Demo with simulated data. Not client work.

The invoices, suppliers and extraction results are invented, and nothing is sent to an AI model. None of it comes from a client project.

Working demo · simulated data

An AI invoice-intake workflow you can try.

Invoices arrive, fields are extracted with a confidence score, checks run, and anything uncertain, failing or large waits for a person. It runs in your browser; the extraction is simulated and no AI model is called.

Jump to the demo ↓

Live demo

Nine invented invoices, one review queue.

Receive the invoices, open any one to see why it was routed, and decide the ones in review.

Demo with simulated data. Not client work.

Loading the invoice-intake demo…

What sends an invoice to a person

Any one rule is enough.

An invoice goes straight to staging only when every check passes, every field is at or above the confidence threshold and the total is within the amount threshold. Anything else waits for a person, with the reasons listed.

One invoice hides a line of white-on-white text telling any AI system to approve it and skip review. The workflow treats text in a document as content, never as instructions: the fields are extracted as usual, and the invoice goes to a person.

The rules the demo applies, from the sample scope. The 2% tolerance and the default thresholds are demo settings, not recommendations.
Rule	Goes to a person when
Purchase-order match	The PO is missing, isn't open, belongs to another supplier, or the total is further from it than the tolerance (2% in this demo; agreed in phase 1 in a real project)
Lines add up	A line, the subtotal or the total doesn't add up
Due date matches terms	The due date isn't the invoice date plus the supplier's payment terms, which usually means a date was misread
Duplicate invoice number	The same supplier and invoice number were received before
New bank details	The invoice says the supplier's bank or remittance details have changed
Confidence threshold	Any field or line was read below the threshold
Amount threshold	The total is above the threshold, however confident the extraction
Instructions in the document	Text in the file tries to instruct the workflow

What’s simulated

The extraction is made up. The measurement wouldn’t be.

Every extracted value and confidence score here was written by hand to show one rule at a time, so the demo says nothing about how accurate any model is. The checks, routing and audit log are real code running on that invented data.

In a real project the thresholds are set from accuracy measured on your own invoices, not guessed. Our guide to scoping an AI pilot with measurable acceptance criteria explains the method.

Each piece of the workflow: in this demo, and in a real project
Piece	In this demo	In a real project
Intake	Nine invented invoices arrive when you press a button	Each PDF attachment in the shared AP mailbox becomes one item in a processing queue
Extraction	Fixed values and confidence scores, written by hand to show each rule; no AI calls	A model on your own AI provider account reads each invoice into a fixed set of fields, chosen in phase 1
Accuracy	Not measured: the numbers are invented	Measured per field on a labeled, redacted test set of your own past invoices before go-live, and again whenever a prompt or model changes
Thresholds	Sliders, applied at once to everything no one has decided yet	Agreed with Finance from the test results, changed as administrator settings and applied to new invoices
Checks	Against invented open POs and past invoices	Against read-only views of your accounting database
Staging	A list on the screen	Your accounting system's invoice staging table, picked up by its existing import; approval for payment stays where it is today
Audit log	Kept in your browser until you reload	Stored with the user and time for every extraction, check, decision and correction

Each piece of the workflow: in this demo, and in a real project

Piece: Intake
In this demo:
Nine invented invoices arrive when you press a button
In a real project:
Each PDF attachment in the shared AP mailbox becomes one item in a processing queue
Piece: Extraction
In this demo:
Fixed values and confidence scores, written by hand to show each rule; no AI calls
In a real project:
A model on your own AI provider account reads each invoice into a fixed set of fields, chosen in phase 1
Piece: Accuracy
In this demo:
Not measured: the numbers are invented
In a real project:
Measured per field on a labeled, redacted test set of your own past invoices before go-live, and again whenever a prompt or model changes
Piece: Thresholds
In this demo:
Sliders, applied at once to everything no one has decided yet
In a real project:
Agreed with Finance from the test results, changed as administrator settings and applied to new invoices
Piece: Checks
In this demo:
Against invented open POs and past invoices
In a real project:
Against read-only views of your accounting database
Piece: Staging
In this demo:
A list on the screen
In a real project:
Your accounting system's invoice staging table, picked up by its existing import; approval for payment stays where it is today
Piece: Audit log
In this demo:
Kept in your browser until you reload
In a real project:
Stored with the user and time for every extraction, check, decision and correction

The demo also leaves out parts of the sample scope: there, invoices from a new or unmatched supplier and files that aren't invoices also go to a person, and invoices above the amount threshold need a second approval by the AP lead before they are written to staging.

See it on paper