The AI Architecture Problem in Customer Service

Klarna recently cut 700 customer service jobs, betting AI could absorb the difference. Within a year, customer satisfaction had dropped and the company was hiring again. Gartner predicts that by 2027, half of the organizations planning similar workforce cuts will abandon those plans too.

This isn’t an AI capability problem. 100% of contact center agents interact with AI daily to handle a wide variety of complex situations. 

Yet zero percent of agents consider AI critical to their work. That gap isn’t surprising. AI keeps getting layered onto the same fragmented systems that were failing agents before advanced AI tools exploded onto the scene. Today, humans still do the connective work, the tab-switching, the copy-pasting, the manual reconciliation that ties one system to another. 

AI was built to replace that labor. Instead, it complicated it. Until someone fixes the connections between systems, agents will keep doing AI’s work by hand, and lose trust in the most important tool at their disposal. 

The Trust Gap No One Is Fixing

Seventy-eight percent of agents say AI tools haven’t meaningfully changed how they work. Ninety-three percent verify AI outputs before acting on them. Agents see enough incomplete or inaccurate responses that double-checking is the default. They aren’t resisting the technology. It hasn’t earned their trust yet.

The consequences become visible in the middle of a customer conversation. A customer calls about a billing dispute, and the AI-generated account summary is three months out of date. They spend the next ten minutes verifying across multiple systems what AI was supposed to surface instantly, all while the customer sits on hold.

Most organizations read that negative experience as a people problem. Agents are scared that their job is at risk due to AI taking on more complex tasks, so they work around it. 

It’s actually a leadership problem. Gartner found that 91% of customer service leaders are under executive pressure to implement AI.

Under that kind of pressure, it’s easy to interpret pushback as an adoption issue rather than a signal that AI is not architected correctly. As a result, the focus remains on accelerating deployment in the wrong areas, and increasing headcount reductions without fixing the underlying problem. 

What Fragmented Deployment Actually Produces

Agents serve as the manual integration layer between a live customer conversation and 4 to 10 disconnected enterprise tools. They “swivel-chair” between billing platforms, ERP systems and ticketing tools while a customer waits on hold. That’s because most enterprises bolted AI onto the same fragmented architecture that was already slowing agents down. 

For example, chatbots can handle the front-end of an interaction but fail when agents need to access the legacy back-office system where the actual transaction lives, forcing agents to complete the task manually anyway. Or, AI summarization tools can generate post-call notes but they write to a separate system rather than the CRM of record. AI inherited the fragmentation instead of fixing it. 

AI is only as connected as the systems beneath it. When systems don’t share context, AI can’t either. It pulls from historical CRM records, but it can’t understand what’s happening in a live customer conversation. By the time it surfaces a recommendation, the underlying data may have changed.

Incomplete context, stale data, and disconnected systems aren’t separate problems. They’re symptoms of an AI architecture failure.  

Building AI Architecture That Works with Human Agents

Building AI that agents actually trust, and that improves customer service outcomes, comes down to four infrastructure requirements: 

1. Provide AI tools with real-time data access. Most AI deployments in CX pull from whatever customer data is available. It’s often outdated. AI needs to be architected to provide real-time access to customer data at the moment a request is made, not a fragmented snapshot assembled from systems holding different versions of the same customer. 
2. Give AI persistent context. When AI is bolted onto existing infrastructure rather than built into it, context doesn’t travel. It gets dropped at every handoff. Persistent AI should be layered across the entire interaction, continuously carrying customer intent and conversational context as work moves between systems, channels, and human agents. When a human agent steps in, they shouldn’t be starting from scratch. 
3. Implement agentic AI for autonomous execution, not just recommendations. The next meaningful leap is AI that can execute the work agents require autonomously while keeping the human in the loop for observability and control. Rather than telling agents what to do next, agentic AI orchestrates workflows across enterprise systems, filing claims, processing refunds, updating records, and completing tasks even across legacy applications without APIs. AI handles execution while human agents focus on judgment, empathy, and building stronger customer relationships. 
4. Choose AI systems that improve from outcomes, not just inputs. Most enterprise AI is static after deployment. The more impactful architecture is one that continuously learns from interaction outcomes, optimizing workflows based on what actually resolved the problem, not just what the training data predicted would. AI that closes the feedback loop between outcomes and future performance is fundamentally different from AI that deploys and holds.

The trust gap agents have with AI isn’t a training problem or an adoption problem. It’s an architecture problem. 

The true ROI of AI isn’t in reducing headcount, it’s in the value created when agents can finally focus on customers instead of systems. That only happens when AI is built into the execution layer, not dropped on top of it. 

The organizations still struggling with adoption aren’t behind on AI. They’re behind on infrastructure. Fix the architecture, and the adoption problem solves itself.