Use cases
AI for legal work
By Ofrai Editorial (Use Case Research)
Legal work is mostly reading a document carefully, then reading it again with a different question in mind. Ofrai puts Claude Opus 5, Claude Sonnet 5 and GPT next to each other so the first pass and the second pass can run on the same credit balance.
Chat and video use credits from your Ofrai balance.
Contract review
Paste a clause and ask for the obligations it creates, the obligations it does not, and the definitions it depends on. The model should quote the sentence it is talking about before it draws a conclusion. A review that does not point back at the source is a review you cannot check.
Redrafting a clause
Give the model the existing wording and the outcome you want, not the wording you want. Ask for three variants at different levels of obligation and pick between them, rather than accepting the first rewrite. Legal drafting is mostly a question of how much you are willing to enforce.
Contract summaries
Ask for a summary with a fixed shape: parties, term, renewal, liability cap, termination triggers, and anything that auto-renews. A fixed shape is what makes a summary comparable across a folder of contracts, and it is the difference between a summary and a paraphrase.
When to bring in a lawyer
Use the model to find the questions worth asking, not to answer them. Anything with regulatory exposure, jurisdiction-specific enforceability, or a number large enough to matter needs a qualified reviewer. The model is good at making you ask the right question; it is not a substitute for someone who can be held responsible for the answer.
AI for legal work: a week of contract turnaround
The pattern that works is a fixed intake rather than ad-hoc pasting. Day one, set one review template with a fixed section list and one instruction that every finding must cite the sentence it came from; that template becomes the thing you reuse, and it is where most of the savings live. Day two, run the template over a small batch, maybe ten agreements of a similar shape, and read every output line by line to find where the model overreaches.
Day three, write down the failure modes you actually hit rather than the ones you imagine, and add them to the template as explicit checks. Day four, build the second template for the other job, which is usually comparison: two versions of the same paper, side by side, with differences listed rather than summarised. Day five, keep the two templates in one file with the names of the documents they work on, so the next person does not rebuild them.
By Friday the batch that took a day and a half is taking an hour to mark up, and the marks are the same because the template is the same. The gain is not speed, it is that two people reviewing the same contract now leave the same notes.
Prompt patterns for clause work
Three patterns cover most of what legal teams ask for. The extraction pattern: 'List every obligation this section creates for {party}, quoting the sentence for each one, and mark any term that depends on a definition elsewhere in the document.' The comparison pattern: 'This is version one. Here is version two. List every change that moves risk, and leave wording changes that do not out of the answer.'
The rewrite pattern: 'Rewrite the clause below so that {specific outcome}, keep the defined terms exactly as they are, and flag any word you removed that carried an obligation.' The third instruction matters more than it looks: defined terms are load-bearing, and a rewrite that quietly swaps a defined term for a synonym is how a definition stops being binding. Keep the variable slots visible so a template can be reused, and do not paraphrase between runs or you lose the ability to diff two outputs.
Choosing between Opus, Sonnet and GPT for legal work
Opus is the deep pass. Give it the document, the background you have, and the question you are actually stuck on, and let it reason before it answers; it is the one to use when a clause is genuinely ambiguous and you need the ambiguity explained rather than resolved. Sonnet is the daily pass: routine clauses, definitions, renewal and notice mechanics, the work you do several times a day and do not want to think about.
GPT is the structured pass, and it earns its place when the output has to be table-shaped, because asking the same three questions across thirty documents and getting thirty comparable tables is a different job from reading one document well. The practical rule is to run Sonnet first on everything, escalate only the flagged items to Opus, and reserve GPT for the moment the output needs to be a table your other tools can read.
Running the expensive model on the easy documents is the most common way this budget gets wasted.
Where AI contract review goes wrong
Four failures account for nearly everything. The first is invented obligations: the model produces a term that sounds like a clause and is not in the document, and the only defence is requiring a quotation for every finding. The second is numeric drift: an amount, a notice period or a date gets altered in a paraphrase, which is why a summary that rewrites numbers instead of copying them is not safe to circulate.
