The Real Problem With Your CX AI Strategy Isn't Automation—It's The Mix

Niranjan Virjayaragavan is the Chief Technology Officer at Five9, an open cloud platform for customer experience.

gettyMost organizations will tell you their customer experience (CX) AI strategy is working, and the numbers back them up.​​

A Gartner survey of service and support leaders found AI spending is up 38% while overall service budgets grew just 2%, a level of commitment that would have sounded implausible three years ago. But that measurement covers only half the question. It shows how aggressively organizations are investing in AI and almost nothing about how well AI and humans work the same queue together, which is the part many organizations haven’t engineered for at all. ​

That’s the real problem hiding underneath years of AI adoption headlines. The industry keeps asking how to automate more, when an equally consequential question is how to get the right mix of human and AI in interactions that need both.

Automation and collaboration aren’t the same investment; optimizing hard for one doesn’t get you the other for free. An organization can hit every automation target and still hand customers off to confused, underinformed agents every day, which is another way of saying the industry is measuring half the problem.

The pace of automation explains why the mix got overlooked. Gartner projects that by 2029, 80% of common customer service interactions will be handled by agentic AI, a shift most industries would take decades to absorb, compressed into a few years by real business pressure. Ninety-one percent of customer service leaders report pressure from senior leadership to implement AI, making automation more of an executive mandate than a strategic choice.

Mandates produce scoreboards, and the AI scoreboard in CX has almost always been “how much did we automate?” It’s an easy number to report up: interactions deflected, cost per contact down and headcount reallocated. What it doesn’t capture is what happens when AI can’t fully resolve alone, where the human half of the mix matters most. A virtual agent that hits an impressive containment number by refusing to escalate isn’t succeeding; it’s just deferring the cost.

Scores go up quarterly while quietly making the job harder for whoever’s left handling what AI couldn’t resolve. If the easy interactions get faster and the leftover ones get more complex and emotionally charged, an organization measuring only the aggregate has no way of seeing what’s “AI success” and where CX is breaking down. The result: AI strategies that look excellent on the chosen scorecard and remain untested on the one that wasn’t.

That gap shows up in deployments that automated first and designed the mix second. Three failure points account for most of it.​

First, context doesn’t survive the handoff. AI knows what the customer said, why they called and what it already tried, but the human agent who picks up next often knows none of that. In my company’s 2026 survey, 34% of leaders said context or data gets lost when a conversation switches channels. A transcript records what was said, not what’s already resolved, which is why a caller has to verify their identity with the AI and the human agent—a duplicate created by the architecture. It’s the most visible failure because the customer feels it directly, undoing whatever goodwill the automation built.

Second is escalation logic that was never defined. Many organizations can’t say when AI should hand off or keep working a problem. Our report found 37% of business leaders experience challenges such as agents lacking visibility into the prior AI interaction even after a transfer, so the agent starts partially blind by design, not by accident. Without an explicit rule, escalation happens inconsistently: sometimes too early, when AI could have resolved things itself, and sometimes too late, after a frustrated customer has already tried repeatedly to reach a human.​​

Third is a supervisor who can’t see the whole queue as one. Metrigy reports 66% of companies either run on a single-server hosted platform managed by a third-party or on-premises infrastructure rather than cloud-native platforms, a reasonable proxy for how many still lack a real-time, unified view spanning human and AI-handled interactions. Most supervisors are stitching together two separate systems mentally, letting problems go unnoticed until a customer complains.

Agents feel that gap first, long before it shows up in a dashboard. A 2026 NYU Tandon and Cornell Tech study found the strongest predictors of a stress spike were long calls with unresolved problems and calls requiring repeated explanations, a fair description of the hardest, least-prepared-for interactions from the inside.

This isn’t a model problem; it’s a design problem. The shift that matters is to stop treating escalation as a single handoff where AI exits and a human starts cold. Design instead for AI that stays present for the full interaction, with humans stepping in where needed. I’ve found that five principles make that work:

1. Context That Travels: What the customer asked, what’s verified and what AI already tried should follow the interaction so no one needs to repeat themselves.

2. Routing By Interaction Type: The human-to-AI ratio shouldn’t be uniform. A password reset and a billing dispute with an upset customer call for different mixes, so set the ratio deliberately for each.

3. Confidence-Based Escalation: Let real-time confidence determine how many attempts AI makes before a person joins, with explicit, auditable thresholds; then, tighten them for high-stakes or regulated issues.

4. AI That Never Fully Steps Aside: While the human handles the judgment call, AI should keep working in the background, pulling records and drafting notes, then picking the thread back up when the human’s part is done.

5. One View Of The Queue: Supervisors should see quality and outcomes across human and AI-handled interactions together.

To find where your mix is breaking, use the data you already have. Pull 50 escalated interactions and count how often customers repeated themselves. Compare containment rates with repeat contacts, since high containment alongside high repeat contact likely means AI is deferring cost, not resolving issues.

The contact centers that are successful won’t be the ones with the highest automation numbers. They’ll be the ones that build seamless coexistence of agentic and human-supported experiences—the difference customers notice and agents feel every time they pick up the phone.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?