ARTEMIS LAB / BLOG
Dharma: turn an invoice into the fields you need
Dharma-AI introduces a model for extracting document fields. How to check its answers before importing them into a spreadsheet.
An invoice arrives as an image, but your spreadsheet needs a supplier, a date and an amount. On November 4, Dharma-AI introduced Field Extractor, a model that turns document fields into structured data. The announcement describes JSON output, a format software can read. Dharma-AI’s introduction.
Reading a page or filling in boxes?
Imagine a blank form sitting beside the invoice. Copying every word onto another sheet does not fill in that form. Someone still has to decide which number belongs in which box. That is the difference between recovering text and recovering usable information.
Dharma-AI describes two modes: let the model identify fields, or give it the fields you want. The company says missing information should return null, meaning no value here. That advertised behavior still needs checking against your documents. How the supplier describes it.
To picture JSON, think of labels attached to values: “supplier” accompanies a name, and “total” accompanies an amount. A tidy answer is not necessarily a correct answer. A number can be readable and placed in the right box while still having been misread from the page.
Start with a small pile
Here is a possible trial method, rather than an experiment carried out by Artemis Lab. Choose a few invoices you can check easily. Write down the expected answers by hand, then compare each field with the tool’s suggestion.
Define every box precisely. “Date” could mean the issue date or the payment deadline. “Amount” could mean the total before or after tax. If your team cannot agree on the question, judging the AI’s answer will be difficult too.
Deliberately include a document with missing information. You can then check whether the box stays empty or a value appears anyway. For this trial, a clearly marked absence is easier to handle than a plausible answer quietly slipping into the spreadsheet.
Include less legible invoices and different page layouts as well. The point of this exercise would be to discover where a review becomes necessary. Keep a record of errors beside the original pages so you can understand each mismatch instead of simply counting failures.
Measure the work left over
Dharma-AI publishes its own quality and cost comparisons. Those are supplier results, rather than independent validation presented here. They do not establish what your organization would save. Comparisons and calculation conditions.
For your trial, count corrected fields and the time spent checking them. Separate a mistake in a supplier’s name from a mistake in an amount: their consequences differ. Keep a human approval step before payments or important imports into another system.
Preparing the spreadsheet would be a useful first step to automate. The criterion for going further is knowing which values can be accepted, which need review, and how to find the original document when someone needs to settle a doubt.