the past decade, enterprises have invested heavily in understanding their customers, whether through journey mapping, sentiment analysis or VoC programs. As a result, many brands can claim to have genuine customer insights. What they can’t claim is the ability to apply those insights at scale.
The reason is structural. Knowing what customers need and understanding where friction exists is of little practical use unless the organization has the culture, operating model, and infrastructure to coordinate across disconnected systems, inconsistent processes, and teams that work with different tools, different data, and different definitions of success. Without those elements in place, it’s impossible to execute resolutions, especially at scale, and that is why so many AI deployments are failing to live up to expectations.
“You can build a perfect process in a controlled environment,” says Tallon. “It will work beautifully in a pilot. But when you move to production, the variation that exists in the real operation — exceptions, ambiguity, the judgment calls that agents make every day — will surface. You can manage it in a pilot because you’re controlling the conditions. At scale, those conditions disappear and the variation could derail what you’ve built.”
At the heart of this is what Tallon calls tacit knowledge — the experiential understanding that employees carry but that is never formally captured.
“A human agent can absorb ambiguity. AI cannot. An agent knows that a particular FAQ is outdated but one line in it is still valid. They know that there’s a conflict between local instructions and corporate policy, but they know how to handle it. They’ve figured things out through experience. That knowledge sits nowhere. It’s not in any system, any document, any dataset. But it’s exactly the knowledge that determines whether a process works in the real world or only on paper.”
The pressure driving poor decisions
Even without these organizational challenges, the environment surrounding AI adoption in CX would be difficult to navigate.
“Our clients are caught between contradictory forces,” says Tallon. “There’s intense pressure to move fast on AI. But because the technology itself is evolving in six-month cycles, this creates a constant concern that any decision made today could be invalidated by tomorrow.”
This pressure is why over half of business leaders admit to deploying AI primarily because their competitors had, not because they’d identified a clear use case. More worryingly, it’s also why 56% of customer service leaders now expect to have their incentives directly tied to AI outcomes.
“Now factor in that most organizations have legacy infrastructure operating on five-to-ten-year horizons,” points out Tallon. “Think about how long it takes just to change a CRM. These systems, which will ultimately need to support AI, operate on a completely different timeline to the technology they’re being asked to carry.”
The stakes are rising and the challenges are mounting, even as the path to returns remains unclear. This is creating a decision-making environment where organizations that are unsure if they have the necessary requirements in place to make AI work at scale are being pushed to adopt and deploy faster and make bigger investments.
Agentic AI increases capabilities and increases the pressure to act
The arrival of agentic AI is making this situation even worse. Agentic AI represents a genuine shift in what AI can do within customer experience. Where previous generations of AI could answer questions, surface information, or guide a human agent through a process, agentic systems act. They can set a goal, plan the steps to reach it, execute across multiple systems, and adapt their future approach based on the outcome. Applied well, that capability can resolve customer issues autonomously, at scale and with consistency.
However, as buyer interest in agentic AI has grown, so have questionable marketing tactics. Chatbots, scripted workflows and robotic process automation are being promoted as agentic when, in reality, they are anything but.
For organizations already under intense pressure to act, this confusion is compounding the problem. Leaders are buying what they believe is agentic capability, deploying it into operations that aren’t ready for conventional AI, and measuring the results against expectations that neither the technology nor the organization can meet.
And because agentic systems act rather than just answer, the consequences of getting it wrong are fundamentally different.
“If a human agent makes a mistake, it might potentially affect a hundred customers,” says Tallon. “If your AI implementation gets it wrong, on day one it could affect ten thousand. It’s as if every agent in the operation made the same mistake at the same time. That’s the amplification problem.”
Three things before you deploy
This amplification risk is why agentic AI demands something many organizations haven’t built: a foundation of operational truth, clear decision logic, and real-time observability, all in place before the system is given authority to act.
1. Discover what agents do
Not what the process documentation says. Not what the journey map shows. What it takes for agents to do the job — i.e., their tacit knowledge that enables workarounds, adaptations and judgment calls.
A third of organizations still say they don’t understand how to make AI work for their business, and 28% still lack a clearly defined deployment roadmap. The starting point for both problems is the same: organizations don’t understand what they’re trying to automate.
The instinct is to start with the most ambitious use case. Tallon argues the opposite.
“People are drawn to the most complex, most ambitious use cases. They want to build the shiny thing,” he says. “But the evidence consistently points in the opposite direction. Is the task simple? Is it repeatable? Can you measure the output? If the answer to all three is yes, that’s where you start. And then you build slowly, deliberately, proving each step before you take the next one.”
The strongest initial deployments tend to be internally focused on supporting agents rather than replacing customer-facing interactions. Use cases like automated post-call summarization, CRM updates, or disposition coding carry minimal customer-facing risk, and build organizational confidence in AI before the system is given authority to act on the customer’s behalf.
