Yes, for routine web forms, with you watching. Browser agents such as ChatGPT Work and Claude in Chrome can type, click and submit, and they pause for logins and payments. They're fine for repetitive, low-stakes forms. For declarations, payments or portals whose terms forbid bots, fill them yourself. For daily repeats, an integration or recorded workflow beats an agent.
So the useful question isn't "can AI do it?" but which of four approaches fits each form: a browser agent, a recorded workflow, an integration that skips the form entirely, or AI that prepares the data for a person to paste. Three things settle it: how often the form comes round, what a wrong entry would cost you, and whether the portal's terms allow automated access at all.
Four ways to get a form filled without retyping it
| Approach | How it works | Best for | Watch out for |
|---|---|---|---|
| Browser agent | An AI model looks at the page, then clicks and types like a person, from instructions you give in plain English | Forms that vary a little each time, done a few times a week | Slow, needs supervision, can pick the wrong option |
| Recorded workflow | Software replays the same clicks and keystrokes you recorded, using data from a spreadsheet | The same form, same layout, many times a day | Breaks when the portal changes its layout |
| Integration or file upload | Your system sends the data directly, or you upload one file instead of typing each line | High volume or high stakes, such as orders | Needs the supplier to offer it |
| AI prepares, a person submits | AI pulls the values from emails, PDFs or notes into a table; someone pastes and checks | Forms with legal declarations or sensitive data | Saves less time, but keeps a person accountable |
The two browser agents most small businesses will meet work differently.
- ChatGPT Work replaced the older ChatGPT agent (Agent Mode), which OpenAI retired on 9 July 2026. It works in a cloud browser, a separate computer on OpenAI's side, and comes with Plus, Pro, Business and Enterprise at no extra cost, though not Free or Go. At a login page it pauses and gives you a secure sign-in form, and it's designed to ask for your confirmation in the chat before actions that are hard to reverse, such as confirming a booking or making a payment. If you find older guides describing ChatGPT agent's "takeover mode", that product no longer exists.
- Claude in Chrome became generally available on 26 August 2026 for paid Claude plans. It works inside your own Chrome browser, on sites where you're already signed in. It asks permission per site, and its "Manually approve" mode pauses before every action, which is the setting to use while you're testing. Anthropic's help centre says it won't make purchases, create accounts or handle card details in any mode.
For the recorded approach, Windows 11 comes with Power Automate for desktop preinstalled. It has a recorder that captures mouse and keyboard steps and turns them into a repeatable flow, which Microsoft describes as suitable for populating forms on web pages and desktop applications. It isn't AI in the chat sense, which is exactly why it's dependable for a form that never changes. If you're weighing it against other automation tools, see when Power Automate beats Zapier.
Where browser agents stall
Agents are improving quickly, but the same problems come up on real supplier and insurer portals:
- Logins, two-step verification and CAPTCHAs. The agent stops and waits for you. On a portal that logs you out after ten idle minutes, that can mean starting again.
- Dropdowns with near-identical options. "250mg tablets" and "250mg/5ml suspension" look alike to a model working fast. This is the most dangerous failure, because the form submits happily with the wrong choice.
- Date and number formats. Day/month versus month/day, commas in quantities, or a field that silently drops anything after a decimal point.
- Multi-page wizards. Validation errors on page four that send you back to page two.
- File uploads. Attaching an invoice or X-ray from your computer is hit and miss, depending on the tool.
- Hidden instructions on the page. A web page can contain text meant to redirect an agent. The explainer on prompt injection covers why this matters when an agent can click Submit.
- Speed. An agent reads the page, reasons, acts, and checks, over and over. For a short form you know well, you'll often be quicker yourself. Agents pay off on tedious forms, not quick ones.
Terms, declarations and who is responsible
Three things stay yours, however capable the tool.
The portal's terms. Many supplier and insurer portals have terms of use that restrict automated access or sharing credentials. Search the terms for "automated", "robot", "scraping" and "credentials" before you point an agent at them. An agent logged in as you is, as far as the supplier is concerned, you.
The wording to look for is usually short and easy to miss. An illustrative clause of the kind many trade portals carry: "You must not use any robot, spider, scraper or other automated means to access the Portal, and you must not share your login details with any third party." Both halves matter: an agent is an automated means, and a cloud-browser agent arguably involves a third party. When you find a clause like that, don't argue with it. Email the supplier's account manager and ask whether they accept orders or claims by file upload, email or a direct link from your system. Suppliers who ban bots often have a bulk-upload page they simply don't advertise.
The declaration. Forms that end with "I confirm this information is accurate" are a statement by you. If the agent chose the wrong option, you still made the declaration. Keep declaration boxes for a person to tick after reading the filled form. If a mistake on a particular form could have legal or insurance consequences, ask your adviser how you should handle it.
