Read a batch of contracts and pull out the clauses that matter, with an AI model running on your own hardware.
A 30-page contract holds about ten facts that a management team actually needs: who signs, for what, from when, for how much, on what payment terms, with what notice period, under which law. This tool extracts them automatically, contract by contract, and lays them out in a table. What sets it apart: the AI model is installed on a machine inside the company (an open model run with Ollama). The text the model analyses is never sent to an online service. It was built for legal and finance departments that, rightly, refuse to hand their contracts to an external AI provider. The tool is currently at the internal test stage: it has been evaluated on a batch of 11 contracts and is not yet installed at any client. Deployment starts with a trial on your own contracts.
For a legal or finance department, the question is not whether AI can read a contract: it is where the contract goes. By placing the model on a company machine, the tool answers that question before performance is even discussed. The expected outcome for the team: a consistent clause table, reviewed by a human, obtained without having to negotiate an exception to the confidentiality policy.
A finance department wants to know which contracts end within the year and under what notice they can be terminated. The batch is processed on the local machine; the resulting table shows, for each contract, the effective date, term, termination conditions and notice period. A lawyer reviews the rows marked "not stated" and completes them by hand.
Ahead of an external audit, the company must present the confidentiality, penalty and governing-law clauses of its supplier contracts. Local extraction produces the table in one pass; the legal department checks and annotates it. The contracts are never shared with a third party for this step.
This reconciliation does not exist in the current tool; it would be built at scoping as an extension. Amounts and payment terms extracted from each contract could then be compared with the invoices received, and discrepancies (different amount, payment term not respected) listed for human verification, alongside purchase invoice processing.
This product applies to the following business functions. Click to discover how we adapt it to each.
These products combine with this one for complete programs.
AI-augmented search on your document heritage — answers grounded in your documents, with sourced citations.
Keep humans in the decision on critical cases, at scale.
Analyst productivity on local workstation, sensitive contexts mastered.
No. The model is installed on a machine you control (a powerful workstation or an internal server). Extraction happens on that machine, with no call to an online service. This is the starting point of the project: it was designed for organisations that refuse to hand their contracts to an external AI provider.
On our internal test bench, 11 contracts of varied nature (including one in English; the scanned contract in the batch had been converted to text beforehand) yielded 107 to 108 correct fields out of 110, with no invented value. That figure applies to this test batch, not to every contract in the world. Before any installation, we repeat the measurement on a sample of your own contracts and give you the figure obtained.
Not yet. The tool is at the internal test stage: it has no graphical interface, runs from the command line and is not installed at any client. Deploying it at your company involves scoping (choice of machine, field grid, output format), a trial on your contracts, then set-up. The duration is set at scoping, after that trial.
You need a machine able to run an open language model: a recent workstation with a graphics card or enough memory, or a small internal server. Exact sizing depends on contract volume and expected speed and is decided at scoping. No subscription to an online model is required.
It does not decide. The model is instructed to answer "not stated" rather than guess, and the table it produces is a draft for review by a lawyer or finance professional. No decision (termination, formal notice, payment) is taken by the tool.
Enterprise document RAG answers a question asked over a document base; this tool fills a fixed grid, contract by contract, with no question to ask. The two complement each other: the clause table serves as an inventory, RAG serves to query the base once it is built.
Free initial scoping, 30 minutes to 2 hours, with a solution lead adapted to your sector.