The Ghost in the Switchboard: Lessons From 81 Million Support Conversations

There is a quiet, dangerous fiction currently circulating in boardrooms and C-suites. It says that AI has finally solved the cost problem of support. The logic is simple: automate the routine questions, watch the volume drop, and reduce the headcount. If you look at the macro-level dashboards, it seems to be working. Efficiency is up, and ticket deflection is reaching record highs.

But if you actually sit with a support team today—if you watch the pace of their internal chats and see the look on an agent’s face at 4:00 PM on a workday—you know the reality is different. Support is not getting lighter. It is getting denser. 

And that density is landing squarely on the humans who remain.

At Hiver, we never set out to build for the bloated complexity of the Fortune 500; we built for the agile, fast-moving teams who want to solve problems without the overhead of a legacy factory. Last year, our platforms helped our customers resolve 81 million conversations. It’s not enterprise volume, but even within this SMB bubble, we are seeing a shift that mirrors, and in many ways, magnifies, the crisis happening in the enterprise world. 

The industry is sleepwalking into a talent crisis. We are witnessing a global erosion of Tier 1 support. In our rush to automate the easy work, the ‘AI-powered’ support platforms have accidentally broken the path to expertise.

You can call this “transformation” if you wish, but the choice of vocabulary does not mask the fact that the entry point into support expertise is quickly and quietly disappearing. What makes this moment different from past cycles of efficiency is that the underlying economics are beginning to move in the opposite direction of the set narrative. 

This is not a temporary management failure. It is a systems shift.

For more than a century, customer service has been governed by the invisible assumption that support is a plumbing problem. That assumption traces back to 1917 and a mathematician named Agner Krarup Erlang. He developed Erlang-C, a formula to tell telephone companies how many operators they needed for a given number of calls. The formula treats humans like pipes: if the water level rises, you simply widen the pipe. Ergo, if you get more calls, you simply add more people to handle the inbound.

That switchboard logic still hums beneath the surface of every modern helpdesk. We talk about AI-first futures and supercharged agents, yet we continue to manage support with relic metrics like Average Handle Time and Occupancy. AI promised escape from the plumbing. Too often, it has simply become the newest pipe fitted into an ancient machine.

Let me illustrate this further.

Last year, Hiver’s infrastructure handled 85 million manual tasks through automations—the password resets, the status updates, and the routing rules that used to be the bread and butter of an agent's day. By the logic of the old switchboard math, removing 85 million units of work should create a vast surplus of human capacity. And if that logic were still valid, this surplus would be visible across the entire support economy.

Instead, industry-wide, 77% of agents report that complexity is increasing, while other reports say that 56% of the support workforce shows record levels of burnout. 

The math fails because it forgot that those 85 million easy tasks were not merely work—they were recovery time, learning time, and cognitive breathing room. Remove repetition without redesigning the human system around what remains, and the result is not efficiency but compression.

The Cornell Global Call Center Project has spent years documenting why this "plumbing" model fails. Their 2007 report research proved for the first time that high-performance teams thrive on discretion, yet the industry continues to manage them for volume. 

It’s taken the support industry almost two decades to catch up, but we are finally talking “cognitive overload” and “human judgement”–catchphrases that other high-stakes systems adopted way earlier as they adopted automation and augmented intelligence. 

In aviation, for instance, autopilot did not eliminate pilots; it transformed them into managers of rare, high-stakes exceptions where judgment mattered more than routine control. In medicine, diagnostic imaging software did not replace radiologists; it concentrated their attention on the ambiguous scans where experience determines outcome. 

In each case, automation removed repetition but preserved the learning pathways that create expertise. Customer support is now entering that same historical phase—only faster, and without having protected the path that produces the experts required.

Can we build the experts of tomorrow if we’ve paved over the path today?

There is a second, more dangerous side effect to this automation that the industry is largely ignoring. We are consciously deleting the apprenticeship ladder. 

According to research, it takes roughly 16 weeks for a new agent to become truly proficient. Historically, those 16 weeks were spent in Tier 1. A junior agent learned the product and the specific language of the customer by handling hundreds of simple, low-stakes requests.

But today, we are throwing new hires directly into the deep end. We see the consequences of this every day. A customer sends an email saying something on the left side of the panel looks broken. To a senior agent, that is a clear reference to a specific activity module. But to a junior agent who hasn't been allowed to handle the 1,000 simple tickets that build that mental map, it is an ambiguous riddle.