The data. An agent sees whatever is on the page, including customer and patient details, and the provider processes it. Use a business plan with terms you've checked, and don't let staff run agents on personal accounts. The tutorial on whether staff should use AI browser agents goes further on the rules.
How it plays out in three businesses
These are illustrations of the decision, not case studies.
An independent pharmacy places orders on three wholesaler portals every day. That's high frequency and high stakes: a wrong strength or pack size matters. A browser agent is the wrong tool here. The first step is to ask each wholesaler whether it accepts orders directly from the pharmacy's dispensing system or by file upload. Where it does, that beats any agent. Where it doesn't, a recorded workflow fed from a checked spreadsheet is more predictable than an agent choosing from dropdowns.
A veterinary practice submits about 25 pet insurance claims a week across several insurers' portals. Each needs a clinical history summary and the invoice. The time goes on writing the history, not the clicking. So AI prepares: it drafts the history from the clinical notes in the format each insurer asks for, a vet or nurse checks it, and a person enters it and ticks the declaration. If writing each history takes 15 minutes by hand and checking an AI draft takes 5, that's roughly four hours a week back, against perhaps half an hour lost if an agent did the clicking badly. The approach in turning PDFs and scans into usable data handles the invoice side.
The drafting step needs the same fencing as an agent run. A request of the kind the practice might use, on a business plan and with the owner's details removed:
From the clinical notes below, write the clinical history for an
insurance claim: presenting complaint, date first seen, tests,
diagnosis, treatment. Use only what's in the notes. Don't describe
any condition as new, ongoing or pre-existing; the vet decides that.
Notes: [PASTE NOTES, owner's name and address removed]An illustrative extract from the draft:
"Presented on 12 August with lameness in the left hind limb. Radiographs showed no fracture. Diagnosis: soft-tissue injury. Treated with rest and anti-inflammatories. Note: a similar lameness was recorded in 2024, suggesting a recurring problem."
The last sentence is the one the vet deletes. The 2024 entry in the notes was for the right foreleg, a different problem entirely, and "suggesting a recurring problem" is the AI drawing a conclusion the prompt told it not to draw, in the one part of a claim that decides whether the insurer pays. The rest was accurate and took two minutes to check.
A hearing-aid practice logs around ten repair and warranty returns a month on manufacturers' portals. Low stakes, the same few fields each time, and a real nuisance. That's a good first agent job: supervised, with the serial numbers pasted in rather than read from a photo, and the practice manager clicking Submit.
What each approach costs
The subscription is rarely the expensive part. Supervision and upkeep are.
| Approach | Money (list prices in USD) | Your time |
|---|---|---|
| Browser agent | Included in paid chat plans, such as ChatGPT Plus or Claude Pro at $20 a month; for customer data, a business plan such as ChatGPT Business or Claude Team at $25 a seat a month billed monthly, with a two-seat minimum | Watching every run at first, then checking each form before submitting |
| Recorded workflow | Power Automate for desktop comes with Windows 11; check Microsoft's licensing if you want to share flows or run them unattended | One to three hours to record and test per form, plus a fix each time the portal changes |
| Integration or file upload | Often free from the supplier; otherwise automation fees, such as Zapier from $19.99 a month billed annually or Make from about $9 a month | Set-up once, then occasional checks |
| AI prepares, a person submits | Any business AI plan you already have | The submitting and checking stay with a person |
Count the supervision honestly. An agent that saves eight minutes on a form but needs six minutes of watching and checking is saving two, and those two minutes may not be worth the risk of a wrong dropdown slipping through on a busy day.
A prompt for a supervised agent run
The difference between a useful agent run and a mess is how tightly you fence it in. This template works in ChatGPT Work and Claude in Chrome:
Task: fill in the repair request form on [portal name] for the item below.
Start at: [exact web address of the form page]
Rules:
- Use only the values in the table. Do not guess or infer any field.
- If a field isn't in the table, stop and ask me.
- If a dropdown has no exact match for my value, stop and show me
the available options. Do not pick the closest one.
- Do not tick any declaration, consent or terms box.
- Do not click Submit or any final confirmation button.
- When every field is filled, stop and list each field with the
value you entered, so I can compare them before submitting.
| Field | Value |
| Model | [model] |
| Serial number | [serial] |
| Fault | [short description]|
| Purchase date | [DD/MM/YYYY] |
| Return address | [practice address] |The "stop and show me the options" line is the one that prevents the near-identical dropdown mistake. The "do not click Submit" line keeps a person responsible for the final step.
