Get a Demo
AI & Automation

You Bought All the AI. Nobody Fixed the Conversation.

Max Noxon Max Noxon
·
You Bought All the AI. Nobody Fixed the Conversation.

Originally published on LinkedIn .

You already know the seats don't stay warm and the agents are stressed, so I'll skip that part. You have the dashboards (maybe too many). You've had the “Outsource to Manila and Bangalore” conversation more times than anyone wants to count, and there's a slide deck somewhere proving you've been investing in the right things.

Here's what the outside data says about how that's going. Forrester's 2026 CX Index found that 26% of US and Canadian brands improved this year and 7% declined. That's a real turn after several years of decline. It's also a small one: 68% of those brands didn't move at all, and Europe and Asia Pacific are mostly holding still. Forrester's own summary is "uneven momentum."

The newer part of Forrester's work is on the employee side. Its 2026 Total Experience Score now includes an Employee Experience Index, and only 25% of US brands came out with a positive employee-experience impact. 37% came out negative.

Bad experiences still cost real money. Qualtrics' 2026 Consumer Experience Trends Report found that for every 10 poor experiences, 5 result in reduced or eliminated spending. AI customer service is where it's going worst: nearly 1 in 5 consumers who used it saw no benefit , a failure rate nearly 4x higher than for AI use on average. Forrester predicted that a third of companies would erode customer trust in 2026 by rolling out AI self-service where it was never going to work.

The AI roadmap is on schedule. For three in four North American brands, customer experience didn't measurably improve. Let's go find out why.

Every Call Left in the Queue Is a Boss Fight

Here's what your vendors don't linger on: voice agents, chatbots, and agentic automation are extraordinarily good at the easy stuff. Password resets. Order status. Balance inquiries. The cheerful category of interactions that a sufficiently determined FAQ page could have handled anyway. The bots ate all of that. Fast, effective, cheap. Then they looked at everything else in the queue (the billing dispute that turned out to be fraud, the customer calling back for the fourth time, the person who is upset in a way that needs human judgment) and punted. Decisively.

TASK_ESCALATE = {
    "status": "UNRESOLVED",
    "next_owner": "HUMANS",
    "message": "Have fun with that, humans."
}

So your human agents now do a different job than they did three years ago. What's left is empathy, lateral problem-solving, and holding contradictory context in your head while staying warm with someone who was frustrated before they said hello. The human brain is brilliant at this. It is not brilliant when it's exhausted, juggling seventeen systems, and absorbing accent and cultural friction on every other call. ContactBabel sees the shift coming. Its 2026 US Contact Center Decision-Makers' Guide expects the ability to understand complex issues to matter more as self-service takes the straightforward requests, and its Emotion-Driven Retention paper says the typical call is now more complex than it used to be, which increases agents' cognitive load.

The clock hasn't relaxed to match. In that paper's data, mean service call duration went from 306 seconds in 2012 to 423 in 2024, average speed to answer went from 31 seconds to 99, and the 2024 call abandonment rate (8.9%) was the highest ContactBabel has recorded. Agents in large US contact centers also use an average of 4.6 applications per call.

There's a second layer most AI roadmaps skip. Agent assist tools (real-time suggestions, automated scoring, live sentiment tracking) don't remove cognitive load. They repurpose it. Instead of creating solutions, agents audit machine outputs and verify the AI's suggestion before acting on it. Checking a suggestion is work, and the checking never stops. Call it the vigilance tax.

It isn't just a contact center problem, either. A Boston Consulting Group study of nearly 1,500 workers, published in Harvard Business Review , found that employees with high AI oversight responsibilities reported 14% more mental effort, 12% more mental fatigue, and 19% greater information overload than those with lower oversight demands. The researchers called it "AI brain fry." Your agents are living it on every shift, and it won't show up on a dashboard. In that same guide, 62% of centers rate their agents' morale good or excellent . That's decision-makers talking. The agents' version shows up in your exit interviews.

You've quietly promoted every agent on the floor into a cognitive athlete. The job got harder at the exact moment AI cleared the easy bits off the deck.

Do your tools, your training, and your operating model reflect that?

Congratulations, You Built a Very Expensive Browser Tab Collection

Here's what I keep noticing when I talk to CX leaders sitting on big AI investments with underwhelming ROI: the stack has a layer for everything. Pre-call intelligence so the agent knows the customer's history before hello. Real-time assist surfacing the right answer mid-conversation. Post-call automation for the wrap-up. Task completion scripts and AI-thingers so nothing falls through the cracks. You've instrumented the entire interaction and you're still eking out a few points on KPIs that were supposed to move a lot more. (If your deck shows a double-digit swing, send it over. I'll wait.)

Most of those tools add what cognitive scientists call extraneous load: mental effort spent navigating systems, switching screens, and verifying AI outputs. What agents need is germane load, the kind that produces empathy, judgment, and the creative problem-solving that makes a customer feel helped instead of processed. The vigilance tax compounds. The screens multiply. And the cognitive athlete you hired for their humanity spends the shift auditing software, in seventeen tabs.

So here's the test for anything new in that stack: does it take a task off the agent, or add one? Most of it adds one: another screen, another output to verify, another thing to remember mid-call.

