Skip to content

A second opinion for tricky form fields (plain)

Engineer version: ../technical/pageagent-advisor-design.md

The idea in one sentence

WorkWingman keeps using its existing, reliable automation to fill out job applications — but for the rare moment a field is genuinely confusing (like a dropdown that doesn't have an exact match for your degree), it can now ask an AI assistant for a second opinion before falling back to asking you.

Why not just let the AI fill out the whole form?

That was tried first, on a private test copy of a real application form. It worked, but two things became clear:

  • The AI's way of "reading" a page skips some of the exact technical markers WorkWingman's existing automation already uses to find the right button or box reliably. The AI just wasn't as precise as the purpose-built automation already in place.
  • The one place the AI genuinely helped was different: when a dropdown had options that didn't exactly match what you'd typed (say, your degree is "B.S. Game Development" but the dropdown only offers "Bachelor of Science" and "Game Design"). Figuring out which option is the closest real match is a language problem — exactly what an AI is good at. Clicking the right button on the page is a mechanical problem — exactly what the existing automation is already good at and shouldn't be replaced.

So the design settled on: keep the reliable automation as the driver, and bring in the AI only as an advisor for that one specific kind of moment.

What actually changes for you

Nothing about how you use the app changes. WorkWingman already pauses and asks you to pick when it isn't confident about a field. Now, before it pauses, it can quietly check with an AI assistant first:

  • If the AI is confident enough about the closest match, WorkWingman fills it in and keeps going — the AI never clicks anything itself, it only suggests; the same trusted automation does the actual filling.
  • If the AI isn't confident either (or isn't available), you still get asked, exactly like before. Nothing is lost — this is purely an extra safety net, not a replacement for you having the final say when it matters.

Guardrails

  • The AI advisor is never allowed to interact with the page directly — it can only read a plain description (a field's label, your intended answer, and the options actually shown) and return a suggestion with its reasoning. It has no ability to click, type, or submit anything.
  • You (or a confidence threshold you control) always still get the final say when the automation or the AI aren't sure.
  • The same design works no matter which company's application system you're using — Workday, Greenhouse, Lever, and others — because the advisor only ever sees the plain, human-readable version of the question, never the technical internals of that particular site.