Ask five firms what “no data leaving” means and you get five different answers.
One partner means the documents never sit on a vendor’s servers. Another means the vendor promised not to train on them. A third means the copy stayed in the country.
Those are three separate requirements. Each one points at a different deployment.
So the honest answer is not a product name. It is a short list of deployment shapes for AI document review, what each one stops, and the questions that show the difference before you sign.
The Short Answer: Three Shapes for Document Review AI
Three setups can satisfy a no-egress requirement for document review. Choose by how literal your rule is.
Your own infrastructure. The application and the model run on servers you own, inside your network or your private cloud. Documents go in, answers come out, and nothing crosses your egress rule. Open-weight models make this practical, because you can host them yourself and audit the weights.
A hard air gap. Same as above, with no outbound connection at all. Updates arrive on media or through a controlled transfer. This is the only shape that survives a strict “nothing leaves, ever” policy, and it costs the most to operate.
Contained execution on a hosted service. The vendor runs the system. You get isolation, a no-training commitment, encryption, and deletion controls. It is quick to start, and the weakest of the three if your rule is literal.
A promise about training belongs in the contract. It does not move one byte off your network. On its own, it does not answer the question you asked.
Four Things People Mean When They Say the Data Must Not Leave
The phrase carries at least four meanings. Getting your risk committee to pick one is most of the work.
Does the vendor train a model on my documents?
This is the meaning most people have in mind, and the easiest to check. Ask for the clause. Then ask whether it covers fine-tuning, evaluation sets, and human review of flagged conversations.
What stays on disk, and for how long?
A no-training promise says nothing about storage. Documents, page images, extracted text, embeddings, and logs can all persist after a matter closes. Ask for the retention schedule per artifact type, and for the deletion path in writing.
Which country holds the copy?
Residency is a legal question about where the storage sits. It is not a question about where the company is registered. A vendor can be local and still replicate to another region.
Does anything derived from the document cross my network boundary?
This is the one that gets missed. An embedding is a numeric summary of your text. A telemetry ping can carry a file name.
An OCR call can ship a page image to a third party. If your rule is literal, derived artifacts count as data.
Write down which of the four is absolute for the matter in front of you. Note which three can be contractual. That page takes ten minutes and settles most internal arguments before they start.
Why the ABA Made Confidentiality a Duty in AI Document Review
On 29 July 2024 the American Bar Association issued Formal Opinion 512 on generative AI tools.
The opinion applies Model Rule 1.6 to AI use. Confidential client information stays protected when it goes into a tool. The lawyer then has to make reasonable efforts to prevent unauthorised disclosure.
Reasonable is the operative word. It is a standard you can meet, and one you have to be able to show you met.
In practice the opinion asks for a small set of things. Know what the provider does with prompts: training, retention, logging, subprocessors. Choose configurations that reduce exposure.
Send less. Control who can reach the output. Supervise the people using it.
None of that requires an air gap. It does require that you know which deployment shape you bought, because the diligence questions change with the shape. Firms working through this often start with the legal practice page, which sets out the confidentiality controls in plain terms.
One boundary worth stating. Whether privilege is waived by a given disclosure is a legal question that turns on your jurisdiction and your facts. Ask your own counsel.
What this article can tell you is what each deployment does with the file.
The Egress Test, and Four Questions About AI Data Handling
AI document review without data leaving comes down to a property you can test. Put the tool on a machine behind a firewall rule that blocks outbound traffic, then give it a review task.
If it works, you have an on-premises deployment. If it stalls, a network call was doing the work, and no amount of contract language changes that.
Run the test before the pilot, not after. It takes an afternoon. It ends the argument.
A failed test is not always fatal. It tells you which shape you are actually buying. That is the fact you need before you can decide whether the shape fits the matter. Four questions to ask alongside it.
Where does inference run? Your servers, the vendor’s cloud, or a third party the vendor has not named. Ask for a diagram with hostnames on it.
What leaves besides the document? Telemetry, crash reports, usage metering, OCR, embeddings, backups. Ask for the list, and for the switch that turns each one off.
Who can read my matter? Support staff, engineers on call, contractors. Ask about access logging, and whether support access needs your approval per session.
What happens at matter close? The delete action, who can trigger it, how long backups persist, and whether you get an attestation afterwards.
If a vendor cannot answer these in writing, you have your answer.
What Document Review Looks Like With the Model Inside Your Network
The work itself does not change much between shapes. The plumbing does.
Start with ingestion. Native PDFs, Word files, scanned images, email exports.
Scanned pages need OCR before any model can read them, and OCR quality decides everything downstream. There is a useful explainer on document compression and OCR that covers why long, image-heavy files are the hard part.
Then chunk, embed, and index. The index sits on your storage. Retrieval pulls the passages that match the question, and the model answers from those passages only.
The difference between retrieval and agent memory is worth an hour before you buy either, because the two fail in different ways.
Then the review tasks. Issue tagging, clause extraction, chronology building from a set of statements, and a first-pass privilege flag with the basis attached. Each document summary arrives with a page citation.
