AI Got as Common as Water. That’s Exactly Why So Many Contact Center Projects Drown.

Three years ago, AI was rare and a little mysterious. Today it’s everywhere, as ordinary as water from a tap. Every vendor has it. Every pitch deck promises it. Every board expects it. And that ubiquity is precisely why so many enterprise AI projects are failing.

DMG Consulting just put a name to it. In June, Donna Fluss and her team published an executive brief, “Beyond the Myths: The Realities of AI in CX,” cataloguing 20 of the most persistent myths driving CX decisions. Their finding: leaders are overestimating what AI can deliver today and underestimating what it takes to make it happen. The consequence, in DMG’s own words, is flawed strategies, misallocated budgets, and initiatives that fail to deliver.

Here’s the uncomfortable truth underneath all 20 myths. Knowing about AI is not the same as knowing how to make AI work inside a contact center. Water is everywhere. Building a hydroelectric dam that runs 24/7, holds up a power grid, and doesn’t kill the fish is still a feat of engineering. Most “AI failures” aren’t AI problems at all. They’re integration, compliance, and operations problems wearing an AI costume.

The failure data is brutal, and none of it comes from us

Three independent numbers tell the same story.

MIT’s NANDA initiative studied 300 deployments, 150 executive interviews, and 350 employee surveys. 95% of enterprise GenAI pilots delivered no measurable P&L impact. Only 5% created real value. They called the cause the “learning gap”: tools that dazzle in a demo but can’t retain feedback, adapt to context, or integrate with the messy reality of enterprise workflows.
Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027. The reasons it names: escalating costs, unclear business value, inadequate risk controls. Notice what’s missing from that list. Model capability. A smarter model fixes none of the three.
From the same Gartner analysis: of the thousands of vendors claiming “agentic” capabilities, only around 130 are building anything that deserves the label. The rest is “agent washing” — a chatbot with a new price tag.

Read those failure modes again. Every one is operational, not intellectual. The models are already good enough. The organizations deploying them underestimated what production actually demands.

The contact center is the hardest room in the building

DMG makes the same point from the inside. Two of its five myth categories are dedicated to it: one on why AI cannot compensate for outdated processes, poor workflows, and unreliable data, and one on the naive belief that AI can build, integrate, maintain, and replace complex enterprise systems on its own. One of the specific myths they call out by name: that AI will cut operating costs by 40 to 50% inside a single year.

That myth dies in the contact center, because the contact center is one of the most hostile production environments in the enterprise. Anything you deploy there has to survive, all at once:

Humans in the loop — live agents, supervisors, QA, and escalation paths that have to hand off cleanly.
Legacy backends — systems of record built decades ago that don’t have clean APIs and aren’t getting them.
A stack of SaaS services and databases that all have to be orchestrated in real time, on every call.
99.999% availability. Five nines. Downtime measured in minutes per year, not hours.
Strict compliance and security — PCI, GDPR, HIPAA depending on the vertical, plus audit trails a regulator can actually read.
Scale — tens of thousands of concurrent conversations with real accents, noise, and people who don’t follow the script. Not one clean demo call.
Cost discipline — a per-inference economic model that no AI lab was ever designed around.

Get any single one of those wrong and the project stalls. Naive deployments get all of them wrong at once, because they start from the premise that a powerful enough model will magically dissolve the complexity. It doesn’t dissolve it. It runs straight into it.

What the survivors actually do differently

The 5% that succeed and the 60% that survive Gartner’s cull share one trait: they treat AI as a component in a well-engineered system, not as the system itself. MIT found the same thing from another angle. Buying from specialized vendors and building partnerships succeeded about 67% of the time. Internal builds succeeded a third as often. Deep, domain-specific engineering beats raw model access, every time.

This is the architectural decision we made years ago, and it’s why we can point to more than 130 enterprise customers running live, referenceable projects — not pilots that never left the lab.

Sit that next to Gartner’s other number for a moment. Gartner could find only about 130 vendors on the planet doing agentic AI for real. We can put 130+ enterprise customers on the phone to confirm we’re one of them. Two different 130s. Make of the symmetry what you like.

We use LLMs where they’re genuinely brilliant: offline. They analyze thousands of real customer-agent conversations, the messy, interrupted, ambiguous ones, and deduce the optimal service strategy from what actually worked and what didn’t. Then purpose-built, SLM-based mini apps execute those strategies in production. The heavy, probabilistic model does the thinking offline. The lightweight, deterministic apps do the doing in the live path. No LLM ever touches a live customer call.

Why that split solves exactly what kills projects

Map it back to Gartner’s three failure modes.

Cost stays predictable. No LLM inference on every live turn means no invoice that scales with call volume and blindsides the CFO. Cost consistency, not cost surprises.
Behavior is deterministic and auditable. Compliance can review exactly what the system said and why. That’s the Glass Box. There is no “the model generated it dynamically” answer when legal comes asking.
Risk is controlled by design. Zero hallucinations in the live path, because there is no generative model in the live path to hallucinate in the first place.

The water metaphor holds all the way down. Everyone now has access to the same water. Frontier intelligence is available to anyone with an API key; the raw model is no longer the scarce ingredient, and its cost per call is trending up, not down. Which means the differentiation has moved entirely to engineering: how you channel it, where you place it, what you build around it, and whether the whole thing keeps running at five-nines while staying compliant and affordable.

The lesson is cheap if you learn it from someone else’s pilot

DMG’s brief is worth taking seriously precisely because it comes from an analyst, not a vendor with something to sell. The expectation gap is real, and it will get more expensive before it closes. Some of that spend was always going to be tuition.

But the core lesson costs nothing if you learn it from someone else’s failed pilot: AI competence and contact center competence are two different skills. The projects that win are run by people who have both. Ask your vendor which one they actually have. Then ask them to put 130 live customers on the phone. We can.

About the Author