Watch an AI agent fail spectacularly to enter a single invoice into an accounting system.
On The Accounting Podcast, I tested Perplexity’s new Comet browser - their AI assistant that promises to work across browser tabs and handle repetitive tasks. The test was simple: take a basic PDF invoice (one line item, $85 for web design) and enter it as a bill in Xero.
I had the Xero dashboard open in one tab and a PDF invoice open in another. Nothing complicated - just one line item for web design at $85/hour. I prompted the AI to enter the invoice in the other tab as a bill in Xero.
What happened next was... interesting. The AI started reasoning through the task, pulling data from the PDF, and navigating to the bill section. It's basically taking screenshots, running prompts, taking actions, over and over.
Something concerning - it's all happening in a hidden browser tab. I can't actually see it clicking around, which makes me nervous. I'm definitely not opening my bank account in this browser anytime soon.
The AI struggled a lot. It kept trying to enter information, clicking in the wrong places, and retrying, just like someone who had never seen Xero before. It managed to create a draft bill with the vendor name, dates, and amount, but it couldn't figure out how to assign an account code—something required to save a bill in Xero. It just kept getting an "account must be valid" error and couldn't work through it.
We're nowhere near AI agents in browsers replacing accountants at the moment. This first version of agentic AI is like GPT-2 - showing potential but struggling with basics. It couldn't even ask for help when stuck, unlike a human intern who'd eventually ask for guidance.
But I see the glimmer of potential. Maybe in a few iterations, we'll have browser-based AI that can monitor an email inbox for bills and enter them automatically. Or better yet, go find missing receipts from vendor websites and attach them to transactions in the accounting system.
But right now, the hype and reality are very, very far apart.
