How is AI being used in customer service?

The Customer Service Pressure Cooker — and the AI Trust Gap

Customer service has become a pressure cooker: expectations keep climbing while budgets and headcount can't keep pace. The result is a market racing to solve the problem with AI — and a growing number of customers who aren't sure they want it solved that way.

The demand side of the equation is unforgiving. According to industry research, 88% of consumers now expect faster responses than a year ago, and 74% expect service around the clock. Meanwhile, support industry data shows 77% of customer service reps report their workload and issue complexity have increased versus a year ago.

Businesses are responding with investment. The AI customer service market is projected to grow from $12.06B in 2024 to $47.82B by 2030 — a 25.8% compound annual growth rate. Gartner forecasts that agentic AI could autonomously resolve 80% of common service issues by 2029, cutting operational costs by 30%.

But here's the tension: the people on the receiving end aren't nearly as enthusiastic as the executives writing the checks.

• 79% of Americans strongly prefer speaking with a human agent, even as 96% of executives believe GenAI will enhance customer interactions.
• 50% of consumers would cancel a service if it were solely AI-driven, and 41% feel customer service has already worsened because of AI.
• 81% of consumers believe businesses adopt AI mainly to save money — not to improve service.

That gap between leadership confidence and consumer trust is the defining challenge of AI in customer service. Implementation quality, not the technology itself, determines which side of the gap a company lands on. As Intercom's Fin team puts it, "AI rarely succeeds when it is added as a thin automation layer" — implementation research consistently shows the operating model matters more than the model.

The stakes are also asymmetric. Consumer research found that 77% of customers say a poor self-service experience is worse than no self-service at all. In other words, bad AI is worse than no AI — it actively damages trust rather than merely failing to help.

This is the tension the rest of this article explores: AI is mission-critical for meeting modern expectations, but deploying it carelessly costs more than it saves. Teams building AI-driven support workflows need ways to verify what actually works before customers find out the hard way — the same principle behind SkillzBench's approach of testing every agent skill in a sandbox before recommending it.

Where AI Actually Works: Tickets, AI Agents, and Self-Service

Not every AI deployment in customer service delivers, but three application areas consistently do: augmenting support tickets, autonomous AI agents, and self-service. The difference between success and failure usually comes down to one philosophy — hybrid AI plus human, not replacement.

First, agent-assist tools are quietly transforming the ticket queue. According to industry data, agent-assist adoption (already at 40% of teams) cuts average handle time by 27%. The agents themselves are believers: support research shows 79% say an AI copilot supercharges their abilities. That matters because 77% of reps report their workload and issue complexity have grown year over year.

Second, chatbots are evolving from scripted decision trees into agentic AI that can actually resolve problems. Gartner forecasts these systems will autonomously resolve 80% of common customer service issues by 2029, cutting operational costs by 30%. Unlike legacy bots, modern agents interpret intent, retrieve account data, and complete actions across systems — Nutribees, for example, reduced human-handled tickets by 77% using one.

Third, self-service deflects tickets before they exist. The numbers make the case:

• 69% of consumers prefer AI-powered self-service for quick resolution
• Companies report up to 70% fewer call, chat, and email inquiries after deploying a virtual assistant
• 92% of consumers would use an online knowledge base if one were available
• 80% of high-performing service organizations offer self-service, versus just 56% of low performers

The caveat: research also shows 77% of consumers say a poor self-service experience is worse than none at all. Quality knowledge content is the foundation.

What ties all three areas together is the hybrid model. As implementation guidance puts it, "the goal is not fewer agents — it is better use of human judgment on complex, high-impact cases." Every credible source agrees: AI handles routine volume, humans handle sensitive complexity. Teams building these agentic workflows can browse and test free agent skills on SkillzBench to automate the routine parts while keeping humans firmly in the loop.

Why AI Deployments Fail: Thin Automation and Broken Handoffs

Ninety-six percent of executives believe generative AI will enhance customer interactions, yet 41% of consumers say customer service has actually gotten worse because of it. The gap isn't the technology — it's how the technology is deployed.

When AI is bolted onto an existing support stack as a thin automation layer, it rarely succeeds. As Intercom's implementation team puts it bluntly: "The technology works. The operating model does not." The AI itself can retrieve information, interpret intent, and complete actions — but without a supporting operating model, it produces frustration instead of resolution.

The single biggest implementation challenge isn't model selection or vendor choice. It's knowledge readiness. Fragmented documentation — where outdated policies sit alongside internal notes and customer-facing content — causes AI to blend it all together, producing answers that are plausible but wrong. That's not hallucination in the technical sense; it's poor resolution quality rooted in messy source material. AI amplifies whatever knowledge base it's given, garbage included.

Escalation design is the second major failure point. According to SurveyMonkey's consumer research, only 15% of customers experience a seamless handoff from AI to a human agent. The other 85% repeat themselves, lose context, or fight the system just to reach a person — even though 89% of consumers say a human option should always be available. A handoff that drops conversation history isn't a handoff; it's a restart.

