Your Customer Support Is Bleeding Money and a Computer Use Agent Can Stop It

The Klarna Warning Everyone Is Misreading

In February 2024, Klarna made global headlines. Their AI assistant was handling two-thirds of all customer service chats in its first month. They had effectively replaced 700 human agents. The CEO went on every podcast that would have him. It was the ultimate AI flex. Then in May 2025, Klarna quietly announced they were rehiring humans for customer service. Every tech journalist wrote the same lazy take: 'See? AI can't replace humans.' That's not the lesson. The actual lesson is that Klarna used a chatbot, not a computer use agent. Their bot could answer questions. It couldn't actually DO anything inside complex systems without breaking. It hallucinated policies. It frustrated customers with edge cases it wasn't trained on. The problem wasn't automation. The problem was shallow automation. A real AI computer use agent doesn't just generate text responses. It navigates your actual support software, looks up the actual order, applies the actual refund policy, and closes the actual ticket. That's a completely different category of tool.

The Air Canada Problem: Why Chatbots Are Legally Dangerous

Air Canada's chatbot told a grieving passenger he could buy a full-price ticket to his grandmother's funeral and apply for a bereavement discount afterward. That was wrong. The actual policy didn't work that way. Air Canada tried to argue in court that their chatbot was 'a separate legal entity responsible for its own actions.' A Canadian tribunal looked at that argument and basically laughed them out of the room. Air Canada lost, had to pay the refund, and got publicly humiliated in every major outlet. This is what happens when you automate the conversation but not the actual workflow. The chatbot had no real access to the policy database. It had no ability to verify what it was saying against live systems. It was generating plausible-sounding text, which is exactly what large language models do when they're not grounded in real system access. A computer use agent connected to your actual knowledge base, your actual ticketing system, and your actual policy documents doesn't hallucinate refund policies. It reads the policy file. It checks the ticket status. It acts on verified information.

What 'Automating Customer Support' Actually Means in 2025

• ●A real computer use agent opens your ticketing system (Zendesk, Freshdesk, Intercom, whatever) and works inside it like a human would. No custom API integrations required.
• ●It reads the customer's message, pulls up their account history in your CRM, checks order status in your fulfillment system, and drafts a response grounded in real data. All in one pass.
• ●Ticket deflection rates of 40 to 60% are achievable with proper AI automation. The average human-handled ticket costs between $15 and $40 depending on complexity. AI handling that same ticket costs roughly $0.60.
• ●Agent swarms let you run multiple computer use agents in parallel. A Monday morning flood of 500 tickets doesn't create a queue. It gets processed simultaneously.
• ●The 20 to 30% of tickets that genuinely need a human get flagged, summarized, and routed. The agent does the research so the human can focus on the actual decision.
• ●Unlike RPA tools that break the moment a UI changes by two pixels, modern computer use agents understand context. They adapt. They don't need to be reprogrammed every time your CRM updates.

Why RPA and Old-School Automation Keep Failing Support Teams

UiPath, Automation Anywhere, the whole legacy RPA stack. These tools were built for structured, predictable workflows. Customer support is neither of those things. A customer writes in about a billing issue that's also connected to a shipping delay that was caused by an address change they made three weeks ago. A UiPath bot sees that ticket and either misroutes it or escalates it to a human immediately because it doesn't match any pre-defined pattern. You paid six figures for the license and it handles maybe 15% of your actual ticket volume. The rest still lands on humans. The fundamental problem with RPA is that it's brittle. It follows scripts. Customer support doesn't follow scripts. Customers are chaotic and their problems are layered. You need an agent that can reason through a messy situation, not one that throws up its hands the moment something unexpected happens. That's precisely what a modern AI computer use agent is designed for.

How to Actually Build This: The Practical Playbook

Stop thinking about this as 'deploying a chatbot.' Start thinking about it as hiring a computer-literate agent that works inside your existing tools. Here's the actual approach that works. First, audit your ticket volume and categorize it. Most support teams find that 50 to 70% of tickets fall into about 10 to 15 repeatable categories: order status, refund requests, password resets, billing questions, shipping updates. These are your automation targets. Second, don't rip out your existing stack. A good computer use agent works on top of what you already have. It logs into Zendesk the same way a human agent does. It reads tickets, opens customer records in Salesforce, checks Shopify for order status, and responds. No six-month API integration project. Third, set clear escalation rules. Any ticket involving a complaint about discrimination, a legal threat, a high-value customer with a complex history, or an emotionally distressed customer goes to a human with a full briefing prepared by the agent. Fourth, measure what actually matters: first response time, resolution time, CSAT scores, and cost per ticket. If your CSAT drops after automation, your agent is doing something wrong and you need to fix it. If CSAT holds or improves while cost per ticket drops by 60%, you've built something real.

Why Coasty Is the Right Computer Use Agent for This

I've looked at the options. Anthropic's computer use is impressive research but it's a capability baked into a model, not a ready-to-deploy agent platform. OpenAI's Operator is promising but still early and limited in real enterprise workflows. Legacy RPA tools are expensive, brittle, and built for a different era. Coasty is the only computer use agent platform sitting at 82% on OSWorld, which is the industry benchmark for how well an AI can actually operate a real computer. Nobody else is close. That score isn't a marketing number. It means the agent can navigate real desktop applications, real browsers, and real terminals without falling apart on edge cases. For customer support specifically, this matters enormously. Your support stack isn't a clean API. It's a messy combination of a CRM your team has been customizing for four years, a ticketing tool with seventeen browser tabs open, and a Slack channel where escalations happen. Coasty works in that environment. It runs on a desktop app or cloud VMs, supports agent swarms for parallel ticket processing, and has a free tier so you can actually test it before committing. BYOK support means you're not locked into someone else's infrastructure decisions. It's the tool I'd recommend to anyone who is serious about automating support without the Klarna disaster or the Air Canada lawsuit.