The third is tone inflation, where 'the parties will consider in good faith' becomes a firm obligation and the legal meaning changes while the reading gets easier. The fourth is broken cross-references: a summary that drops one clause leaves every 'as set out in clause five' pointing at nothing. The way to catch all four is the same: make the model quote before it concludes, make the human verify every number, and keep the source open next to the output the whole time.
AI for legal work: credits per contract, not per seat
Legal tooling is usually priced per named user, which means a read-only colleague still costs a full licence. On Ofrai the cost is the work, not the seat. A short clause review and a long agreement review draw different amounts of credit because they consume different amounts of model, and the balance is shared with everything else you run in the studio. Two practical consequences follow.
First, adding a reviewer to a review batch costs only what that reviewer actually reads, so the expensive model can be reserved for the documents that need it. Second, the marginal cost of the second read-through of a document is a fraction of the first, which means the cheapest legal win available is usually reading a document twice rather than reading it once faster.
What to take from the model, and what to sign off yourself
Split the output into two piles and keep them separate in the document. Into the first pile go the questions: which defined terms are used without being defined, which obligations have no deadline, which clauses auto-renew without notice, which liability caps do not cover the biggest exposure on the page. Into the second pile go the answers, and the answers stay empty until a person with the authority to accept risk has filled them in.
This split is worth the extra minute it costs because it makes the handoff unambiguous. A reviewer knows from the first page alone whether the document is asking them to check an argument or to sign a position, and the model's output never has to be read as if it were advice.
Who should build this in, and who should not
Build this in if your documents repeat. If the same clause shapes come back every month, if several people review the same kind of paper, or if the bottleneck is finding the right question inside a long document, then a fixed template plus a fast model will remove real hours and the risk stays with the reviewer.
Stay out of it if your work is one-off and high-stakes: a bespoke transaction where the whole point is that no two clauses are alike, or a matter where a single phrase carries a specific legal consequence that the model will not know to flag. There is also no point running a model over a document you already know well. The case for it is volume and repetition, and if you do not have that, the setup cost is not worth it.
A first afternoon with AI for legal work
Start with one document you already know, because you need a baseline you trust. Paste in the agreement you reviewed last month, ask for the obligation extraction with a quotation for every finding, and read the output against your own notes. You are looking for two things: the places it found something you missed, which is the case for keeping it, and the places it found something that is not there, which is the case for tightening the template.
Once the extraction template holds up on a document whose answers you know, add the comparison template and run it on a redline where you already have the change list. By the end of the afternoon you should have two templates, a list of the failure modes you actually hit, and a clear sense of which parts of your review this can carry. That last part is the deliverable, and it is what tells you whether the second week is worth it.
Open studio
FAQ
- Does this replace a lawyer?
- No, and the page is written so it cannot be read that way. The models are good at finding the question you should be asking inside a long document, and bad at being accountable for the answer. Use the output as a set of issues to raise. Anything with regulatory exposure, jurisdiction-specific enforceability, or real money attached needs a qualified reviewer before it goes anywhere.
- Can I paste a whole contract at once?
- You can, and for a first pass it is usually worth it. For anything you intend to act on, split it: review the definitions, the commercial terms, and the risk provisions separately, because a single pass over a long document will weight the early pages more heavily than the last ones. The chat-with-PDF tool is a better fit when the document is a PDF with pages you need to cite back to.
- How much does a review cost?
- It depends on the document length and which model you point it at, and the live price is shown on the model pages in the studio. The pattern that keeps cost predictable is running the fast model over the whole batch and escalating only the flagged items to the deeper one, rather than sending everything to the strongest model.
- How do I start on Ofrai?
- Open Text Studio, paste a clause you already understand, and ask for obligations with a quotation for each. Keep the answer, then paste the same clause into a second request asking for three redraft options. That single document gets you the two templates you will actually reuse, and it is short enough to finish before you close the tab.
- What is AI for legal work on Ofrai?
- One workspace holding Claude Opus 5, Claude Sonnet 5 and GPT, plus a chat-with-PDF tool, on a single credit balance. You draft the review template once, run it over a batch, and the model you use for each document is a choice you make rather than a contract you sign.