Perhaps the most striking failure is organizational. Zendesk's CX research reveals a stark perception gap between leadership and frontline agents:

• 72% of CX leaders believe they've provided adequate AI training
• 55% of agents say they've received no AI training at all
• Only 21% of agents are satisfied with the training they did get
• Only 34% of agents even understand their department's AI strategy

When agents don't understand the AI they're working beside, resistance follows — and it's operational, not cultural. Teams push back when AI creates cleanup work or threatens service quality, not because they fear change. The fix is equipping agents properly: clear training, real tools, and workflows where AI handles routine volume while humans keep judgment on complex, high-impact cases.

The pattern across every failed deployment is the same: leadership bought technology, but nobody built the operating model around it. Teams assembling that model — connecting agents to vetted, tested capabilities rather than raw, unverified tooling — can evaluate and benchmark agent skills for free on SkillzBench before putting them in front of customers. Skills, proven — not just installed.

Your Implementation Playbook: 5 Steps to AI That Customers Actually Like

The difference between AI that customers tolerate and AI they appreciate comes down to sequencing, not technology. Teams that bolt AI on as a thin automation layer fail; teams that follow a deliberate rollout consistently see results like 90% positive ROI, according to 2025 CX research.

Step 1: Start with routine, high-volume work. Ticket routing, FAQ deflection, and agent-assist tools deliver value without gambling on the customer relationship. Roughly 75% of inquiries can be resolved by AI without human intervention, yet half of consumers say they'd cancel a service that became fully AI-driven, per SurveyMonkey's consumer research. That gap is your roadmap: automate the repetitive volume, keep humans on everything complex or sensitive.

Step 2: Fix your knowledge base first. Fragmented documentation is the top implementation challenge, causing AI to blend outdated, internal, and customer-facing content into plausible-but-wrong answers — not hallucination in the technical sense, just poor resolution quality. And the stakes are high: 77% of consumers say a poor self-service experience is worse than no self-service at all. Clean, authoritative content is the foundation everything else sits on.

Step 3: Design escalation paths with full context handoff. 89% of consumers say companies should always offer a human agent option, but only 15% experience a seamless AI-to-human handoff. Build escalation so the agent sees the complete conversation history, customer data, and any actions already taken. Customers should never have to fight the system to reach a person.

Step 4: Be transparent about AI use. 74% of CX leaders agree AI transparency is paramount, and 14% of consumers would lose trust in a business whose AI agent doesn't disclose it's AI. Label the bot, offer a human option, and keep people in the loop for sensitive cases. Automation that hides itself reads as deception, not efficiency.

Step 5: Close the training and tooling gap. Here's the uncomfortable one: while 72% of CX leaders say they've provided adequate AI training, 55% of agents report receiving none, and only about 20% have generative AI tools at their disposal. Meanwhile, 79% of agents believe an AI copilot supercharges their abilities. The appetite exists — leaders just aren't feeding it.

• Audit which ticket types are truly routine before automating anything.
• Treat knowledge base cleanup as a prerequisite, not a parallel project.
• Test your escalation flow by trying to reach a human yourself.
• Survey agents on the tools and training they actually have.
• Measure resolution quality, not just deflection rates.

For teams building their own support agents and workflows, the quality of the underlying agent skills matters as much as the rollout plan — with published research finding roughly 26% of available skills carrying at least one vulnerability, it pays to vet what you install. SkillzBench is building a free directory that benchmarks and security-scans agent skills so you can equip your support agents with confidence; join the waitlist to be first in line.

Equip Your Agents — Human and AI Alike

There's a quiet gap at the heart of customer service AI: leaders believe in it, but the people using it every day often don't have it. While 72% of CX leaders say they've provided adequate AI training, only 21% of agents are satisfied with what they received — and barely 20% have generative AI tools at their disposal. Closing that tooling gap is where AI adoption actually succeeds or stalls.

The case for equipping agents is strong. 79% of support agents believe an AI copilot supercharges their abilities, and agent-assist tools already show a 27% reduction in average handle time. The winning model is hybrid: automation for routine, high-volume work, and human judgment reserved for complex, sensitive cases where empathy earns loyalty.

But handing agents — or AI agents — new tools without vetting them creates new risk. Support teams building AI workflows face the same question as any agent builder: which skills and tools actually help, and which introduce security or quality problems? That's the discipline SkillzBench is being built around. Every skill listed goes through the same gauntlet:

• Security scanning and license checks, since published research found roughly 26% of agent skills carry at least one vulnerability
• Sandbox benchmarking with the same model, with and without the skill, so the measured Lift in percentage points is published separately from ratings
• Version pinning to a commit SHA, so what you tested is exactly what you deploy
• A 0-100 trust score built from security, quality, license, maintenance, and community signals — with money never touching the scores

That last point matters for a field where consumer trust is fragile. 14% of consumers would lose trust in a business if an AI agent doesn't disclose it's AI, and 50% would cancel a service that's solely AI-driven. Deploying unvetted automation into that environment is how the "thin automation layer" failure happens — the technology works, but the operating model doesn't.

The same verification mindset applies whether you're equipping a human agent with a copilot or wiring an AI agent to handle ticket triage. Know what a tool does, prove it improves outcomes, and give it only the access it needs. SkillzBench's free MCP server is designed for exactly that: connect any agent, toggle only the skills and tools you've chosen, and the agent sees nothing more.

If you're an agent builder or a support team assembling AI workflows, browse the verified skills directory, test what fits, and join the waitlist for launch. Your agents — human and AI alike — deserve tools that are proven, not just published.