What the case study says

OpenAI’s write-up on Legora describes a legal platform that uses OpenAI models, ChatGPT and Codex to speed up legal workflows, extend contract analysis, and free lawyers to spend time on higher-value work. Published 2026-10-28.

That is the whole of the public claim. There are no accuracy numbers, no cycle-time measurements, and no customer counts in the summary. So the useful question is not “how fast is it” - it is “what would have to be true for this to be safe to use”.

Contract review is a citation problem

A contract summary feels impressive and is nearly worthless on its own. What a lawyer needs is not “this agreement has a 30-day termination clause” but “clause 7.2, second sentence, says 30 days, and here is the exception in 7.4”.

That distinction turns the engineering problem into three parts, only one of which is the model:

  1. Retrieval. Find the passages that could answer the question, including the definitions section that changes what a word means.
  2. Attribution. Make every generated sentence carry a pointer to the passage it came from - and refuse to emit a sentence that cannot be pointed at anything.
  3. Review. Show the pointer next to the sentence, so a human can verify in seconds instead of re-reading the contract.

A model that is 5% more accurate but does not attribute is worse than a weaker model that cites. In legal work, unverifiable output has negative value, because checking it costs more than doing the work yourself.

Why “freeing lawyers for higher-value work” is the honest framing

The phrase is doing real work here. It is not claiming the model replaces review; it is claiming the model absorbs the mechanical layer - the first pass over 200 contracts to find the ones that deviate from the template, the extraction of renewal dates, the diff between two redlines.

That is the same pattern as every credible agent deployment: automate the volume, escalate the exceptions. The lawyer’s judgement moves from reading everything to reading the flagged subset, and the product’s value is measured by whether the flagged subset is actually the right one.

Where Codex fits

Case studies list tools together and readers assume one story. Here there are two:

  • Models and ChatGPT are the product surface - what the legal team uses.
  • Codex is the engineering surface - how the product gets built and maintained.

Conflating them leads to the wrong conclusion (“the AI writes legal advice and the AI writes the code, so both are automated”). Separating them is also the more useful lesson: an AI-native company uses models on both sides of the line, in the product and in the toolchain that produces it.

What to copy if you work with documents on a desktop

Most people handling contracts do not have a legal platform; they have a folder. The transferable rules are the same at that scale:

  • Never accept a summary without a pointer. If the tool cannot name the file and the location, treat the sentence as unverified.
  • Diff against the template, not against memory. The valuable output is “this clause differs from your standard”, which is a comparison, not a generation.
  • Keep the original untouched. Review should read the file, not rewrite it. If the tool edits in place, you have lost your reference copy.
  • Log which documents were opened. In legal work, the audit trail is often as important as the answer.

MOOGH is built around those constraints for ordinary documents: it shows which files it intends to open before acting, keeps the original intact unless you approve a write, and records what was touched. Download MOOGH if you want a document workflow where the evidence is visible.

FAQ

No. It means the mechanical layer - extraction, comparison, first-pass review - is being absorbed, while judgement and accountability stay with the lawyer. The case study’s own framing says “free lawyers to focus on higher-value work”.

What accuracy would make this trustworthy?

The question is the wrong shape. Trustworthy legal output is not a percentage; it is a property of the workflow: every claim traceable, every exception surfaced, and a human who can verify in seconds. A 99% accurate system with no citations is still unusable.