Data extraction from any document, into clean rows
PDFs, invoices, emails, web pages — agents extract the fields you define and write them where they belong, flagging anything they are not sure about.
Opens in AGNT and takes about a minute.
Didn’t open? AGNT may not be installed here yet. Download AGNT free · Download the .json
What this actually involves
| Tools it needs | Google Drive, Dropbox, Google Sheets, Airtable, PostgreSQL |
|---|---|
| Setup | 4 steps, about ten minutes |
| Approval | Yours, per category — nothing is sent on your behalf unless you say so |
| Where it runs | Your own machine. Credentials stay in a local vault. |
What the agent actually does
Clean rows, and a flag on anything uncertain
Scanned PDFs, email bodies, spreadsheets that were never designed to be parsed, web pages, attachments in formats a vendor invented years ago. The agent reads for meaning, so a new supplier’s layout does not require a new parser.
Your schema
You define the fields and their types; the agent fills them. Where a value needs deriving rather than copying — a total that must be computed, a date in an ambiguous format — it does the work and records how it arrived at the answer.
Uncertainty is flagged, not hidden
Anything the agent is unsure about is marked for review rather than guessed. A field flagged low-confidence costs somebody five seconds; a field silently wrong costs a great deal more, much later.
The tools this job needs
Google Drive
Where the work arrives. Where the documents live — and pile up. The agent watches it and reads what turns up in full, rather than matching a rule against a subject line.
Dropbox
Context the agent pulls in before deciding. Shared files that arrive needing processing. It is read, not just referenced — which is what lets the decision account for it.
Google Sheets
Context the agent pulls in before deciding. The database every team already has open. It is read, not just referenced — which is what lets the decision account for it.
Airtable
Where the result lands. Structured records with an API worth automating. Nothing is written here until the agent has formed a view and, where you asked for it, you have approved it.
Set it up
- Connect the source — Gmail, Drive, Dropbox, or a folder the documents land in.
- Define your fields in plain language, including what makes a value suspicious.
- Choose the destination: Google Sheets, Airtable, PostgreSQL, or your own system by API.
- Run a batch of real documents, check the flagged rows, and tighten the field descriptions.
You can see exactly what it did
The brief it works from
This is the actual instruction set the template installs — what the agent is told to do, and what it is told never to do. Every line of it is yours to edit in AGNT after install.
Read the full brief
You turn unstructured documents into clean structured rows.
Read for meaning, not position
You locate an invoice total because you understand what an invoice total is, not because you know where the last one was printed. That is the whole reason you exist instead of a template parser: a new supplier, a redesigned form or a page scanned slightly rotated should just work.
Fill the user's schema
Extract the fields they defined, in the types they defined. Where a value must be derived rather than copied — a total that needs computing, a date in an ambiguous format, a name split across two lines — do the work and record how you arrived at it.
Dates deserve special care. 03/04/2026 is two different dates depending on where the document came from. If you cannot establish the convention from the document itself, flag it rather than guessing; a silently wrong date propagates further than almost any other error.
Confidence is a required output, not a courtesy
Mark every field you are unsure about. A flagged field costs somebody five seconds. A field that is silently wrong costs a great deal more and is discovered much later, usually by someone who then stops trusting the whole dataset. Never invent a value to complete a row. An empty flagged field is a correct answer.
Validate where you can
If line items should sum to a stated total, check. If a reference should match a known format, check. A failed cross-check is an exception for a human, not a reason to discard the document.
Report per run: how many documents, how many clean, how many flagged and why. If a particular source is consistently producing flags, say so — that is a field description that needs improving, and it is fixable.
Shown exactly as it ships. The agent inherits whichever model you already use.
Why template parsers keep breaking
Traditional extraction encodes positions and patterns: the total is in the box at these coordinates, the reference matches this expression. It works beautifully for the documents it was built against and fails on the next supplier, the next redesign, the next scanned copy that arrived slightly rotated. Maintenance grows with the number of sources, which is exactly backwards from what you want.
Reading for meaning changes the maintenance curve
An agent locates the invoice total because it understands what an invoice total is, not because it knows where the last one was printed. New formats generally work on the first attempt, and the ones that do not are handled by improving a description rather than writing another parser. The cost of adding the fiftieth source is roughly the cost of adding the second.
Confidence is the feature that makes it usable
The reason extraction pipelines get abandoned is rarely accuracy in aggregate — it is that nobody can tell which rows to trust. An agent that reports per-field confidence turns the output into something a person can work with: accept the clean rows, spend attention only on the flagged ones, and watch the flagged proportion fall as the descriptions improve.
Common questions
Does it work on scanned documents?
Yes. Scans and photographs are handled through vision-capable models, including the awkward cases of skewed pages and mixed print and handwriting.
What happens to a field it cannot find?
It is returned empty and flagged, never invented. An honest gap is recoverable; a plausible fabrication in a financial system is not.
Can it validate as well as extract?
Yes — cross-checks like "line items must sum to the stated total" are exactly the kind of rule worth adding, and failures become exceptions for a human rather than silent bad data.