Can You Use ChatGPT or Claude to Review Business Contracts? An Honest Guide for SMBs
You have a service agreement sitting in your inbox and your lawyer charges $400 an hour. Before you forward it, you open ChatGPT or Claude and paste the contract in. The model comes back with a tidy summary and a few flagged clauses. But is that actually useful, or are you just getting confident-sounding text that misses the clause that will cost you six months of revenue?
The direct answer: ChatGPT and Claude are genuinely useful for a first-pass contract review, but they have specific failure modes that are invisible unless you know what to look for. Used correctly, they can save you time and sharpen the questions you bring to a lawyer. Used naively, they give you false confidence.
What do these models actually do well on contracts?
Plain-language translation is where both models shine. If a contract is dense with legalese, asking Claude to "explain each section in one plain sentence, written for a business owner who is not a lawyer" produces genuinely readable output. This is reliable and repeatable. You are not getting legal advice; you are getting comprehension, which is valuable on its own.
Flagging common problematic structures is also solid. Paste a vendor agreement and ask: "Identify any auto-renewal clauses, unilateral change rights, limitation-of-liability caps, and indemnification obligations that fall on my company." Both models will find these when they are present. They will also catch obvious asymmetries, such as a termination clause that lets the other party exit with two weeks' notice but requires you to give ninety days. These are the kinds of things a rushed owner misses when skimming.
Comparison is another underused application. If you have a previous contract with a vendor and you are reviewing a renewal, paste both and ask the model to list every clause that changed. This takes a human twenty minutes of careful side-by-side reading. The model does it in seconds and is accurate on structural changes, though it can misread heavily reformatted documents.
Where do ChatGPT and Claude quietly fail on contract review?
The first and most important failure: the model does not know your jurisdiction's law. A limitation-of-liability clause may be enforceable in Texas and unenforceable as written in California. The model will describe the clause but is unlikely to tell you it would not hold up in San Diego County Superior Court. It does not have reliable knowledge of how local courts have interpreted specific contract language, and it will not consistently flag when state law overrides a contractual term. This is the gap that costs money.
The second failure is context blindness. The model knows nothing about your relationship with the other party, your industry's norms, or what you actually negotiated verbally. It cannot tell you that the payment terms in section 4 conflict with the side letter you signed last year. It also cannot tell you that in your industry, a 30-day cure period is standard and anything shorter signals a counterparty trying to set up an exit. That kind of pattern knowledge lives in a lawyer's head or in your own institutional memory, not in a general-purpose language model.
The third failure is the one most owners discover painfully: missing what is not there. Models are reasonably good at analyzing text that exists. They are unreliable at noticing absent protections. A contract that says nothing about data ownership, intellectual property created during the engagement, or dispute resolution is a contract with dangerous gaps. When you ask "does this contract protect me?" the model tends to evaluate what is present rather than inventory what is missing. Ask explicitly: "List standard protections that a business owner in a services agreement would typically want, and tell me which ones are absent from this document." That prompt gets you much closer to useful output.
A worked approach for the Friday afternoon contract that just landed
Here is a sequence that makes both tools genuinely useful without crossing into relying on them for legal judgment. First, paste the full document and ask for a plain-English section-by-section summary. Read it. Second, paste the document again and ask the model to flag: auto-renewal terms, indemnification obligations, limitation-of-liability language, termination notice requirements, and any clause granting the other party unilateral rights to change terms or pricing. Third, ask the model to list standard protections that are absent. Fourth, use that output to write a short list of specific questions for your lawyer. You have now turned a $400 general review conversation into a targeted 30-minute call. That is the actual value.
One thing to avoid: asking the model "is this contract fair?" or "should I sign this?" Those prompts produce plausible-sounding answers that are not grounded in your situation. The model will hedge, but not always enough, and a business owner who wants reassurance can easily read confirmation into a hedged answer.
Getting this kind of consistent, structured output from ChatGPT or Claude requires standardized prompts that your whole team uses the same way every time. The DSE Group AI enablement program is built around exactly this: engineering the prompts, context, and workflows once so that every person in your company gets reliable output instead of random quality depending on how they happened to word the question that day.
If you want to talk through how to build contract review, proposal drafting, or other legal-adjacent workflows into a system your team can actually trust, reach out to the DSE Group team. We work with businesses across Southern California to turn one-off AI experiments into repeatable processes that hold up under real workloads.