Having Claude review a contract can save hours of clause-by-clause reading, as long as privacy is settled before the first page gets pasted in. This guide covers the settings to check, how to prepare the document and a multi-pass review method built for professional use.
Key Takeaways
- On Free, Pro and Max plans, chats only feed model training if the improvement setting is on, and incognito chats are excluded either way.
- Anthropic itself advises against sharing confidential business documents, so a contract gets redacted before it goes in.
- A solid review runs in separate passes, and every quote Claude returns gets checked against the original.
Have an AI Sum Up This Article
ChatGPTSettings to Check Before Handing Claude a Contract
The first step is knowing when your conversations are kept and how. On the consumer Free, Pro and Max plans, Anthropic separates two situations based on the model improvement setting, which sits in the account’s privacy settings.
Who can read your chats? By default conversations stay private and are not automatically reviewed. If model improvement is switched on, access is limited to a small number of people involved in training, after the data has been de-linked from the user ID. That alone is a good reason to check this setting before every sensitive file, especially on a shared or newly configured account.
Check the paperwork too. Many commercial agreements include their own confidentiality clause, and some restrict passing the text to third parties or third-party tools. Read that clause first, because the contract you want reviewed may itself set limits on how you are allowed to review it, whatever Claude’s privacy settings say.
With the setting off, your chats are not used to train models, and a deleted chat leaves your history immediately and Anthropic’s back-end storage within 30 days. With it on, new or resumed chats may be kept in a de-identified form for up to five years in training pipelines.
For legal documents the call is easy: keep the setting off on whichever account handles this kind of work. If you turned it on in the past, switching it off stops future use, but data already used in completed training runs stays there.
Two exceptions are worth knowing. A conversation flagged by safety systems can still be analyzed to enforce the usage policy, and the inputs and outputs involved can be kept for up to two years. Thumbs up and thumbs down feedback is stored separately for up to five years and can be used for training once it is de-linked from your identity.
The practical takeaway: never rate a Claude answer inside a conversation that holds a contract. A rating sends that exchange into a retention track that runs far longer than the rest of your chats.
Incognito mode adds another layer. Incognito chats are never used to improve Claude, even with model improvement on, and they do not appear in your history or in Claude’s memory entries. They are kept for 30 days by default, and Enterprise organizations can set a custom retention period.
Incognito has its limits. It only works outside projects, an incognito chat cannot be converted into a regular one or reopened once closed, and the new Claude experience blocks file creation and code execution inside it. For a legal review, those constraints barely matter.
Memory deserves a look too. Claude’s memory now follows users into Cowork, so a confidential detail mentioned in a regular chat can resurface somewhere else. Incognito avoids that spillover, since those chats do not feed memory.
The picture changes for companies. On Team and Enterprise plans, governed by commercial terms, inputs and outputs are not used for training by default. If your organization pays for one of those plans, that is the account to use for a client or supplier contract, even if a personal account is quicker to open on a busy day.
One last point concerns connectors. Raw content from a connected Google Drive or MCP server is excluded from training unless you copy it directly into the conversation. Leaving the file in its storage space and letting Claude read it through a connector limits what travels through the chat itself.
Keep work and personal use apart as well. A contract reviewed from a personal account ends up mixed with your private conversations, projects and memory, while a dedicated work account keeps each matter in a space whose settings you control and can audit.
Prepare the Contract Before You Submit It
Even with the right settings, one rule applies. Anthropic recommends against sharing sensitive financial information, login credentials or confidential business and personal documents. A real commercial agreement often checks several of those boxes at once.
The fix is to redact before sending. Replace company names with “Party A” and “Party B”, signatory names with their job titles, amounts with ranges or “[amount]”, and strip out addresses, registration numbers, bank details and file references.
In practice, a sentence such as “Dupont Industries shall pay 48,500 euros to Martin Consulting no later than March 30” becomes “Party A shall pay [amount] to Party B no later than [date]”. The legal meaning of the obligation survives intact, while nothing left in the text identifies the signatories or the deal.
Redaction keeps most of the value of the review. The risks in an agreement rarely come from the parties’ names. They sit in how the clauses work: durations, exit conditions, liability caps and cross-references between sections.
Keep the structure whole, though. Section numbering, schedules, definitions and internal references let Claude follow the logic of the document and catch a clause that contradicts another one several pages later.
Keep a mapping table on your side that links the neutral labels to the real names, amounts and dates. It stays on your machine and never enters the conversation, and it lets you put the right details back into Claude’s findings when you write your internal memo or negotiation email.
Format matters as well. Claude accepts files of up to 30 MB. For a PDF of 101 to 1,000 pages, only the text is processed and visual elements are not analyzed, and a PDF over 1,000 pages is rejected.