Take one concrete case. Two hundred scanned invoices, a disputed schedule of works, and a question about when the delay started. The useful output is a dated table where every row links back to the page it came from.
A fluent paragraph with no page numbers is worthless in front of a judge.
Cite everything. A summary an associate cannot check against the page is a liability. The output has to point back to the source text so a human can confirm it in seconds.
Sign-off stays human. The model flags, the lawyer decides. Every coding decision lands in an audit log with the model version attached.
Then it feeds back. Every accepted or rejected flag is a label, and labels improve the next pass. That loop is where the time saving comes from, and it only works if the loop stays inside your environment.
Where Hosted AI Document Review Still Fits
Not every firm needs an air gap. Pretending otherwise wastes money.
Plenty of matters are sensitive without being catastrophic. Plenty of work also sits around the review rather than inside it. Building the chronology, drafting the cover memo, testing your own argument before opposing counsel does.
That is the layer OpenCraft AI’s sandbox is built for. The hosted product runs in the cloud, and the containment is stated plainly. It runs real code and real files inside an isolated sandbox.
It cannot escape. It cannot delete your files. You can leave it running unattended, and the same page commits to open-weight models with failover between them.
The legal page for law firms adds the controls a firm will ask about. Encryption end to end. AES-256 across files, chats and documents, no training on client content, one-click deletion at matter close with retention policies respected.
One setting deserves attention, because it is the part firms get wrong.
Ocai remembers your work across sessions. That is what makes the tenth brief sharper than the first. It is also a retention decision you just made.
The write side is the part that needs a policy, and the comparison of retrieval and memory explains why.
So treat memory as a matter artifact. Open one workspace per matter, then delete it when the matter ends. Point your retention policy at the workspace, not at the user.
Where a firm’s rule is literal rather than contractual, the same vendor publishes a different deployment. The enterprise page describes the application and the AI running on infrastructure the client owns, in the client’s country.
Its own summary line reads: “The data never leaves.” It is positioned as a proof of concept on your own files. For a firm with a genuine egress rule, that is the version to test.
Cost follows tokens rather than seats. The pricing page lists rates per million tokens, by model.
If your finance partner wants a per-matter number, run one real matter and read the usage report. Projections guessed in advance tend to be wrong in both directions.
What Belongs in the Contract for Document Review AI
Clauses are cheap at signing and expensive to add later. Six lines carry most of the weight.
One: no training on client content, covering fine-tuning, evaluation, and any human review of conversations. Two: a retention schedule per artifact type, with a deletion attestation on request.
Three: named subprocessors, with notice before a change. Four: support access only with your per-session approval, and an access log you can export.
Five: a notice window for a security incident, short enough that you can act on it. Six: the right to run your own test on your own hardware, at any time, without asking permission.
That last one is the one to fight for. If a vendor will not let you run the test, treat their architecture as unverified, because that is what it is. The enterprise deployment is the one to point at when your risk committee asks whether the clause is enforceable in practice.
What Self-Hosting Document Review AI Actually Costs
Self-hosting is the honest answer for a firm with an absolute rule. It is also a second job.
Start with compute. GPU capacity sized for the models you plan to run, plus headroom for the busy week. Then a serving stack, and a person who can keep it patched.
Then an evaluation set: documents your lawyers have already coded, so you can measure whether a model change helped or hurt. Then a path from the output back into your document management system. Without that path, nobody will use it.
Maintenance runs from there. Model drift, prompt changes, retrieval quality, storage growth, incident response, and the version log your audits will ask for.
You stop paying a licence. You start paying for hardware, for a person who can run it, and for the weeks it takes to wire into the DMS.
The trade is real in both directions. A 30-document matter does not need this. A production review with a rule that says nothing leaves probably does.
If you want the component list before you commit, there is a guide to running AI on your own infrastructure that names the parts and prices the overhead.
There is a wider argument for owning the stack in Open Source AI Is No Longer Optional and in AI Model Sovereignty. The short version: access you rent can be withdrawn. Weights you hold cannot.
A 30-Day Test Before You Sign a Document Review Contract
Thirty days is enough to find out whether a deployment is real.
Days 1 to 3. Write the rule. One page, stating which of the four meanings is absolute, which matter types it covers, and who signs off on an exception. Without that page, every demo looks compliant.
Days 4 to 7. Pick one closed matter. Past work, already billed, no live client exposure, with scans, emails, and at least one contract. Give the same document set to two candidates.
Days 8 to 10. Run the egress test. Firewall rule, outbound blocked, one review task. Record what happens.
Days 11 to 20. Run the work. Ask each candidate the same five questions. Start with the chronology, then move to the privilege flags.
Then the harder tasks: extract indemnity and change-of-control clauses, summarise fifty documents with page cites, and list every place two documents contradict each other.
Days 21 to 25. Score it. Time to first useful output. Error rate against your own coded set.
Then the softer signals. Citations that resolve to the right page. Number of times a lawyer had to re-explain the task.
