AI Data Extraction
Invoice data extraction is powered by Brikly's AI engine. The system continuously improves as it processes more invoices from your suppliers.
When you upload an invoice, Brikly's AI reads the entire document and converts it into structured, editable data. This replaces the manual process of keying in supplier details, line items, and totals by hand.
What the AI extracts
The extraction engine identifies and pulls out the following fields:
Header information
- Supplier name - the company that issued the invoice.
- Invoice number - the supplier's reference number.
- Invoice date - the date printed on the invoice.
- Delivery date - if different from the invoice date.
- Invoice total - the grand total including VAT.
Line items
For each product listed on the invoice, the AI extracts:
- Description - the product name as printed by the supplier.
- Quantity - number of units ordered.
- Unit of measure - e.g. kg, litres, cases, each.
- Unit price - price per unit.
- Line total - quantity multiplied by unit price.
- VAT rate - the VAT percentage applied to that item.
- Item type - whether the line is an ingredient, a consumable, or equipment (reusable kit, which is set aside from matching). Delivery and other charges are left unclassified, so you can set them aside once and Brikly remembers.
The unit a supplier billed in
Some wholesale suppliers bill the same product per tub on one delivery and per case on the next, and print which one they used in a column of its own. That might be a unit-of-measure column reading "split / TUB", a Unit column sitting beside a Box column, or a Kg column sitting beside a Pieces count. Where the document prints it, Brikly reads it and records the billed unit alongside the line, then checks it against the printed line amount before using it. Without that column the two readings are impossible to tell apart from the line alone, which is how a per-tub price ends up costed against a case.
This matters because it is not a price change. If a supplier bills you per tub this week and per case next week, the cost per kilo has not moved. Brikly treats the billing unit as a fact about the paperwork, not as a price movement, so your margins stay steady and you are not asked to review an alert that was never real.
List prices and per-line discounts
Where a line carries its own discount column ("Disc %" or "% Disc"), the unit price printed beside it is the list price and the amount at the end of the row is already net of that discount. Brikly costs on the net figure and keeps the list price and the percentage as a record of how the line was priced. Per-line discounts of this kind are not added to the invoice's discount list, because they are already reflected in the line total.
Catch-weight and variable-weight lines
Deli, meat and cheese invoices usually print two weights on the same row: the size of one piece, which is normally in the product name, and the total weight actually delivered, which is roughly the quantity multiplied by the piece size. Only the first of those describes the product. Brikly keeps the piece size as the pack measurement and never lets the delivered total become it, so a case of ten 2.5kg salamis is not recorded as a single 23.5kg item.
The exception is a line priced per kilo. There the delivered weight is what the supplier charged for, so it stays as the quantity and the unit price stays as the rate per kilo.
Totals and VAT
- Subtotal (net amount before VAT).
- VAT breakdown by rate (e.g. 0%, 5%, 20%).
- Grand total (including VAT).
Confidence scoring
Every extracted field is assigned a confidence score that tells you how certain the AI is about its reading:
| Confidence | Indicator | What it means |
|---|---|---|
| High | Green | The AI is very confident. Typically no action needed. |
| Medium | Amber | The AI's best guess - worth a quick check. |
| Low | Red | The AI struggled with this field. Manual review recommended. |
Fields with high confidence are usually correct, but it is good practice to glance over the extracted data before confirming - especially for the first few invoices from a new supplier.
Reviewing extracted data
After extraction, you are presented with a review screen showing all the data the AI found. From here you can:
- Edit any field - click on a value to correct it.
- Confirm the invoice - accept the extracted data and move to the matching stage.
- Reject the invoice - discard the extraction and start again if the results are unusable.
Fields that need attention are highlighted with their confidence colour so you can focus your review time where it matters most.
How extraction improves
The instructions Brikly gives its extraction engine are versioned, so every improvement is a deliberate, dated change rather than a silent drift. When a new version ships, the first invoice you upload afterwards may take a few seconds longer than usual while the engine warms up. Nothing you have already extracted is altered.
Handling extraction errors
Occasionally the AI may misread a value or miss a line item entirely. Common causes include:
- Poor image quality - blurry or low-resolution photos.
- Unusual invoice layouts - heavily designed or non-standard formats.
- Handwritten annotations - the AI focuses on printed text and may skip handwritten notes.
If extraction results are poor:
- Check whether a higher-quality version of the invoice is available (e.g. the original PDF instead of a photo).
- Re-upload the better version.
- If the same layout consistently causes issues, contact support - the team can investigate and improve handling for that supplier's format.
If you correct an extracted value, make sure to update the related fields as well. For example, if you change a line item's unit price, verify that the line total and invoice total still add up.
Next step: reviewing line items
After extraction, some line items may be missing fields that the AI could not read (e.g. pack size or unit of measure). See Reviewing Line Items for how to quickly fill in missing data before matching.
How corrections improve the system
Every correction you make feeds back into Brikly's learning engine. Over time, the AI becomes more accurate for the suppliers and formats you work with regularly. See The Learning System for more on how this works.