2. Codify the decision logic
Before giving any AI system the authority to act, the rules of action need to be explicit. What triggers which action? Under what conditions does the system escalate? What happens when two rules conflict? What’s the rollback if the system gets it wrong?
This is where the transition from conventional AI to agentic AI becomes most consequential. A chatbot that gives a wrong answer creates a bad interaction. An agentic system that executes the wrong action creates an operational problem that scales instantly.
The codification of decision logic is also where organizations must confront the difference between what they think they want and what genuinely works. Full autonomy is not the goal. The most effective deployments use a graduated approach: AI proposes actions and humans approve them, then AI acts autonomously but with human oversight able to intervene. Only then, with confidence and trust established, should the system handle routine resolution independently, escalating only when complexity or risk thresholds are exceeded. Each stage must be completed successfully before the next can start.
3. Build observability from the start
An AI system that acts autonomously needs to be observable in real time. That means transparency into what the system is doing, why it’s making the decisions it’s making, and how to intervene when something goes wrong.
The financial argument alone demands it. Nearly two-thirds of customer service AI spending is operational expenditure — cloud compute, subscriptions and ongoing model operations. AI is a living system with recurring costs that need to be justified by ongoing, measurable outcomes.
Observability also means audit trails for every decision the system makes, rollback capability for every action it takes and continuous monitoring that flags anomalies before they reach scale. These are architectural requirements. Autonomy without oversight creates the conditions for exactly the kind of scaled errors that make AI deployments fail.
However, who has responsibility for that oversight is of equal importance. “The pilot team has all the knowledge: the context, the intent, the understanding of what the AI is supposed to achieve for the customer,” says Tallon. “The moment that project transfers to IT for production management, all of that knowledge walks out of the room. And that’s exactly the point where projects start to degrade because the organizational handover stripped away the expertise that was making it work.”
The implication is that AI in CX — and agentic systems in particular — can’t be managed in the way traditional technology is managed.
“We need to reinvent what it means to manage AI agents,” says Tallon. “It’s not IT maintenance and it’s not traditional operations management. It’s a new discipline that combines operational expertise — understanding what a good customer interaction looks like — with AI configuration expertise. Those two capabilities need to be working together.”
Measure what matters
Even with the right foundations in place, organizations could still undermine a deployment through incorrectly measuring performance.
Against the goal of increasing productivity, 54% of customer service leaders report that their C-suite views progress as above target. But against the bottom line, only 16% rate customer service as performing significantly above target on cost reduction. Productivity is improving. Value is not following.
“Too many organizations are focusing on deflection rates,” says Tallon. “Platforms are promising 80% deflection. But deflection means nothing on its own. If you deflect a call, but the customer has to call back, or jump to another channel, you haven’t resolved anything. You may have created more frustration. What we should be measuring is resolution. How many customers across all channels are resolving their issue on the first contact? That’s the metric that matters.”
And the downstream consequences of automation are often invisible in traditional metrics.
“There’s a full equation that most organizations aren’t calculating,” Tallon continues. “They measure the AI project in isolation. The deflection rate, automation percentage, cost per interaction. But they don’t look at what happens to the rest of the operation. If AI handles all the simple queries, every interaction reaching a human agent is now an escalation. That means longer handle times, more complex conversations, agents who need more training and more experience. The cost structure of the entire operation shifts, and if you haven’t modeled for that, your ROI calculation is incomplete.”
Scaling is rebuilding
The final misconception that undoes deployments is the belief that a successful pilot can be extended across the organization like a software rollout.
“Scaling AI in CX is much more like scaling a human operation than deploying a piece of software,” says Tallon. “You don’t replicate it. You rebuild, with an understanding of the specific inputs each time. Sometimes you can reuse components. Sometimes you can’t. It’s not a question of drawing a line and extending it across the organization. Each new use case, each new line of business, starts from understanding the specific call drivers, customer expectations, required expertise and desired outcomes.”
That reality demands alignment across functions that don’t traditionally work together.
“When one part of the business says this is a top priority, that signal needs to reach every team and every partner involved,” says Tallon. “But, in practice, this is not always the case. There are silos in these organizations, and if there aren’t strong bridges across them, even well-designed projects will stall.”
“You can work around the transformation challenge at the pilot stage,” he adds. “But when you scale, it becomes an organizational challenge — process transformation, infrastructure transformation, the way you measure, the way teams work together. It’s a transversal transformation of the company. It means doing business differently.”
No matter how much the technology advances or how much the pressure to act intensifies, the route to a successful implementation will remain the same. It starts with operational discipline and an understanding that in CX there are no shortcuts between a pilot that works in a controlled environment and a deployment that delivers at scale.
Where AI fits the bigger picture
AI is one of three challenges showing up at the top of every CX leader’s agenda right now. Download “The CX complexity guide” to see how 221 leaders across eight industries are approaching AI, attrition, and cost — including what separates the organizations already seeing returns from those still stuck in pilot.