Two years. That's how long one developer held out against car ownership after moving somewhere that absolutely required one. The plan was simple: let AI do the heavy lifting. Just describe what you need, get perfect listings, pick a winner. It didn't quite work out that way. The experiment started promising enough. Claude whipped up model recommendations based on stated requirements—genuinely helpful for someone who admits total car ignorance. But then came greed. The developer asked for actual Autotrader listings to browse directly through the chat interface. What came back were dead links, each one redirecting back to the search page instead of individual vehicle ads. The culprit? Claude wasn't pulling live data at all. It was reading stale, cached snapshots of Autotrader's results. Individual listing pages are blocked from automated access entirely, returning a clear ROBOTS_DISALLOWED error whenever the web_fetch tool tried to open them. There's a workaround—the official Claude Chrome extension can run on Autotrader's search results page—but it's clunky and doesn't sync with the main conversation interface (hence the abysmal 2.6 rating in the Chrome Web Store).

MOT Results Hit Another Brick Wall

The developer then attempted something clever: fetch registration numbers from listing photos, then pull each car's UK MOT inspection history from gov.uk to assess vehicle condition and value. The prompt was elegant in theory. In practice, every single MOT lookup triggered a CAPTCHA challenge, completely blocking automated access. Undeterred, the developer discovered that HM Government's own API exposes this data—but only behind authentication. A local proxy with API credentials solved that problem, creating a working pipeline for vehicle history checks. Progress!

The Gap Between AI Hype and Reality

The final workflow required running searches manually for each car model because the Chrome extension crashes frequently during extended sessions. For someone wanting to compare a dozen options across different manufacturers, this meant repetitive manual steps rather than a single automated sweep. The developer admits the real ending: frustration won out. They went with a listing their wife found instead of waiting for the AI workflow to stabilize. The system got built properly only after the purchase was complete, as a proof-of-concept for next time.

Key Takeaways

  • AI agents still struggle with live e-commerce data—caching and bot blocks defeat basic shopping tasks
  • CAPTCHAs remain highly effective against automated vehicle history lookups without API access
  • The gap between "Claude can help" and "Claude works in production on real tasks" is substantial

The Bottom Line

This isn't a failure of AI capability—it's a reminder that the web wasn't built for agents. Until sites open APIs or change bot policies, your digital assistant will keep hitting walls when you need it most. Build your own proxies or stay frustrated.