Here's a number that should give every operations director pause: 87%.
That's the median increase in human agent workload after customer-facing AI deployments, according to Zendesk's latest research cited in CX Foundation's workforce analysis published this week.
Read that again. Not a decrease. An increase. By nearly double.
For two years, the promise has been simple: deploy AI, reduce call volumes, cut costs. But something else is happening in practice. And understanding why matters more than chasing the next chatbot feature.
Why workloads are going up, not down
The explanation is more interesting than a simple "AI doesn't work." It does work—and that's precisely the problem.
Customers now trust brand AI enough to ask questions they previously wouldn't have bothered with. When your chatbot can actually answer "where's my order" reliably, customers start asking follow-ups. "Can I change the delivery address?" "What if I'm not home?" "Can I add something to the order?"
Each of those becomes an escalation to a human. The easy queries get handled automatically; the edge cases land on your agents' desks.
According to Valoir's October 2026 benchmarks study, almost three-fifths of contact centres have no plans to reduce headcount despite 95% running at least one AI capability. Many are actively hiring. The work hasn't disappeared—it's transformed.
The concentration effect
What used to be a mixed bag of simple and complex queries is now concentrated difficulty.
Think about the typical agent day before AI: answer fifty calls, thirty of them straightforward ("what's my balance?"), fifteen moderately complex, five genuinely hard. The easy calls gave breathing room. The rhythm was manageable.
Now? The thirty easy ones never arrive. The AI handles them. Your agent's day is fifteen moderate calls and five hard ones, back to back, with no relief.
That's not less work. It's more demanding work, compressed into the same hours.
52% of organisations report that AI-to-human escalations are increasing, per Frost & Sullivan. The escalation isn't a failure—it's the system working as designed. But it creates a new problem: how do you support agents doing nothing but complex, emotionally charged work all day?
What the successful contact centres are doing
The contact centres getting this right aren't fighting the workload shift. They're redesigning around it.
They've restructured their teams. Firstsource, for example, now requires team leaders to spend one day per month handling customer contacts themselves and receiving QA on their own interactions. The goal: leaders who actually understand what agents face, not managers who haven't taken a call in years.
They've changed their metrics. Thirty percent of contact centres are revising KPIs because of AI, according to Valoir. The shift is away from efficiency measures like average handling time and toward outcome measures like first-contact resolution and customer sentiment. If every remaining call is complex, measuring talk time punishes agents for doing their job well.
New metrics emerging include:
- Case complexity scoring — predicting difficulty before routing, so agents don't get five brutal calls in a row
- AI adoption tracking — measuring whether agents actually use the assist tools available
- Coaching success rates — tracking which interventions actually improve performance
They've invested in better handoffs. The AI-to-human transfer is where most customer experiences break down. The best implementations pass full conversation context so agents don't start from zero. Some are going further: letting agents interact with the AI to get additional context before accepting the transfer, rather than being thrown into the deep end.
The tools that actually help
Agent Assist started as a way to surface knowledge articles during calls. It's becoming something more useful: a Swiss Army knife that handles the administrative burden agents resent most.
Real-time knowledge surfacing — relevant information appears alongside the conversation, eliminating the "please hold while I look that up" pause.
Auto-generated summaries — the call ends, and the CRM notes are already drafted. Agents review and tweak rather than type from scratch. That 30-60 second saving per call adds up to hours per week.
Cross-system actions — the emerging "headless" capability where agents can pull data and perform actions across different tools through natural language commands. Still early days, but the contact centres piloting this report meaningful efficiency gains.
The key distinction: these tools reduce administrative work without adding complexity to the customer conversation itself. They make the hard work easier to execute, rather than trying to eliminate the hard work entirely.
What to actually do
If you're planning AI deployment or already live with disappointing results, the research points to specific changes:
Audit where escalations land. Map the queries that transfer from AI to human. Are they edge cases that genuinely need human judgment? Or gaps in your AI's training that could be closed? Often it's both—but the ratio matters.
Redesign agent schedules around complexity. If every call is hard, agents need more recovery time between them. Shorter shifts, more breaks, rotation into offline work. The old scheduling assumptions don't apply.
Train for the new job. The work has changed. Agents now need deeper product knowledge, better de-escalation skills, more autonomy to resolve issues. Treating them as interchangeable cogs handling standardised queries doesn't match reality anymore.
Share productivity gains. If AI makes your operation more efficient, some of that should flow to the people doing the remaining (harder) work. The facilities management company that shifted 43% of calls to AI saw agent satisfaction rise—partly because the job improved, partly because compensation reflected the increased skill requirement.
Measure what matters. Track resolution, not containment. Track agent satisfaction alongside handle time. If your AI containment rate is excellent but your best agents are leaving, you haven't solved anything.
The honest assessment
AI in customer service works. It handles routine queries reliably. It reduces overall costs when implemented thoughtfully. The organisations getting value from it—and there are many—aren't doing anything magical.
What they're not doing is assuming AI deployment automatically makes agent work easier. The workload paradox is real: successful AI increases customer confidence, which increases engagement, which increases escalation complexity.
The contact centres thriving in this environment have accepted that reality and redesigned around it. They treat AI as a way to shift agent work toward higher-value activities, not eliminate agent work entirely. They invest in their people because the remaining work actually requires skilled people.
That's not the headline the AI vendors want to write. But it's closer to what actually happens.
Hostcomm provides AI voice agents and contact centre technology for UK businesses. If you're navigating the workload paradox—and want to get the agent experience right alongside the automation—we'd welcome the conversation.