Quote request intake
Quote requests arrive as free-form emails. Someone reads each one, finds the product codes and quantities, checks them, and chases whatever is missing. This demo does the reading and checking, hands the decision to a person, and measures its own accuracy. Every email, customer and product here is made up.
Honest label: no AI model runs in this demo. The reader is a set of plain rules, because the point is everything around the reader: the checks, the hand-off to a person, and the measurement. An AI model could replace the rules and the rest would stay.
Read a request
It runs in your browser. Nothing you type is sent anywhere. Please do not paste real customer emails.
What it read
Checks
Draft reply
Drafts are never sent automatically. A person reads and approves them.
Where a person fits
- Read. The reader pulls out the customer, products, quantities, needed-by date, PO and ship-to.
- Check. It tests the request against the catalogue: is the product code real, is the quantity at or above the minimum, does the date allow for the lead time.
- Route. A complete request goes to a person to price. An incomplete one gets a drafted follow-up. Anything the reader is unsure about is handed straight to a person instead of guessed.
- Approve. Nothing is entered into an order system or sent to a customer without a person saying yes.
How well does it work?
Automation you cannot measure is automation you cannot trust. This runs the reader, live in your browser, over 80 labelled made-up requests where the right answer is known.
By kind of request
By field
Fields it did not read correctly
Read this before trusting the score
The same person wrote the reader and the test emails, which flatters the reader. A real test uses real past requests, including ones nobody thought of in advance. And a good score on the cases you can imagine says nothing about the ones you cannot. These three realistic requests fool the reader without any warning:
This is why a person approves every request. The goal of the automation is to remove the routine reading and checking, not the judgment.
How it works
- Labelled test data. A seeded Python script writes 80 made-up emails in four styles (clean, informal, tricky and incomplete) with the correct answer for each, so the reader can be scored. It is not based on any real company’s emails or catalogue.
- One reader, shared. The same JavaScript reads the sample you pick and the whole test set, so the score on this page is the score of the code you are using.
- Rules decide the route. Required fields, the catalogue, minimum quantities and lead times are explicit checks. When the reader is unsure (a quantity written in words, a date like 28/11) it says so and hands the request to a person.
- The page measures it. Decisions are scored as correct, handed to a person, or wrong, and the one that matters most, marking a bad request as ready, is counted on its own.
In a real engagement the reader is usually an AI model, which copes with messier text than rules can. The checks, the hand-off, the human approval and the measurement stay the same, and the same test set tells you whether the model is good enough before it touches live work. The source code is public on GitHub.
Got a pile of requests someone retypes by hand?
Tell me what arrives and where it ends up.
Start a conversation