What AI invoice processing reads, and what it hands to a person
A model reads the invoice. We check the numbers and build it in a business day, free until it runs.
AI invoice processing uses a model to read an invoice and return its fields: supplier, invoice number, dates, totals, tax and every line. It takes over the keying, and the checking stays. What a system does with the invoices it cannot settle decides how much checking you still owe, because a wrong value arrives looking like a right one.
Fabrx builds the API. Landing it in a given system is done by partners, so the only requirement is a system that accepts an API call.
What moves, and what it becomes
A photo taken on a phone
Skewed, lit from one side, with a thumb across the corner. The totals are usually legible. The line table is where the angle starts costing you rows.
A payable, with the lines kept
Purchases and payables
A scan with no text layer
An image of a page, with no text layer behind it. Every character comes from reading the pixels, including the ones in the tax box.
A payable, read from the image
Purchases and payables
A line table drawn without rules
No gridlines, so the rows exist only as spacing. A description that wraps onto a second line is the common place two items collapse into one.
A payable, one row per line
Purchases and payables
An invoice in a language nobody downstream reads
The field names are translated and the numbers are not. Nobody in the approval chain can sanity check the extraction by glancing at the page.
A payable, fields labelled in yours
Purchases and payables
A new supplier on a layout you have never seen
Nothing about it was known in advance. This is the case a template based tool has no entry for, and it is the ordinary case in a long tail of small suppliers.
A payable, same as any other
Purchases and payables
What happens to one document
Your documents
Invoice arrives. Your documents
Email, supplier portal or scan. Nothing changes about how.
T+0
Fabrx
Identify the document. Fabrx
Before anything is read as an invoice, we establish that it is one.
If this is not an invoice
Hold for review. Fabrx. Needs a person.
Your AP clerk sees the document and the thing that stopped it.
needs a person
Then it rejoins the path above.
Read the header. Fabrx
Supplier, invoice number, dates, currency, totals, tax, your reference.
Read every line. Fabrx
Fields are read by what they are, so an unfamiliar layout is an ordinary one.
Check the numbers agree. Fabrx
Lines against the total, tax against the rate. A number that disagrees with itself stops here.
If the numbers do not agree
Hold for review. Fabrx. Needs a person.
Your AP clerk sees the document and the thing that stopped it.
needs a person
Then it rejoins the path above.
Send the payload. Fabrx
A posting-ready record at the endpoint you nominate.
Where we hand over
The handoff is a JSON payload to an HTTPS endpoint you own, or a webhook we call on your side. We hold no credentials to your ledger.
Your AP system
Becomes a payable. Your AP system
Your matching rule and your approval chain run where they run today.
What we do, and what stays yours
Fabrx builds and runs
- Identifying what each document is, so an AP inbox does not need sorting first
- Reading the invoice, in whatever layout it arrives in
- Pulling the header fields and every line item, including tables that cross pages
- Checking the numbers against each other, lines against the total and tax against the rate
- Stopping anything that does not settle, with the invoice attached to it
- Delivering a posting-ready payload to an endpoint you nominate
- Adding fields when what you need changes
You keep
- The rule that decides what counts as correct for your ledger
- Your purchase order data, and what you treat as a match
- Your approval chain, which does not move
- The decision to pay
- Your ERP or AP system, and who is allowed to see it
- Somebody to look at the output and tell us whether it is right
When a document is wrong
A pipeline with no failure path is a diagram of a good day. These are the ones that happen.
The lines do not add up to the total
Extraction finishes and the invoice stops before your endpoint, with the sum we read and the total printed on the page shown side by side.
Who sees it. Your AP clerk, in the review queue.
- Corrected and released
- Sent back to the supplier
- Released as it stands, with a note
It is not an invoice
A statement, a remittance advice or a duplicate of one you already have. It is identified as something else and set aside, and we are told. Nothing reaches your endpoint.
Who sees it. Us first, then you, usually the same day.
- Added as its own document type, usually in one business day
- Confirmed as out of scope
The page cannot be read
A scan too dark to resolve, or a photo too blurred at the line table. The invoice stops before extraction and no value is guessed at.
Who sees it. Whoever sent it, by reply, with the page that failed.
- Resent and reprocessed
- Keyed by hand, the way it is done today
The supplier is not one of yours
The invoice is held with the supplier details we read off it. We never create a supplier record in your system.
Who sees it. Your AP clerk, with the name, the bank details and the address.
- Matched to a supplier you already have
- Supplier created by you, then released
Four approaches to AI invoice processing
The same invoice, four approaches, and what each one asks of you.
What it costs to start
By hand
No new spend. The cost is hours your team is already paid for.
OCR with templates
A licence and an implementation, before you know it reads yours.
An LLM you wire up yourself
Model credits, plus the engineering time to build and keep it.
Fabrx
Nothing. The build is free and you pay per document once it runs.
Who defines the fields
By hand
Whoever is keying, differently each time.
