Guide

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

Alex Thompson||8 min
Ctrl+R

U.S. businesses lose $856 billion every single year because of bad customer service. Not bad products. Not bad marketing. Bad support. And the wild part? Most companies trying to fix this are making it dramatically worse. Klarna is the poster child. They fired 700 human support agents, replaced them with AI chatbots, bragged about it publicly, and then spent 2025 quietly rehiring humans after their customer satisfaction collapsed. The internet noticed. Their CEO backpedaled so hard he basically admitted the whole thing was a PR stunt. But here's what nobody's actually saying out loud: Klarna didn't fail because they used AI. They failed because they used the wrong kind of AI. There's a massive difference between a dumb chatbot that reads from a script and a real computer use AI agent that can actually navigate software, update tickets, process refunds, and close issues the same way a trained human would. One of those things works. The other one is why 64% of customers now tell Gartner they'd prefer companies didn't use AI for support at all. That statistic should scare every product and ops leader reading this.

The Chatbot Graveyard Is Full of Good Intentions

Let's be honest about what most 'AI customer support' actually is in 2025. It's a decision tree dressed up in a chat bubble. It's a FAQ bot with a personality. It's the thing that makes customers type 'AGENT' in all caps until something human picks up. These tools have been around since 2016 and they've trained an entire generation of consumers to hate automated support on sight. The Gartner survey didn't come out of nowhere. Customers aren't anti-AI. They're anti-bad-AI. They've been burned too many times by bots that can't actually do anything, that loop them through the same three options, that cheerfully say 'I'm sorry to hear that!' while solving absolutely nothing. The real failure here isn't automation. It's that companies keep automating the conversation without automating the work. A bot that can chat but can't touch your CRM, can't issue a refund, can't update a shipping address, can't escalate with context, is just a very expensive way to make your customers angrier. You haven't automated support. You've automated the part where customers feel ignored.

What Automating Support Actually Requires

  • Reading and writing inside your actual support tools: Zendesk, Freshdesk, Intercom, whatever legacy system you're stuck with
  • Pulling up order histories, account records, and transaction data across multiple tabs and platforms simultaneously
  • Executing real actions: issuing refunds, updating account details, flagging fraud, escalating tickets with full context attached
  • Handling unstructured requests that don't fit a template, because real customers never write support tickets like FAQ answers
  • Working across browser, desktop, and internal tools without needing a custom API integration for every single system
  • Doing all of this at the speed of a fast human, not the 45-second lag of a chatbot waiting for a webhook to fire
  • Knowing when to hand off to a human and doing it with full context so the customer doesn't have to repeat themselves

Klarna fired 700 support agents, replaced them with AI chatbots, and then rehired humans after customer satisfaction tanked. The lesson isn't 'don't use AI.' It's 'don't use a chatbot and call it an AI agent.' There's a difference, and your customers can feel it immediately.

Why Traditional Automation Tools Keep Failing Support Teams

RPA tools like UiPath were built for structured, predictable workflows. Fill this form. Copy this row. Move this file. That's fine for back-office finance processes where nothing ever changes. Customer support is the opposite of that. Every ticket is a slightly different problem. Every customer has a slightly different account state. Every resolution requires judgment calls that a rigid automation script simply can't make. So companies spend six figures on RPA implementations, build hundreds of brittle rules, and then watch the whole thing break the moment a UI changes or a customer asks something slightly off-script. OpenAI's Operator and Anthropic's Claude computer use features are more interesting, but they're still research previews with real reliability gaps. Independent reviewers in 2025 called Operator 'unfinished, unsuccessful, and unsafe' for production use. Anthropic's computer use is impressive in demos and shaky in production. Neither was built specifically to be the best computer use agent for real enterprise workflows. They're general-purpose models with computer use bolted on. That distinction matters enormously when you're running support at scale and a bad resolution costs you a customer worth $500 a year.

The Right Way to Automate Customer Support With a Computer Use Agent

Here's the actual playbook. Step one: stop thinking in terms of 'chatbot' and start thinking in terms of 'AI employee who can use a computer.' A real computer use agent doesn't need API access to your tools. It sees what a human sees, clicks what a human clicks, and works inside your existing software stack without a six-month integration project. That means you can deploy it against Zendesk, Salesforce, your internal order management system, your shipping portal, and your billing software all at once, without writing a single integration. Step two: start with the highest-volume, lowest-complexity tickets. Shipping status updates. Password resets. Subscription changes. Refund requests under a certain dollar threshold. These are the tickets that eat 60-70% of your team's time and require almost no judgment. A good computer use AI agent handles these end-to-end, not just the conversation but the actual resolution, in under two minutes. Step three: run the agent in parallel with your human team first. Don't pull a Klarna. Let it handle tickets autonomously while humans review the outputs. Tune it. Build trust in it. Then scale. Step four: use the humans you still have for what humans are actually good at: complex escalations, angry customers who need empathy, edge cases, and relationship-building with high-value accounts. The companies doing this right aren't replacing their support teams. They're making their support teams 10x more effective by removing all the repetitive garbage from their queues.

Why Coasty Exists and Why It's Different

I've looked at most of the serious computer use AI options out there. Coasty is the one I keep coming back to, and the reason is simple: it's the only one built from the ground up to be the best computer use agent, not a language model that learned to click things as an afterthought. It scores 82% on OSWorld, the industry-standard benchmark for AI computer use performance. Nobody else is close. That's not a marketing claim. OSWorld is a third-party benchmark that tests agents on real-world computer tasks, and 82% is a genuinely hard number to hit. In practical terms, that means Coasty can navigate your actual support tools, handle multi-step workflows across different applications, and execute resolutions reliably enough to trust with real customer interactions. It controls real desktops and browsers, not sandboxed demos. It supports agent swarms for parallel execution, so if you have 500 tickets in the queue, you don't wait for them to process one at a time. There's a free tier to actually test it before you commit, and BYOK support if you care about where your data goes. The honest pitch is this: if you're going to automate customer support, you need a computer use agent that can actually use a computer at a professional level. Coasty is currently the only one that clears that bar in production.

Here's my take, and I'll stand behind it: the companies that are going to win at customer support over the next three years aren't the ones with the biggest human teams, and they're not the ones with the most aggressive chatbot deployments. They're the ones that figured out how to use a real computer use AI agent to handle volume without sacrificing quality, and then pointed their human talent at the problems that actually need humans. The Klarna disaster wasn't a warning against AI. It was a warning against lazy AI. Against treating a support ticket like a chat conversation instead of a task that needs to be completed inside real software. If you're still running manual support queues, you're burning money that your competitors are saving. If you're running chatbots that can't actually do anything, you're actively training your customers to distrust you. Neither of those is a strategy. Go try Coasty at coasty.ai. Start with your ten most common ticket types. See what an 82% OSWorld score looks like when it's working through your actual Zendesk queue. Then decide if you still want to pay a person to copy-paste order numbers all day.

Want to see this in action?

View Case Studies
Try Coasty Free