The phone is still where the interaction happens. Live agent calls are 61.6% of inbound interactions in ContactBabel's 2026 US Customer Experience Decision-Makers' Guide , and with phone self-service it's over 70%. In most verticals, digital channels (under 30% of interactions) still get more of the CX investment. And every tool in your stack quietly assumes the customer can understand what the agent is saying. In the same survey's 1,000-person customer panel, 19% said they can't hear clearly or have to repeat themselves very often, and another 34% said fairly often. ContactBabel is careful to say the question covers bad audio and accent difficulty alike, onshore or offshore. Its own back-of-envelope estimate, built on stated assumptions (30% of calls, 15 extra seconds each), puts the cost of all that repetition at $2.1B a year across US contact centers.

A peer-reviewed study, Wang et al. in the International Journal of Research in Marketing , shows what happens when they can't. Customers rated call center agents with non-standard accents lower only when the service outcome was unfavorable. Compare that against the boss-fight section above. Unfavorable outcomes are what hard calls produce, and hard calls are what's left for your humans.

Now Do the Call, but Also Be Someone Else

Taking load off an agent means removing a task, not adding a layer. There are plenty of ways to do it: fewer systems per call, less machine output to audit, hard calls routed to the right people. A hidden example that percolates budgets is a line item called “Accent Training,” often aimed at outsourced call center teams.

This looks responsible but gets complicated on a live call. Standard programs typically run three to six months before producing noticeable results, and outcomes vary a lot between agents based on aptitude and how much they practice off the clock. There's a hidden problem: it adds a third simultaneous task to an already overloaded call. The agent is managing the customer's problem, managing their own emotions, and now consciously performing a version of their own voice, all with a frustrated stranger on the line. That's an extraneous load, and under pressure on a hard call, agents revert to natural speech. The training performs in the classroom and evaporates on the floor. If yours is moving CSAT, keep it. Not sure? Get an answer before the next renewal.

There's a human cost here too: asking someone to suppress their natural voice as a condition of employment can wear down the confidence and presence you hired them for. On paper, it's a loop: training meant to help retention adds to the pressure that pushes agents out. The contact center guide ranks excessive pressure or stress fourth among reasons agents quit , behind wrong fit for the job, low pay, and lack of promotion.

Your agents are brilliant. You selected them for empathy and problem-solving, trained them, and handed them an impressive stack of little bots and AI tools. Then you sent them into the hardest calls in the queue, the ones the bots couldn't handle, carrying a communication barrier none of those tools address. Qualtrics' 2025 State of the Contact Center report found that nearly half of contact center employees say that if they had AI tools that made their work faster or easier, they would use the time saved to improve the quality of their work. Only 20% were actively using AI to resolve customer issues.

Your agents aren't resistant to technology. They're waiting for the right kind.

So the solution can't live in the agent's head. If the problem is structural (a cognitive task added to an overloaded call, a behavioral fix that evaporates under pressure, an identity cost that compounds), the fix has to operate at the structural level too. Not in the classroom. In the call itself.

Real-time speech technology doesn't ask the agent to change. It changes how the agent is heard, preserving voice, personality, and pace while removing the perceptual friction that triggers customer frustration. The three-task juggle collapses to two. The classroom-to-floor gap disappears because there's no classroom. The agent shows up as themselves, which is who you hired.

That's one example, but the test should apply to everything else in the stack: does it take a task off the agent, or add one?

Mind the Human-Shaped Gap in Your Model

Forrester's 2026 customer service predictions were blunt: the year would be about hard work, not AI transformation. Simplify, restructure, and lay the unglamorous plumbing AI needs before it can do anything useful. The organizations positioned to realize their AI investment are the ones doing that groundwork, not the ones stacking more layers on an unstable base.

The variable most AI ROI models leave out is the quality of the human interaction that catches everything the AI couldn't. Say your agentic layer deflects 40% of volume. The remaining 60% (the hard stuff, the emotionally loaded stuff, the relationship-defining stuff) lands on a cognitive athlete who is exhausted, under-supported, and working through communication friction on every call. The net customer experience doesn't improve. It fails at a different point in the journey. Forrester's 2026 numbers point the same way: 68% of brands didn't move on CX , and only 25% of US brands have a positive employee-experience impact . Whatever is holding those scores in place isn't in your routing logic. It's in the conversation.

The human side of that ledger has a price you can already calculate. ContactBabel's retention paper models the cost at $14,324 per agent who leaves, which comes to about $2.9M a year for a 500-seat center at the 41% attrition typical of large US centers. Adding absence together with attrition jumps that up to 3.5-5.5M a year. Yeah, they are estimates, but the vector is clear.

The balance you're looking for isn't more AI versus fewer humans. Americans would choose AI over a live agent in some situations , mostly for quick answers to simple questions. 34% never would, and 63% worry AI will make it harder to reach a human. Customers are least willing to use AI for ongoing issues and complex queries, which are exactly the calls your human agents are left with. The Qualtrics report lands in the same place: only 29% of consumers trust organizations to use AI responsibly, and 50% worry it will leave them with no human to connect with. That's a conversation problem.

The organizations that close the gap first won't be the ones who added the most layers. They'll be the ones who noticed the ROI was sitting in the conversation the whole time.

That's an operations story, not a technology one. Pick one queue, pull the hardest calls from last week, and listen to them the way your customer does. You'll know within a few calls.

Sources

Liked What You Read? See Balto in Action.

Balto helps leading contact centers turn insights into outcomes, in real time. Book a live demo to discover how our AI powers better conversations, coaching, and conversions.

Book a Demo Now