Receive
Purchase orders arrive in one agreed inbox as PDF, email text or Excel.
SYSLUME Order Intake
Emailed purchase orders, turned into checked orders your team approves.
Order Intake reads purchase orders from PDFs, email text and Excel files, checks them against your product catalog and customer list, and puts each one in a review queue. Approved orders export as a CSV for QuickBooks Online. Nothing is written to your accounting system without approval.
The problem
Someone on your team opens each purchase order, finds the product codes, checks the prices and types the order into QuickBooks. It is slow, it piles up at month end, and one wrong code or quantity can mean a wrong shipment or invoice. The work needs care, but most of it is copying.
How it works
Five steps, one of them always human.
The agent prepares the order. A person decides what happens to it.
Extract
The agent turns each document into a structured order: customer, lines, quantities, prices, dates.
Check
Product codes can only come from your catalog. Customers are matched to your customer list. Anything that does not match is flagged as an exception.
Approve
Your staff approve, edit or reject each order in a review queue. Exceptions are shown first.
05 / Export. Approved orders are exported as a CSV file for QuickBooks Online import. Only approved orders are included.
What it does
Prepares, checks, flags
- Reads typed PDFs, email text and Excel purchase orders
- Uses only product codes from your catalog
- Flags unknown codes, price differences and missing fields
- Keeps a dated record of every approval, edit and rejection
- Runs with least-privilege access to accounts you own
Human approval
What it never does
Acts on its own
- Never sends emails or replies to customers by itself
- Never writes to your accounting system without approval
- Never trains models on your documents
- Never handles handwritten or badly scanned documents in version 1
How we measure it
Order Intake is built on the Claude API, using structured outputs and product codes limited to your catalog. Every change is checked against an evaluation suite before it is used.
- Field-level accuracy: every field (customer, product code, quantity, unit price, date) is compared with the correct answer on a fictional test set of purchase orders.
- Cost per document: the API cost of processing each document, measured on the same set.
- Exceptions: how often the agent flags a document for a person instead of guessing.
First results, 10 October 2026
Test set: 36 fictional purchase orders (12 email text, 12 PDF, 12 Excel; 136 order lines), 13 of them with one planted problem: an unknown product, a price that differs from the catalog, an unknown sender, a wrong stated total or a quantity written in words. Model: Claude Opus 5.5.
- Every planted problem went to review: 13 of 13 caught, and none of the 23 clean documents was flagged.
- 33 of 36 documents fully correct (customer, PO number, and every line’s product, quantity and price). The other 3 contained a quantity written in words (“two dozen”); the agent does not convert it, it sends the order to a person.
- Line accuracy: product codes 136 of 136; unit prices 136 of 136.
- Cost: about $0.011 per document on average ($0.020 at most); about 4 seconds per document.
Read these numbers with care. The documents are fictional and cleaner than a real inbox, so treat the results as an upper bound until measured on your own orders. Results also vary between runs: in an earlier run one 12 oz cup was read as the 8 oz product, and the price check caught it and sent the order to review. That is why nothing is exported without a person approving it.
Demo
A one-minute screen recording of the real product running on Claude: 8 fictional purchase orders arrive, the review queue flags what needs a person, a reviewer fixes a misread line, and approved orders export for QuickBooks Online.
Try 5 documents free
See the extraction on your own documents before you buy anything. $0, no build commitment.
- Pick 5 purchase orders of the kind you receive most often.
- Redact anything you do not want to share, such as bank details, signatures or personal phone numbers. Keep product codes, quantities and prices if you can.
- Use the start form and choose “Try 5 documents free”. Describe the documents; do not attach them there. We reply in writing with how to send them.
- Within 48 hours of receiving the documents, you get the extraction results back in writing.
- Your documents are deleted within 30 days. You can ask for earlier deletion at any time.
Typed PDFs, email text and Excel files only. Handwritten or badly scanned documents are out of scope for now. See the privacy notice for the document trial.
Start the free trialQuestions
What happens to my documents?
They are used only to produce your results. They are processed through the Claude API, with Anthropic as our sub-processor. We do not use them to train models. Trial documents are deleted within 30 days, or sooner if you ask.
Which systems does it work with?
Version 1 works on purchase orders forwarded to one agreed address; we connect that inbox during the build. It exports a CSV for QuickBooks Online import. Your product catalog and customer list can come from a QuickBooks export or a spreadsheet. Other systems can be discussed in an Automation Fit Check.
What if the extraction is wrong?
Anything the agent is unsure about, or that does not match your catalog or customer list, goes to the exception queue for a person to fix. Nothing is exported until someone approves it, so a wrong extraction is caught in review rather than in your accounts.
Who owns the accounts and data?
You do. The agent runs with the least access it needs, on accounts you own, and you can remove that access at any time.
Do I need to get on a call?
No. Everything is written and async: scope, questions, results and support.