Review Contracts with AI Without Uploading Them to the Cloud

Most AI tools handle a contract the same way: you upload the file, a vendor’s servers read it, and an answer comes back. Enterprise plans add contract terms around that exchange, such as limits on training and retention. Those terms are real, and they are worth reading closely. They are also promises about what a vendor will do with a document it has already received.

Private contract review removes the exchange. The model runs on a server in your building, and the network is built so that server has no route to the internet. The contract is read where it already lives, and nothing is sent anywhere to be read. That is the difference between a contractual control and an architectural one: terms can be amended, while a firewall rule that blocks outbound traffic does not change when a vendor updates its terms of service.

In the system we build, production documents, prompts, embeddings, model inference, chat history, logs, and backups stay on hardware inside your network. Approved maintenance uses a separate channel you control, opened for a scheduled window and disabled again when the work is done.

Neither approach is right for every office. For a head-to-head on data path, cost shape, and when each one fits, see private on-premise AI vs ChatGPT Enterprise and Claude Enterprise. If a client agreement or your own counsel says a contract cannot be disclosed to a third party at all, private AI for data that can't leave the building is the page that covers that case.

The system is a private chat on your internal network, with individual accounts, pointed at the folders where your contracts already live. A document watcher notices new and changed files and keeps the index current, so nobody uploads anything. Three kinds of work come up most.

• Clause lookup: ask what the vendor services agreement says about termination notice, and the system retrieves the relevant clauses and answers from them rather than from the whole file.
• Comparison against your standard terms: put an incoming draft beside your own standard terms and ask where they differ, so the attorney starts with the differences instead of reading cold.
• Cited answers: every answer links the clause and the source file it drew from, so the reader opens the contract and confirms.

Under the hood this is retrieval-augmented generation. Before the model writes a word, the system retrieves the passages from your own contracts that best match the question. Then a local model writes an answer grounded in those passages and cites them. Chunking and retrieval are tuned against your real contracts during onboarding. The full pipeline, from a file landing on your share to a cited answer, is walked through in how on-premise RAG works.

The comparison workflow is scoped with you in discovery: which clause types matter, and which of your documents count as the standard. The diagram shows the idea with illustrative clause types.

We do not promise that the system surfaces every difference. Answer quality depends on retrieval quality, which depends on tuning against your real documents, and that is why a person reads the cited source before relying on anything.

The system assists. It does not give legal advice, and it is not built for unsupervised legal decisions. The platform is designed for assistive work with a human in the loop.

In practice an attorney still reads the contract, decides what a clause means for the matter, negotiates, advises the client, and signs off. What changes is the time spent finding and comparing. Citations make checking fast: if retrieval surfaces the wrong passage or an answer over-summarizes, a reader who opens the cited clause sees it in seconds.

Citations also give us a tuning loop. When an answer cites the wrong passage, that is a retrieval problem with a visible trace, and we adjust it during onboarding. Grounded systems fail where you can see them.

Hardware

The machine that runs a private AI system for a small office sits on a desk or a shelf; it does not need a data center. We choose the exact machine only after the prototype shows how many people will use the system and how many contracts it has to index, so you do not overbuy or underbuy. The options we quote, and why memory is the spec that matters, are in the private AI server hardware guide.

Cost

On-premise cost has three lines: the server, which you buy once and own, with current hardware in the $2,500 to $6,000 range; a one-time implementation scoped in discovery; and a recurring operations line that your IT team can own or we can run. There is no per-seat license and no per-token meter, so adding a reviewer does not add a bill. The breakdown, and how it compares with subscription seats, is in what on-premise AI actually costs.

It starts with discovery, and discovery ends in a working prototype. We interview your team, take a sample or de-identified set of your contracts, and build a prototype you can watch answer real questions. No server is purchased until that use case is proven.

From there the steps are the ones on the Private On-Premise AI Solutions page: size and procure the server from what the prototype showed, install it on site with network isolation, load your contracts, and train your team to verify answers against their cited sources. If you want it, we can then handle patches, model upgrades, and monitoring as a managed service. If you stop managed operations later, you keep the hardware and the data.

We already build public sites and internal tools for law, accounting, and advisory firms, described on our professional services page. If a firm later wants private AI on its own documents, the service described above is the one it would use.

Does the AI see contracts from other clients?

Only the people you allow can reach a document through it. The system uses individual accounts and role-based permissions, and it enforces them at answer time: a question from someone without rights to a document does not retrieve from it, and an audit log records who accessed what. Which roles see which folders is designed with you during discovery. Because everything runs inside your network, no vendor or outside organization receives your contracts either.

Can it work fully offline?

Answering questions does not need the internet. The models, the index, and the chat all run on the server, general outbound access is blocked at the firewall, and the server has no persistent public endpoint. The one exception is maintenance, which uses a separate channel you control, opened for a scheduled window and disabled again afterward. The exact mechanism is agreed during security design.

Is this the same as ChatGPT Enterprise?

No. ChatGPT Enterprise is vendor-hosted, so your files are processed on the vendor’s systems under enterprise terms that can limit training and retention. A private on-premise system keeps the models and documents on hardware you own, with outbound internet blocked. Neither is better for every office, and the comparison article covers data path, cost shape, and when each one fits.

What file types does it read?

The pipeline starts at your file share, and the documents we show there include Word files, PDFs, spreadsheets, and scans. Which formats you actually have, and how much of each, is one of the first things discovery pins down, and we test those formats on your sample set before anything is purchased.