On a long agreement, a price grid or penalty schedule embedded as an image can therefore slip through unread. Retype those elements as text, or split the document into parts of under 100 pages before submitting them.
Schedules deserve the same care as the main body. General terms, price lists, service levels and data protection clauses often live there, and a review that skips them misses part of the real commitments. Submit them alongside the main document, redacted the same way.
Redact in the source file rather than by editing text inside the chat. Working on a copy of the contract in your own editor gives you a clean version you can reuse for every pass, and it guarantees that the unredacted original never sits in the conversation history, even for a moment.
Redaction itself can be automated. If you handle agreements regularly, a local script that swaps names and amounts from a fixed list avoids missing an occurrence deep in the document. Our Claude Code guide to automating repetitive office work shows how to hand that kind of routine to an agent working on your files.
Finally, read the redacted version once more before sending it. A company name left in a header, a footer or the file name itself is enough to undo the whole exercise.
Give the redacted file a neutral name as well, something like services-agreement-draft-2 rather than the client’s name and the deal amount. File names show up in the conversation and in any shared link, and they are the detail people forget most often.
More articles on Horizon
- Claude Cowork and Chat Merge Into One Claude App
- Gemini 3.8 Live Undercuts OpenAI on Voice AI Pricing
- Siri AI on iPhone Leaves EU Users Waiting
Review a Contract With Claude in Four Passes
Asking Claude to “review this contract” in one go produces a general summary that helps but rarely goes far enough. We get better results by splitting the review into four passes, each with a precise instruction and a required output format.
The first pass maps the agreement. Ask Claude to list, section by section, who commits to what, by when and for what in return, laid out as a table. That overview often exposes an imbalance in obligations before the detailed read even starts.
The second pass targets risky clauses. Spell out the list to examine: term and automatic renewal, termination rights, penalties, liability caps, intellectual property, confidentiality, governing law and jurisdiction. For each item, require the section number and the exact wording of the passage.
The exact quote is your safeguard: every passage Claude returns has to appear word for word in the original contract. A model can paraphrase or misattribute a clause, and only that check turns a plausible analysis into a reliable one.
Use the same pass to have Claude pull out the definitions. A term defined early on, such as “Deliverables” or “Customer Data”, can widen or narrow an obligation several sections later. A list of defined terms with their cross-references makes those effects visible at a glance.
The third pass compares the text with your baseline. If you have an in-house template or the terms your company usually accepts, add them to the same redacted conversation and ask for a list of deviations, ranked from the most to the least binding for you.
The fourth pass prepares the negotiation. Ask for the questions to put to the other side on every vague clause, then for alternative wordings to propose. Claude is useful for drafting several versions of the same clause, more or less protective depending on your leverage.
If the contract is written in another language, ask for the findings in your own language but keep every quote in the original wording. A translated quote cannot be checked word for word against the source, and that check is the whole point of the method.
When negotiation produces a new draft, put the old and the new version in the same conversation and ask for every change, clause by clause. That comparison catches quiet edits slipped in between two rounds, such as a shorter deadline or a lower cap, which the eye easily misses after several reads.
A fifth, optional pass helps when the people signing off are not lawyers. Ask Claude for a one-page summary of the contract in plain language: what the company commits to, what it gets, what it risks and which points remain open. Decision-makers read that page, while counsel works from the detailed findings.
Add a caution instruction to every prompt: ask Claude to say clearly when a passage looks ambiguous or when it lacks the information to reach a conclusion. An answer that owns its uncertainty beats a confident analysis built on a misread clause.
Here is a sample instruction for the second pass: “Here is a redacted services agreement. For each of the following clauses, give the section number, quote the exact passage, explain the risk for Party B in two sentences and suggest one question to ask. If a clause is missing from the contract, say so.”
That last line matters. Explicitly asking Claude to flag a missing clause prevents an answer that fills the gap with boilerplate, and a missing liability cap or exit clause is often the most important finding in the whole document.
Work in a single incognito conversation per contract. Context stays consistent from one pass to the next, and nothing mixes with another matter in your history or memory. Copy useful results into your own working document as you go, since a closed incognito chat cannot be reopened.
Keep a short log of what each review of a contract turned up and what was renegotiated. After a few deals, that log shows which clauses keep coming back in your sector, and your second-pass checklist gets sharper every time you update it.
Keep the tool in its place. Claude speeds up reading and negotiation prep, yet signing off on a binding contract remains a job for a lawyer or in-house counsel. An assisted review is how you walk into that meeting with the right questions, and pay for fewer billable hours of groundwork that has already been done on your side of the table.
Once the method is settled after a few agreements, save your four prompts in a separate file, ready to paste. The next review then becomes a routine of a few minutes of setup, with the rest of your time going to the deviations that actually matter for your business and for the relationship that follows.
Follow the story on Horizon.
