Can an AI fill in a form for you?

How should an agent read a form the way a person does?

A form is not a bag of empty boxes. Each control has a visible label, a name in the markup, a placeholder, and sometimes helper text underneath. A careful fill starts with the label next to the box, not with guessing from the field name alone. If the label says “Preferred name” and the name attribute says “legal_name”, the label wins unless the page itself says otherwise.

Dropdowns need the option text, not the hidden value, unless the page shows codes. Checkboxes and radios are on or off, never half-filled. Dates belong in the format the control already uses: day-month-year in Britain unless the field says ISO. If a field is required and the source data is missing, the honest move is to leave it blank and say so, not invent a value.

What should be checked after each field is filled?

Filling is not the same as being done. After a value is placed, the agent should read the field back as it now appears on the page. Autocomplete, masks, and country pickers often rewrite what was typed. A postcode with a space, a phone number with a leading zero, or a date that snapped to the wrong century should be caught here, not after you have already sent the form.

If the page shows an inline error, that message is part of the form. The next action is to fix that field or stop, not to press on. A second pass over required fields, then optional ones, is slower than spraying every box at once and far less likely to leave a wrong value sitting next to a correct label.

Why stop before the submit button?

Submit is the point of no return for most public forms. An agent that can type can also send. The useful default is to fill, re-read, and then wait. You look at the page the way you would if a colleague had filled it: names spelt as you intended, address lines in the right order, marketing boxes left as you prefer.

A simpler tool is often better here. Your browser’s own saved addresses and a password manager already fill many fields without an agent in the middle. Use those for logins and payments. Use an agent when the form is long, unfamiliar, or drawn from notes you already have in chat — and still keep the last click. GROX is a conversational surface that can draft and act in other tools; it is not a reason to skip that last look.

Label-first fill

Match each box to the visible label and helper text, not only to the field’s internal name.

Read-back

After typing, read the value as the page now shows it, including masks and rewritten dates.

Stop before submit

Leave the send button for you so a wrong line is still a draft, not a filing.

Leave sensitive fields

Passwords, one-time codes, and card details are typed by you, never pasted from a chat.

When should a form be left entirely to you?

Some fields are not “data”. A password, a one-time code, a card number, a CVV, and a banking sort code should not pass through an agent’s memory of the session if you can avoid it. Type those yourself. The same applies to medical identifiers, national insurance numbers, and anything the page marks as confidential.

If the form is a legal filing, a tax return, or a contract, the agent can prepare a draft in a document, but the official form is still yours. A password manager, the browser’s autofill, or simply typing is the better choice whenever the risk of a wrong send outweighs the time saved. GROX can keep a thread across research and drafts; it does not replace your judgement on what must never be submitted by mistake.