Turn incoming documents into reviewed data.
Document intelligence
Read supplier invoices and other business documents, then check extracted fields against the relevant ERP records. Your team can review uncertain or mismatched details.
Does this invoice agree with the purchase order?
Extract the supplier, reference, quantities and prices. Match those fields to the order and receipt, then show the differences for review.
What needs checking on the supplier document?
Extracted fields are compared with the ERP records before use.
Document intelligenceIllustrative records| Field | Document value | Check |
|---|---|---|
| Supplier | Northwind | Resolve to the supplier account |
| Invoice reference | 7734 | Check for duplicates |
| Quantity | 100 units | Compare with 96 received |
| Unit price | £25 | Compare with the purchase order |
The received quantity needs investigation.
The document supplies candidate values. Cobalt checks the fields and related records before preparing the exception.
- Fixed rules handle documents whose layout is already known.
- Text reading happens inside Cobalt ERP first. Sending pages to an outside AI provider is off unless you switch it on, has a page limit, and every send is logged.
- A quality check decides whether the text is clear enough to use, rather than confidently reading a blur.
- External AI only when needed and enabled — local text recognition uses your plan's document-page allowance without using external-model credits.
- Strict format checks — anything that doesn’t fit the expected format stops there.
- Matching stays exact — reading gives names and numbers. Finding the right supplier or product is a lookup in your records, limited by your permissions, never a guess.
- A person reviews it, then it becomes a draft — created the normal way, and not posted.
The safeguards around it.
- Duplicate checks before anything is created, so the same invoice can’t come in twice by two routes.
- No automatic posting, ever.
- Clear rules on keeping, deleting and sharing documents, because they contain other people’s details.
- It remembers your corrections — how you fixed something and why, and lets that fade over time rather than turning it into a fixed rule.
- Limits on processing, so a broken document can’t tie up the system.
- Protection against harmful documents — every document is treated with caution, and text inside it that looks like an instruction is treated as plain text.
Field by field, and document by document.
Accuracy is measured against a set of checked documents. Otherwise you are only guessing how well it works.
Each field
How often each field is read correctly and how often it is missed: supplier, reference, date, currency, amounts, line description, quantity, unit cost and stock code. It also counts made-up values, meaning fields filled with something that isn’t on the document at all.
Each document
How often the result is in a valid format, how often it can’t read the document, how often it is ready with no warnings, how often reviewers correct it, and how long it takes to create the draft.
What it reads is a suggestion until it is checked.
Nothing read from a document is treated as fact. Each value is a suggestion, showing how sure it is and which page it came from, until your rules confirm it or a person accepts it. That is how it saves time without being trusted blindly.
Wherever paper arrives.
- “Pull the totals off this supplier invoice” — into a draft bill, with a card showing what to check.
- “Which purchase order is this delivery note for?” — it reads the reference, then looks up the order.
- “Does this match what we ordered?” — the lines on the document against the order and what was received.
- “Is this a duplicate?” — checked before the bill is created, not after it is paid.
- “What did the signed proof of delivery say?” — the signature and photo, checked against the order.
Frequently asked questions.
Does a document ever post automatically?
No. Reading a document creates a draft the normal way, and a person confirms it. No setting changes that.
Are my documents sent to an outside AI provider?
Only if you enable an external service, with a page limit and a log of what is sent. By default, text is read inside Cobalt ERP using your plan's document-page allowance, without using external-model credits. An enabled external vision service may also consume your plan's external-model allowance.
How often does it make things up?
We measure it: the share of fields filled with a value that isn’t on the document. It is checked field by field against a set of checked documents, because a made-up value can cost you money, where a missed one only costs time.
Why doesn’t the AI decide which supplier it is?
Because there is one right answer, and it comes from looking up your records. The AI reads the name off the page. An exact lookup, limited by your permissions, decides which supplier that is.
Try it on your own records.
Book a demo and ask it a real question about your business. Every answer shows how it was worked out and which records it used.
Book a demo