You hand Claude a spreadsheet with 300 rows and ask it to submit each one through a partner's web form. The first two go through while you watch.
Then the login times out. A cookie banner lands on top of the submit button. Somewhere around row 40, a phone number goes into the company field. Nothing throws an error, so the bad submissions often go unnoticed until later.
Claude is good at deciding what belongs in each field. Filling web forms at scale depends on the browser.
- Claude handles the judgment inside a form well. Filling forms at scale depends on the browser runtime.
- An "LLM-per-step" loop re-reasons every submission, so cost grows with every run.
- Logins, CAPTCHAs, and one-session-at-a-time local browsing cap your throughput.
- Compile the submission once, then keep the model for the judgment calls.
Most teams reach for Claude here for good reason. A form submission starts as a judgment task: which record maps to which field, what to write in the free-text box. Claude is good at that.
The work then becomes repetition rather than judgment. Planning is a judgment task, so do it with a reasoning model. But execution is a repetitive task – so it makes sense for it to be compile it.
Why filling web forms at scale breaks in Claude-only workflows
The model re-reasons every submission
A fully agentic run treats every execution as if it were the first one. "You are paying a language model to relearn the task, forever." That cost is small on the first row and large by row 300.
An LLM-in-the-loop agent's cost is linear in steps and runs, while a code-first agent's cost is dominated by infrastructure. On one multi-step task, how Airtop works puts a compiled agent at 1m 21s and $0.063. Claude Code on Opus 4.7 took 7m 58s and $6.26.
Sessions and logins do not persist
An LLM-per-step setup runs locally, on your computer. No cloud browsing, no password vault, one session at a time.
The login you established by hand is gone by the next batch, and every CAPTCHA becomes a manual interruption. With Agent Builder, you sign in once and stay signed in. Local browsing serializes the work, so 300 submissions become 300 turns you sit through. Compiled agents run up to 100 simultaneous sessions instead, and big jobs finish in minutes instead of hours. Throughput depends on how many sessions the runtime can run at once.
Can’t we just use Computer Use?
Computer Use is a reasonable option for a first pass. It adapts to layouts it has never seen, self-corrects when a click misses, and needs no maintenance when a site ships a redesign. For the first form, that flexibility is worth paying for.
The cost shows up when the same form is submitted many times. Every action requires a call to a slow model, so API calls add up fast for a task that runs on a schedule.
The Agency Spectrum: Why AI Agents Don't Need to Think Hard Every Single Time gives the range. Even state-of-the-art computer use models "experience randomness, producing unexpected or incorrect behavior between 20% to 60% of the time (depending on complexity)." Run a task 1,000 times a day and a small probability of error compounds into a large number of failed attempts.
Agency is most useful the first time you do something, and least useful the hundredth time. Computer Use is better suited to a first pass than to a scheduled batch of a thousand submissions.
What works for filling web forms at scale
Reason once, at build time. Compile the result. Then run the compiled artifact, calling the model only where the work changes from run to run.
You describe the job in plain English, the way Agent Builder's own example prompt does: "Fill out these forms for me." Agent Builder builds and tests the automation, then compiles it into reliable, reusable code. That compiled agent runs like software rather than an LLM guessing every step. The Act step handles the mechanical half: "It clicks through real flows, fills forms, and extracts exactly the data you asked for."
That uses authenticated cloud browsing: browsers that run in the cloud, not on your computer. Built-in proxies, CAPTCHA solving, password vault, and up to 100 sessions come with it.
Form filling and submissions is a named workflow, and The Form Submitter is a ready template. It parses plain-text field values, dismisses cookie banners and popups before retrying, submits, and confirms from the resulting page. The old way was precise selectors, error-prone workflows, and constant maintenance, which Say Goodbye to Manual Form Filling with Airtop's New AI-Powered API still describes.
Reliability on a long batch comes from session persistence, self-healing, and run traces. Broken runs heal themselves, monitoring is compiled alongside the agent, and every run leaves full traces and video.
Keep Claude as the reasoning layer. Bring your own agent and add web automation to the agents you already run in Claude, n8n, Make, or Codex. Airtop + Claude Code installs with a single command. From there, execution stays reliable and cost-effective while your coding agent handles the reasoning and orchestration.
Form fill use cases that deterministic agents can solve
Enter weekly hours in the PSA
Someone needs their week's time entered into a PSA or ERP portal from a source of truth, on a set cadence, without doing the data entry by hand. A chat session can't do this because it ends when the conversation ends — there's no standing process to wake up every Friday, log in, and submit; a person still has to remember to ask. A compiled agent can, because it runs on a schedule, holds portal credentials and session state between runs, and executes the same login-fill-submit sequence unattended, every time the cadence calls for it.
File e-invoices and taxes on the deadline
Government e-invoicing and tax portals need someone to log in and submit or download filings by a fixed deadline. That's a bad fit for a chat, which has no way to persist across days or weeks to hit a filing date — and a missed or late filing has real financial and legal consequences that a one-off conversation can't be accountable for. A compiled agent works instead because it's triggered by the filing calendar itself rather than a person remembering to ask, and it captures confirmation of submission as proof it happened on time.
Handle license renewals and disclosures on the regulator's calendar
Periodic disclosures, license renewals, and reporting forms need to be filled, submitted, and confirmed on the regulator's schedule. Claude can't own this because compliance calendars run for months or years, well past the life of any single conversation, so nothing ties the deadline to an action. A compiled agent can, since it persists as a background job tied to the regulator's calendar and logs the confirmation automatically, turning a compliance risk into a routine background task.
