Process your documents,at a fixed amount
Ten thousand leases to review before closing, a mailbox to classify, scanned forms to key in. The tool your provider or your team puts in place reads them, pulls out fields, a category or a summary, and calls a language model for each document. With Agent Libre behind it, that model is served in Canada for a fixed monthly amount, and the batch’s end date is known before you launch it.
Talk about your documentsWhat breaks when a batch is billed by usage
Three situations the person launching the batch knows well.
The batch cannot be budgeted before it is launched
Nobody knows the average length of the contracts before reading them. Billed by usage, the price of the February statements or the June renewals shows up at month end, when the year’s budget is already closed. So the team launches a small sample, waits for the invoice, then negotiates the rest.
A hundred-page contract is read in pieces
A lease with its schedules or a collective agreement exceeds what many models accept at once. The tool splits it, and the clause on page 84 that refers to a definition on page 3 is extracted without its context. The reviewer redoes those cases by hand.
Sorting what may leave the country takes longer than the processing
Before sending a batch to an American provider, someone decides which files contain personal information and pulls them out. For client files or emails, that sorting happens file by file, and it delays every batch.
You, your tool, and the model that answers it
Your tool stays the same. What changes is the model that answers it.
Your organization
Your teams, your data and your rules. You decide what the tool does, and with which data.
Your document processing tool
The software that takes in your contracts, forms and emails by the thousand and returns the fields to key in or the category to file each one under. Your provider or your team sets it up and decides what it looks for in each document.
The model that answers the tool
Every time the tool needs an answer, it calls a language model. That model can come from different providers. This is where Agent Libre comes in.
What Agent Libre changes behind your tool
The tool keeps the reading, the extraction and the classification. What changes shows up in four places.
The end of the batch is computed before the start
A processing speed is reserved for you. The batch volume divided by that speed gives the end date, and that date lines up with your team’s calendar: a batch that must finish before the quarterly close is planned backwards from it.
The queue waits, the bill does not climb
When several batches leave the same day, the tool’s requests wait their turn behind the reserved speed, with no cutoff and no extra fees. The month of tax statements costs the same amount as the month of July.
A long file fits in a single request
A hundred-page lease with its schedules is read in one go, cross-references included. The tool can also send the same document a second time, for a finer extraction or after a validation reject: re-read text is included in the plan.
Data in Canada
The model is hosted and operated in Canada. Sorting out the files that cannot leave the country goes away, since none of them leave. Documents are not kept beyond processing and are never used to train a model.
Three rules for a batch
A sample first
The free 24-hour to one-week audit measures what your tool consumes with your current provider, on a sample of documents. The speed to reserve and the duration of the full batch are computed from that measurement.
Documents already in text form
The model reads text. Scanned PDFs and photos of forms go through your tool’s character recognition first.
Twelve months, then a review
The monthly amount is written in the contract for twelve months. At renewal, measured consumption revises it, up or down.
What each team decides before connecting the tool
Four decisions come before changing the model behind your processing tool.
Take two hundred documents already processed and validated by your team, run them through the tool again with the new model behind it, and compare field by field. Set in advance the error rate you accept on the fields that matter, a contract amount or a due date. The switch of the full batch is decided on that result.
Check three things in your tool: that it lets you change the model’s address, that it converts scans to text before sending, and that it asks for a document again when the answer does not match the template. The model change takes three lines of configuration; those three checks take longer than the change itself.
List the fields that play no part in the extraction (signatory’s name, address, account number) and have Agent Libre’s gateway mask them before they reach the model. That list covers the whole batch, replaces file-by-file sorting, and serves as the basis for the assessment required by Quebec’s Law 25. Details are on the Governance page.
Decide with operations whether the batch must finish in two weeks or in three months. The reserved speed, and therefore the monthly amount, depends on that answer. A spread-out batch costs less per month; a rushed batch costs more per month, but it finishes before the close.
What we get asked on this topic
Does Agent Libre provide the processing tool?
No. The tool that reads your documents, extracts their fields and writes the results into your systems comes from your provider or your team. Agent Libre serves the language model that tool calls.
Does extraction quality hold up on our documents?
It is checked on your documents. Put in the comparison sample the cases that make extractions fail today: multi-page tables, clauses that refer to one another, schedules scanned askew. The field-by-field result on those cases decides the switch.
Our volume doubles next year. What happens to the plan?
The amount written in the contract does not move during the twelve months. If the volume doubles, each batch takes twice as long at the same speed. At renewal, measured consumption is used to set the new speed and the new amount.
What does the model see of our documents?
The text the tool sends it, after the masking set in the gateway. It sees neither the original file, nor the images, nor the documents the tool keeps on your side. Nothing is kept beyond processing, and requests are never used to train a model.
Do poorly scanned handwritten forms get through?
The model receives the text produced by character recognition. A poorly scanned handwritten form yields text full of holes, and the extraction suffers, whatever the model. Put those cases in the comparison sample and plan a check by a person for them.
See also
Internal document assistant
What each team decides before putting a model served in Canada behind the assistant that answers from your policies, procedures and contracts.
Professional services
What a law, accounting or consulting engineering firm decides before putting a model served in Canada behind its contract review and file search tools.
What “hosted in Canada” actually means
Where documents sent to a model end up, and what a server in Canada changes.
Your company has two supply chains
The first one you already know. Materials come in, they are transported, stored, transformed, delivered. Every step has a supplier, a lead time, a unit cost. When the margin tightens, you walk down that chain step by step, and you generally know where to look.
The second one moves information instead of materials. A price request comes in and becomes a quote. An order comes in and becomes a production order, an invoice, a confirmed delivery. A site report becomes a change order, or it does not. This chain runs through every one of your departments and it employs a large share of your people.
The 24-hour to one-week audit
We measure your actual consumption at your current provider: who calls what, what it costs, and what is leaving your walls. Free, no commitment, and the report is yours even if you go no further.