This gap in language creates a massive expertise debt. We have no way to grow the next generation of seniors because we have automated away all the entry-level lessons. It is why Gartner predicts that 50% of organizations that cut staff for AI will be forced to rehire for those same roles by 2027. 

They are realizing that while AI can talk, it cannot replace the "expertise, empathy, and judgment" that is only built through experience. We have paved over the stepping stones and now wonder why no one can cross the river.

Every profession relies on a visible path from novice to expert, a period where mistakes are small, stakes are low, and pattern recognition quietly accumulates. When that middle ground disappears, expertise stops renewing itself. The immediate symptom is slower resolution and higher escalation. The long-term consequence is far more severe: a hollowed-out workforce where senior judgment becomes both scarce and fragile. By the time the shortage is visible in executive dashboards, the learning cycle required to fix it has already been broken for years.

So, how do teams survive when no single person can hold all the answers?

This pressure has triggered a visible survival mechanism inside the inbox. This year, Hiver users created 5 million internal notes directly on top of customer conversations. This is not just a collaboration statistic. It is a signal that the era of the lone wolf agent is over.

In the old model, one person owned one ticket until it was closed. But when every ticket is a complex exception, no one person can hold all the answers. Those 5 million notes represent a collective huddle. By swarming around the problem inside the conversation, teams are building a "shared brain" in real time. They are redistributing the cognitive load so that no one individual has to carry the weight of a complex crisis alone. 

But collective intelligence is only one side of survival. The other is economic reality.

Because we have commoditized speed through AI, the metrics we have used for decades are losing their meaning. There is a new metric emerging that should worry every CFO: the rising cost of automation.

Gartner’s latest analysis suggests that the dream of "free" support is ending. By 2030, the cost per GenAI resolution is expected to exceed $3; surpassing the cost of many offshore human agents. The subsidies are ending, and the infrastructure costs are mounting.

Remember how I said earlier that the underlying economics are beginning to move in the opposite direction of the narrative? For more than a decade, automation has been justified primarily through labor arbitrage—the assumption being that software becomes cheaper as it scales. But when the marginal cost of automated resolution begins to rival or exceed human labor, the strategic question changes entirely. 

AI can no longer be defended purely as a substitute for people. It must justify itself as a multiplier of human capability. That distinction is profound. Substitution reduces cost in the short term. Multiplication determines resilience in the long term. The organizations that understand this shift early will stop asking how to remove humans from support and start asking how to make the remaining humans dramatically more effective.

And effectiveness, in this new landscape, is no longer defined by correctness alone.

Last year, our customers used a feature called “Sentiment Analysis” over 3.6 million times to understand emotional context in support conversations. This shows us that a correct answer is no longer enough or effective. 

Customers arrive at a human interaction with a sentiment deficit. They are tired and frustrated by the bots they just fought. In this environment, the job of the agent has shifted from information delivery to emotional restoration. The goal is no longer “Resolution”; it is “Emotional Equilibrium.”

So are you still managing a factory, or are you ready to support experts?

If this diagnosis is correct, then the implication for leadership is not incremental improvement but managerial reorientation. The central question of support is no longer, “How do we handle more tickets with fewer people?” It is, “How do we compound human judgment where automation stops while ensuring that the right handoffs happen at the right moment and context is never lost through the tool sprawl and the shared brain?” And that is a long question, indeed. 

Organizations that continue to optimize for throughput will appear efficient right up until the moment customer trust erodes and talent flight accelerates. Organizations that instead invest in context, collaboration, and experiential learning will look slower in the short term but structurally superior in the long term. The difference between those paths will define the next decade of customer support and experience.

Support leaders are simultaneously facing three converging pressures: the rising cognitive cost of each remaining interaction, a shrinking pipeline of entry-level talent capable of growing into expertise, and an automation layer whose marginal savings are flattening as infrastructure and model costs mature. 

When complexity rises faster than productivity, the system constricts. This is why the coming disruption in support will not be gradual. It will feel sudden, not because the forces are new, but because they have quietly crossed the threshold where incremental optimization can no longer compensate for structural strain.

The crisis we are facing is not a lack of technology; it is an attachment to a 100-year-old management philosophy. If we want to solve the expertise debt and prevent the L1 erosion, we have to stop building tools that isolate agents and start building environments that recognize and facilitate the shared brain.

Support can no longer scale by deflecting customers; we can only scale by empowering the people who act as trust keepers in high-risk situations. The next decade won’t be won by the company with the most efficient bot, but by the one that preserves the most capable, calm, and context-rich humans.