Re-run quote forms with current parameters
Pricing-tool forms need to be re-run with current parameters to refresh quotes and rates, often daily. Claude can fill the form once when asked, but it has no way to independently re-run it every day to keep client-facing numbers current. A compiled agent handles it because it repeats the exact same parameter-entry and submission sequence on a timer, so quotes stay fresh without anyone re-triggering the work.
Keep job posts from expiring across boards
The same requisitions need to be re-filled and re-posted across multiple job boards so listings don't get buried or expire. This breaks down in a chat because there's no persistent memory of which listings are live where, or any mechanism to check back in and re-post before expiration. A compiled agent solves it by tracking state across runs and re-executing the posting flow on a schedule, keeping listings visible without manual re-entry.
Batch onboarding and offboarding into HR systems
Each day's onboarding, offboarding, and work-authorization changes need to land in HR and government systems as a batch. A chat can't wait for those changes to accumulate and then act on them — someone has to gather the data and ask, every time. A compiled agent can, running as a nightly job that pulls the day's changes and processes them into each system automatically.
Push new leads into CRM and partner portals
New leads need to move from a form or sheet into a CRM, dealer network, or partner portal as they arrive. Claude only acts when prompted, so leads sit in a queue until someone notices and asks it to process them — the opposite of what's needed. A compiled agent avoids that by polling the source on an hourly or daily cadence and filling the destination forms itself, so new rows never wait on a human trigger.
Sync prices and stock to marketplace portals
Prices, stock status, and availability need to be pushed into every marketplace and supplier portal each morning. There's no standing job in a chat that runs each morning on its own — it would need to be re-invoked and re-supplied with current values every single day. A compiled agent runs on a daily trigger instead, pulling current values from the source and pushing them into each portal's forms without anyone kicking it off.
Reconcile bookings against the source of truth
Bookings need to be reconciled against the source and reservations created, confirmed, or updated in a scheduling system every day. Claude can't take that initiative on its own — it needs a person to notice something changed and ask it to fix it. A compiled agent checks the source on a schedule and fills and submits the necessary changes itself, keeping the calendar in sync without oversight.
Test signup and checkout forms on a loop
Critical forms need to be filled out end-to-end on a loop, with an alert the moment a field breaks or a submission fails. A chat has no way to run that test continuously and no channel to proactively alert anyone — it only responds when asked, by which point a broken checkout may have already cost customers. A compiled agent runs the test continuously on a schedule and can push an alert the instant something fails, catching the break before customers do.
Try it on your next batch of forms
Describe the submission once and compile it. The agent signs in and stays signed in, runs the batch across parallel sessions, and leaves Claude the judgment calls that need a model. Try it for free; you can spin up your first agent in five minutes. Start for free.
FAQs about filling web forms at scale and Claude
Can Claude fill out web forms?
Yes, and for one or two forms it does the job well. Claude reads the record, decides which value belongs in which field, and writes the free-text answer.
An LLM-per-step setup browses locally, one session at a time, with no password vault, so a login you set up by hand won't survive to the next batch. The web automation layer changes that: sign in once, stay signed in.
Why does form filling break between 10 submissions and 1,000?
Because the loop re-solves the same problem on every run. Nothing carries over from submission 10 to submission 1,000, so your bill tracks steps and runs instead of dropping. At 10 submissions the per-run cost is easy to miss. At 1,000 submissions the per-run cost is obvious.
Error rates add up the same way. The Agency Spectrum: Why AI Agents Don't Need to Think Hard Every Single Time puts unexpected or incorrect behavior between 20 percent and 60 percent of the time, depending on complexity. At 1,000 runs a day, even a small per-run error rate becomes a queue of bad submissions to chase.
A compiled agent is deterministic at runtime: same input, same output, with no prompt being re-evaluated on every run.
Is Computer Use good enough for repeated form submissions?
It depends on how many times you plan to run the thing. For a handful of submissions, yes: the adaptability that handles an unfamiliar layout or a missed click earns its cost while you're still working out the flow.
Each repeated action spends another call on a slow model, and the variance that makes Computer Use adaptable also means each run can fail independently. For a submission that goes out on a schedule, compiled execution is more predictable.
The work runs in two phases. Spend the capable model at build time, working out the flow. Then run a compiled agent at execution, keeping model calls for the steps that change from run to run, like reading a confirmation page and judging whether the submission landed.
Do I have to stop using Claude to fill web forms at scale?
No. Keep Claude as the reasoning layer and give it a browser. Bring your own agent is the default, and web automation attaches to the agents you already run in Claude, n8n, Make, or Codex.
Airtop + Claude Code is a one-command install. Claude keeps the reasoning, orchestration, and data processing, and the browser steps run in the cloud.
You can start from a template instead of a blank file. The Form Submitter template already dismisses the cookie banner sitting on your submit button before it retries, and it confirms the result from the page that loads after submitting.
What happens when the form changes?
Minor drift gets absorbed inside the compiled agent. Broken runs heal themselves, and the model is used to relocate an element that moved or got renamed. A moved or renamed element is the kind of change that breaks a hand-maintained selector.
A hard failure is caught rather than left unreported. Monitoring is compiled alongside the agent, so repeated failures notify a human rather than piling up bad submissions that look successful. Agent Builder leaves full traces and video of every run, showing you where the flow stopped.
A full redesign still requires a rebuild. Rebuilding from a plain-English description with failure context still beats the selector-by-selector rework in Say Goodbye to Manual Form Filling with Airtop's New AI-Powered API.