The field list at the end is where you catch what the rules didn't. At the hearing-aid practice, an illustrative run on a manufacturer's warranty portal ended like this:
"All fields are filled. Model: [model as given]. Serial number: [serial as given]. Fault: intermittent sound in the left aid, no fault with the right. Purchase date: 03/04/2026. Return address: entered as provided. I have not ticked the declaration or clicked Submit."
Every value matches the table, and one is still wrong. The practice wrote the date day-first, meaning 3 April. The portal's date field expected month-first, so it read 4 March, which on this model would put the aid a month closer to the end of its warranty. The agent copied the value faithfully, exactly as told. It showed up only because the practice manager compared the portal's own summary page, which displayed "4 March 2026", with the table before clicking Submit. Two changes fixed it for good: dates in the table are now written as words ("3 April 2026"), and the prompt says "enter dates in the format the field's placeholder shows". Check the portal's summary page, not just the agent's list.
The ten-form test before you trust it
- Pick one form you fill in often. Time yourself doing three by hand.
- Run ten real ones with your chosen approach, watching each.
- For each, record: minutes from start to ready-to-submit, how many fields were wrong, and how many times it needed you.
- Keep it only if there were no wrong values you'd have missed without checking, and the average time, including your checking, beat doing it yourself.
- If the agent passes, keep reviewing every submission for the first month, then spot-check one in five.
Here are the hearing-aid practice's results, filled in (illustrative). By hand, three returns took 8, 10 and 9 minutes, so about 9 each. With the agent, ten runs averaged 4 minutes to ready-to-submit plus 2 minutes of checking against the summary page, so 6 minutes each. One field was wrong (the date above), and it was caught. The agent needed her four times: twice for logins, and twice when a dropdown had no exact match, which is the rule working as intended. Three minutes saved on each of about ten returns a month is 30 minutes. That's a pass, but a small one. The practice kept it because the returns no longer pile up waiting for the one person who knows every manufacturer's portal. If the time saving is all you're after, a result like this says the form was barely worth automating.
If it fails on accuracy, move one row down the table: from agent to recorded workflow, or from recorded workflow to AI-prepared data. If it fails on speed, the form probably wasn't worth automating. The broader guide to stopping retyping data between apps covers the integration route when a supplier offers one.
Matching each form to the right approach
| The form | Use |
|---|---|
| A few times a month, low stakes | Whatever is quickest: often just doing it, sometimes a supervised agent |
| Several times a week, low stakes, fields vary | Browser agent with the fenced prompt, after the ten-form test |
| Many times a day, same layout | Recorded workflow, or an integration if the supplier offers one |
| High stakes (orders, clinical or money details) | Integration or file upload; if none exists, AI prepares and a person submits |
| Has a legal declaration, a payment, or terms that forbid bots | A person fills or at least submits it |
More on AI and form filling
Can AI fill in PDF forms?
ChatGPT, Claude and Gemini can read a PDF form and give you the value for each field from your notes or documents. Whether a tool can write those values back into a fillable PDF varies by tool and changes often, so open the file it returns and check every field before you sign or send it. For scanned paper forms, extracting the values into a table is usually the reliable part.
Will a supplier portal block an AI agent?
Some will. Portals that use CAPTCHAs or bot detection may stop an agent or force you to take over. Separately, a portal's terms of use may forbid automated access even if nothing technically blocks it. Read the terms first, and if they say no, ask the supplier whether they offer a file upload or a direct link from your system instead.
Does ChatGPT see my password when it logs in to a portal?
When ChatGPT Work's cloud browser reaches a login page, it pauses the task and shows you a secure sign-in form. OpenAI says the username, password and any two-factor code you enter there go straight to the remote browser, aren't visible to the model and aren't stored by ChatGPT. Even so, use a separate portal login for this where the supplier allows it, and never let an agent handle card details.
Further reads
- What Is an AI Agent? A Plain-English Guide for Business Owners — A plain explanation of what an agent actually is.
- What Can Go Wrong When AI Agents Take Actions for You? — The wider risks of letting software act for you.
- How to Turn Handwritten Forms Into Spreadsheet Data With AI — When the form starts life on paper.
- How to Automate Purchase Orders and Supplier Emails With AI — Take supplier ordering out of portals altogether.
- Does AI Stock Ordering Pay Off for an Independent Pharmacy? — The pharmacy ordering question in more depth.
- How to Pilot Your First AI Agent Without Risking Customers — Run a first agent trial without risking customers.
- AI Submission Intake: Stop Re-Keying Data Into Insurer Portals — Key once, check once: extract client documents into a master record, validate it with simple rules, then feed insurers by API, rater or copy-ready blocks.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help centre pages on ChatGPT Work's cloud browser (secure sign-in and confirmations); Anthropic's Claude in Chrome general availability announcement (August 2026) and Claude in Chrome permissions guide; Microsoft Learn on Power Automate in Windows 11.