Days 26 to 30. Price it properly. Licence or credits, hardware, staff time, integration, and the cost of the next model upgrade. Take the per-token rates from the published rate card, then replace them with your own measured usage.
Put the four numbers on one page, and decide. A firm that runs this sequence usually finds one candidate that clears the rule and one that only sounds like it does.
FAQ: AI Document Review for Law Firms
Can we use a general chatbot for document review in a law firm?
Only where the deployment keeps documents inside an environment you control. A hosted chat service processes your files on the provider’s servers, so it fails a literal no-egress rule however good the answers are. For matter work under a strict policy, treat it as unusable and give your team a sanctioned path instead.
Does “we never train on your data” mean the data stays with us?
No. It answers one of the four questions and leaves three open. Storage, residency, and derived artifacts are separate commitments, and each needs its own line in the contract.
How do we verify a no-egress claim?
Block outbound traffic at the firewall and run a real review task. Then ask for three things: the deployment diagram with hostnames, the list of outbound calls the product can make, and the switch that disables telemetry. A vendor who can produce all three is telling the truth about their architecture.
Is on-premises AI cheaper than an eDiscovery licence?
It depends on volume, and on how you count staff time. On-premises moves cost from a per-gigabyte licence into GPU hardware, serving infrastructure, and the engineer who runs it. Ask both vendors for a twelve-month total cost that includes hardware, storage, integration and internal hours, then compare the two numbers.
Does using AI waive privilege?
That is a legal question for your own counsel, because the answer depends on jurisdiction and on facts we cannot see from here. ABA Formal Opinion 512 sets out the confidentiality duty around AI use, and it asks for reasonable efforts to prevent unauthorised disclosure. Build the record that shows those efforts.
What encryption should we require?
At minimum, encryption in transit and at rest, with keys you control where the deployment allows it. The OpenCraft AI legal page states AES-256 across files, chats and documents, and end-to-end encryption. Ask any vendor for the key management detail behind the claim.
Can the AI remember a matter and then forget it?
Yes, where the product supports deletion by workspace. Ocai’s memory is what makes repeat work faster, and it is also stored content. Keep one workspace per matter, delete at close, and confirm the deletion in writing before you archive the file.
What hardware does on-premises document review need?
It scales with model size and page volume, so there is no single answer. Smaller open-weight models run on modest hardware, and larger ones need serious GPU capacity. The self-hosting guide gives sizing bands by parameter count, so start there and test on the closed matter from the 30-day plan.
Making the AI Document Review Choice Deliberately
Two firms can run the same tool and end up with different risk. One wrote down what “no data leaving” meant. The other assumed.
Decide which of the four meanings binds you. Pick the deployment shape that satisfies it. Run the firewall test before the pilot. Keep the citations. Keep the human sign-off. If you want to see the work before committing to a deployment, start with free credits on the hosted sandbox and run the same five tasks. For a firm that needs the application and the model on its own infrastructure, the proof of concept is the route.
Blanket bans and unexamined defaults both cost you, and the analysis of what blanket bans actually do covers the second failure mode in detail.
References
- American Bar Association, Standing Committee on Ethics and Professional Responsibility. “Formal Opinion 512: Generative Artificial Intelligence Tools.” 29 July 2024. Primary text. Summarised and relied on in sources 2 to 4 below.
- University of North Carolina School of Law Library. “ABA Formal Opinion 512: The Paradigm for Generative AI in Legal Practice.” UNC School of Law Library, 2025. https://library.law.unc.edu/2025/02/aba-formal-opinion-512-the-paradigm-for-generative-ai-in-legal-practice/
- Maryland State Bar Association. “The ABA’s Stance on AI: Formal Opinion 512.” MSBA, 7 Aug. 2024. https://www.msba.org/site/site/content/News-and-Publications/News/General-News/The%20_ABAs_Stance_on_AI_Formal_Opinion_512.aspx
- National Conference of Bar Examiners. “Generative Artificial Intelligence Tools: ABA Formal Opinion 512.” The Bar Examiner, Fall 2024. https://thebarexaminer.ncbex.org/article/fall-2024/generative-artificial-intelligence-tools/
- OpenCraft AI. “Enterprise AI: Runs on Your Infrastructure.” Accessed 18 Nov. 2026. https://opencraftai.com/enterprise
- OpenCraft AI. “Legal AI: Best Associate, Every Matter.” Accessed 18 Nov. 2026. https://opencraftai.com/solutions/lawyers
- OpenCraft AI. “AI Copilot Pricing: Pay As You Go.” Accessed 18 Nov. 2026. https://opencraftai.com/pricing
- OpenCraft AI. “Self-Hosted AI Assistant: Run AI on Your Own Infrastructure.” 18 Nov. 2026. https://resources.opencraftai.com/blog/self-hosted-ai-assistant/
- OpenCraft AI. “RAG vs Agent Memory: Why Retrieval Was Never the Hard Part.” 6 Oct. 2026. https://resources.opencraftai.com/blog/rag-vs-agent-memory/