OCR with templates
You do, in their configuration, and you maintain it.
An LLM you wire up yourself
You do, in a prompt and a schema you keep tuning.
Fabrx
You tell us on the call, and we build to it.
An invoice layout nobody has seen before
By hand
A person reads it the same way they read the last one.
OCR with templates
No template exists for it, so it drops out until one is built.
An LLM you wire up yourself
Usually handled, and you find out which way per invoice.
Fabrx
Read like any other. Fields are read by what they are, wherever they sit on the page.
How a wrong value reaches you
By hand
As a typo, caught by whoever checks the batch.
OCR with templates
As an empty field or a misaligned one, usually visible on sight.
An LLM you wire up yourself
As a confident, well-formed number. A schema validates the shape of the answer and has no view on whether it is the right one.
Fabrx
It stops at the arithmetic. Lines against the total and tax against the rate is a check on the value, and anything that fails it waits for a person.
Where the data lands
By hand
Retyped into the system you already use.
OCR with templates
Their platform, then an export or a connector into yours.
An LLM you wire up yourself
Wherever your script writes it, which you also own.
Fabrx
A record at an endpoint you own, inside the system you already use.
Someone has run this
They built solutions for any situation encountered along the way. For us, this type of partnership matters.
- 300+
- Field agents using it every working day
- 4,000+
- Documents processed every month
- 4 yrs
- In production, on the same workflow
Blitz is a real estate field agency network, and these figures come from ID documents turning into signed contracts. They are not invoice figures. The pipeline is the one described above, running on a different document type.
What a pilot measures
Measured on your documents
Field-level accuracy
Every field on every sample invoice, scored against your correction.
Sample size
How many of your real invoices we ran. You choose which.
Exception rate
The share that stopped for a person, and what stopped them.
Line-level completeness
Lines returned against lines printed, counted per invoice.
Build time
Calendar time from your samples reaching us to the pipeline running.
What we ask you for
How many invoices do you process in a month?
It sets the pricing band, because the rate depends on volume.
How many need the line items as well as the total?
Line extraction is the harder half and it changes the build.
What gets checked today before an invoice is posted?
It tells us which checks to run before the payload leaves us.
Who looks at a posted value when somebody thinks it is wrong?
That person is who the review queue is built for.
How often does a supplier send you a layout you have not seen?
It tells us how much of your tail is new suppliers every month.
What it costs
- 01
We build it
We build the extraction, usually within one business day of receiving your sample documents.
- 02
You check it
It costs you nothing until it runs on your real documents. If the accuracy is not good enough, nothing goes into production and nothing is charged.
- 03
Then you pay
A prepaid package priced per document, valid for a calendar year. The rate depends on the document type and your monthly volume.
- No subscription
- No per-seat charge
- No setup fee at any volume
Questions about AI invoice processing
What is AI invoice processing?
It is the use of a model to read an invoice and return its fields as data, so that supplier, invoice number, dates, tax, totals and line items arrive in your system without anybody keying them. The reading is the part the AI does. Matching, approving and paying stay where they are.
What is the difference between OCR and AI invoice processing?
OCR turns the picture of a page into characters. It tells you a string says 4,812.00 and has no view on what that number is. AI invoice processing reads the same page for meaning, so it can tell the net from the gross and a line from a subtotal. Most template based tools pair OCR with a map of where each field sits, which is why they break when a supplier moves one.
How accurate is AI invoice processing?
It depends on your documents, which is why we will not publish a number here. A vendor quoting one measured it on their sample and not on yours. We build the extraction on your invoices first and show you what it got right, field by field, and you decide from that. The build is free, so finding out costs you time and no money.
Can AI read invoice line items?
Yes, and it is the part worth testing. Header fields sit in roughly the same places across suppliers. Line tables wrap across pages, carry descriptions that run onto a second row, and are often drawn with no gridlines. When we run your samples we count lines returned against lines printed, per invoice.
Can AI read an invoice layout it has not seen before?
Ours does. The extraction reads fields by what they are, wherever they sit on the page, so a new supplier and a redesigned form are ordinary cases. There is no template per supplier to build and none to repair. A rebuild happens when what you need changes, such as a new field or a new document type.
What happens when the AI gets an invoice wrong?
Some wrong values fail the arithmetic and some do not. A sum that disagrees with the printed total, or tax that disagrees with the rate, fails the check and the invoice waits for your AP clerk with both numbers shown. A document that turns out to be a statement is set aside before extraction. A value that is wrong and still internally consistent is why somebody on your side reviews the output, and we say so on the call.
Related
- invoice processing, step by step
The stages an invoice moves through, and where the hours go before any of it is automated.
- how the build works
The call, the sample documents, and what happens in the business day between them.
- a four-year extraction in production
Different documents, same pipeline. 300 field agents, 4,000 documents a month, running since 2022.
Send us twenty invoices, including the awkward ones
We build the extraction and show you what it got right, field by field, before you pay anything.