Converting plain english descriptions into working browser automation—how much tweaking do you actually need?

[PixelPioneer88]

I’ve been experimenting with AI copilot workflow generation for the past few weeks, specifically trying to automate some repetitive form filling on a vendor portal we use. The promise is that you describe what you want in plain English and it generates a ready-to-run workflow.

Here’s what I found: I wrote something like “log into the portal, navigate to the purchase orders section, and fill in the order details from a CSV file.” The system actually generated a workflow that was maybe 70% there. The login part worked fine, but it had trouble with some of the dynamic dropdown menus and the form validation logic.

I ended up having to tweak the workflow manually to handle edge cases—like when the site returns an error message and you need to retry. It wasn’t terrible, but it wasn’t as automated as I’d hoped either.

The real question I have though: does this match everyone else’s experience, or am I just not writing my descriptions well enough? I’m wondering if there’s a sweet spot for the complexity level of tasks where the AI-generated workflows actually work without much modification.

[EmberCloud]

I run into this exact situation constantly. The plain English descriptions work best when you’re specific about the steps and the expected behavior at each point. Instead of “fill in the order details,” say “locate the order amount field, clear any existing text, and enter the value from column B.”

That said, dynamic dropdowns and form validation are exactly where you’d want to add a touchup. The workflow generation gets you 90% of the way there on simple scenarios, but real-world sites have quirks.

With Latenode, you get both the AI copilot for the heavy lifting and the ability to layer in custom code for those edge cases. No need to start from scratch when you hit a snag. Plus you can test different approaches in dev mode before pushing to production.

Check out https://latenode.com to see how the hybrid approach handles this kind of thing.