The AI automation tension in CX isn't whether — it's how

Everyone in customer support wants the same thing: AI automation that serves customers faster, better, and cheaper. Vendors are building toward it, buyers are budgeting for it, and market observers keep pointing the industry further down the road. And they're right to. The demand is real, the economics are compelling, and the technology is genuinely improving.

So it's worth being honest about an uncomfortable fact: leaning hard on AI — asking it to carry the heavy automation largely on its own — is what's produced results that are mixed at best, and in places, troubling. The good news, which I'll come back to, is that this points toward a better way to build rather than a reason to retreat.

The evidence is hard to wave away

Customers are running out of patience with automation that doesn't work. In a recent Parloa survey of about a thousand U.S. consumers, 60% said they'll repeat themselves only once before abandoning an automated support attempt, and more than half won't spend even three minutes in an automated system before asking for a person. The single biggest pain point wasn't wait times or transfers — it was talking to a bot that doesn't understand them. Only 14% said they completely trust future AI to handle complex requests better than a human, and more than a third expect that adding automation will make things worse before it makes them better.

The failures carry a price. CX Dive collected three numbers that stick: half of consumers blame company leadership — not "the technology" — when AI support fails them; Google's AI, even at a reported 91% accuracy against trillions of queries, still produces billions of wrong answers a year, with real legal exposure now attached to them; and a single rogue dealership chatbot handed a customer a buyback offer roughly $5,000 above what the business intended to pay. Meanwhile, a striking share of enterprises have quietly rolled back AI agents they'd deployed with confidence.

It would be easy to read all this as a case against automation. I think that's the wrong conclusion — and the wrong analysis.

The pitfall: mistaking the model for the machine

The mistake underneath most of these failures is a conflation. We've started using "AI automation" and "large language model" as if they were the same thing. They aren't.

Zeynep Tufekci made the broader version of this point in a recent New York Times opinion essay: today's large language models are "plausibility engines," not reasoning machines. They don't verify their outputs against truth or logic — they produce what's probable given their training data, and they can't do otherwise on their own. Which means the ugliest failures aren't aberrations to be debugged away; they're the technology doing exactly what it was built to do. Reasoning about the future of automated work as though an LLM is the automation therefore leads you to bad predictions.

The same trap shows up in customer support. When a company decides to "add AI," it too often means pointing a language model at customers and hoping fluency substitutes for reliability. It doesn't. LLMs are extraordinary at language — understanding intent, summarizing a messy thread, drafting a response, searching knowledge by meaning rather than keyword. They are not, on their own, a dependable system of record, a transaction engine, or a source of guaranteed-correct answers. Ask one to be all of those things and you get exactly the failure modes above.

Automation, properly understood, is much bigger than any single model. It includes deterministic rules, workflow logic, system integrations, and human judgment — technologies we've had and trusted for years. The problem isn't that we automated. It's that we asked one probabilistic component to carry the whole load.

Parts of the market are starting to move past the confusion

None of this means the market has sorted itself out. Plenty of vendors are still pushing language models to do more automation than they can reliably carry. But there are signs that some are shifting the weight onto complementary technology — orchestration, integration, and governance — to do the work the model can't. CX Today recently argued that most buyers are still shopping for the last AI cycle, evaluating tools on conversational polish and demo-day charm, while the more serious competition has quietly moved underneath the experience layer — to orchestration, governance, and execution depth. As Genesys CTO Glenn Nethercutt framed it, agentic AI is shifting CX "from assisted engagement to governed execution." The important question is no longer whether the assistant sounds smart. It's whether the system can connect to live systems, act across them, stay inside guardrails, and do something sensible when a workflow fails.

That's not an argument for less automation. It's an argument for more — but of a different and more serious kind. The frontier isn't a more eloquent chatbot. It's automation that can reliably resolve and act.

A measured approach beats a maximalist one

The path forward isn't to retreat to human agents, and it isn't to hand everything to a model and cross your fingers. It's to layer the technologies according to what each is actually good at:

Use LLMs where language is the hard part — interpreting what a customer means, pulling the relevant history, drafting a clear reply, surfacing the right knowledge. Wrap them in rules-based logic and real integrations where correctness and consequence matter — pricing, account changes, anything that touches money or state — so that outcomes are governed rather than guessed. And keep human agents in the loop where judgment, empathy, and accountability carry the day, with automation handling the volume and setting them up to resolve the hard cases well.

It's worth taking the skeptic's objection seriously here. Tufekci is doubtful that wrapping models in deterministic rules and harnesses really solves the problem, on the grounds that hand-specifying every rule and boundary is exactly what caused symbolic AI to stall decades ago. She's right that you can't enumerate the entire universe of customer interactions in rigid logic. But that's not the job. The job is narrower and very achievable: identify the specific, high-consequence actions where correctness is non-negotiable — issuing a refund, changing an account, quoting a price — and bound the automation to deterministic paths there, while letting the model handle the open-ended language work and routing genuine ambiguity to a person. You don't need to formalize everything. You need to formalize the parts where being wrong is expensive, and design the system to know the difference.

Done this way, automation and human assistance stop being rivals and become complements. The data even points there: three-quarters of consumers in that same Parloa study said they'd prefer automation — if it could anticipate their needs and act proactively rather than making them repeat themselves. Info-Tech's Julie Geller put the operating principle well: treat customer friction "as a system signal rather than just a metric," and build automation from real interaction data instead of assumptions. And it can be done: when Chime deployed its AI agent thoughtfully, satisfaction went up, with its COO noting that automation and cost savings don't have to come at the expense of a great experience.

Why the next few years are the exciting part

The tension in the market right now — enthusiastic vendors, impatient customers, uneven results — isn't a sign that CX automation is failing. It's what the middle of a genuine transition looks like. The industry is moving from "does it sound intelligent" to "does it reliably do the job," and that's a far healthier question to be arguing about.

The teams that win the next few years won't be the ones that automate the most, or the least. They'll be the ones that stop treating a language model as the whole answer and start engineering automation as a system — LLMs for language, deterministic logic and integrations for reliability, humans for judgment. Set expectations at that level, and the "troubling" results become solvable engineering problems rather than reasons for despair.

We're past the point where fluent chat impresses anyone. What's coming is automation that actually resolves things — and that's a much more interesting future to be building toward.

‍