What automated review actually does
Contracts are long, and the failures are boring. Nobody misses the liability cap because it is hidden; they miss it because it is clause 14.3 in a document they read at the end of a long day. Repetition and consistency checking is exactly what automation is better at than a human being.
Four things worth automating
The clause that should be there and is not: a limitation of liability, a data protection appendix, an IP assignment, a termination right.
PDFnoted: Checked against what a contract of this type normally contains, and reported as an absence. Absences are the hardest thing for a human reviewer to notice.
An obligation that binds one party only, an asymmetric termination right, a cap that protects one side.
PDFnoted: Flagged with the clause quoted, so you can decide whether the asymmetry was negotiated or simply went unnoticed.
A defined term used two ways, a cross-reference that points nowhere, a cap that contradicts an indemnity.
PDFnoted: Found by comparing occurrences across the whole document — the task humans are worst at precisely because it requires holding two distant passages in mind at once.
A sanctions exit clause, a data-transfer term, or a governing-law choice that collides with mandatory local rules.
PDFnoted: Checked against 219 regulation instruments and 22,671 provisions across China, the US and the EU, with the matched article cited. This is the part a general-purpose model cannot do, because it needs the actual texts.
Where a human still has to decide
✓ Worth doing
- Use it as the first pass, before counsel sees the document
- Check every finding against the clause it cites
- Use cross-border mode when the deal touches China, the US or the EU
- Export the annotated file so the next reader sees the same findings
- Treat a "no findings" result as a reason to read the unusual clauses anyway
✗ Not worth the risk
- Treat the output as a legal opinion
- Accept a risk score you cannot trace to a clause
- Assume silence means the contract is fine
- Skip the commercial judgement on whether a term is acceptable
- Let the review be the last word on anything you are about to sign
How PDFnoted reviews a contract
- Read the documentThe file is parsed; a scan gets OCR first so there is text to work from.
- Choose the lensGeneral review, or cross-border review against the Chinese, US and EU corpus.
- Produce anchored findingsEach finding names the clause and cites its source — regulation articles where relevant.
- Annotate and exportMark up the document and export it in the same workspace, so the findings travel with the file.
Compared with the usual approaches
| Manual read-through | General-purpose chatbot | PDFnoted | |
|---|---|---|---|
| Findings anchored to clauses | By definition | No | Yes |
| Missing clause detection | Depends on the reader | Sometimes | Yes |
| Cross-border regulation check | Needs counsel | Recollection only | 219 instruments, cited |
| Chinese-language contracts | Needs a reader | Varies | Yes, English report |
| Annotated document out | Manual mark-up | Copy and paste | Export from the workspace |
| File retained afterwards | N/A | Varies | Not retained |
Frequently asked questions
What is AI contract review good at?
The things that are tedious and easy to miss on a first read: a clause that is absent when it should be present, a term that is one-sided, a defined term used inconsistently, a liability cap that contradicts an indemnity, and — in a cross-border deal — clauses that collide with mandatory local rules. It is a structured first pass, not a replacement for judgement.
What is it not good at?
Commercial context. Whether a five per cent cap is acceptable in this deal with this counterparty is a business decision, not a text analysis problem. It also cannot tell you what was negotiated verbally, what the parties meant, or what a court would do. Anything that depends on facts outside the document is outside its reach.
Why does traceability matter?
Because an uncheckable conclusion is unusable. If a tool says a contract is high risk and cannot show which clause drove that, you have to read the whole document anyway to find out — so the tool saved you nothing. Findings that cite the clause can be verified in seconds and dismissed just as fast when they are wrong.
Can it review a scanned or signed contract?
Yes. Scanned pages are OCR'd first to build a text layer, then reviewed. Recognition accuracy becomes part of the pipeline, so figures and party names in a scan deserve the same spot-check you would give any OCR output.
Does it work on Chinese-language contracts?
Yes, and this is where the tool is unusual: a Chinese-language contract can be reviewed with an English-language report, with each finding anchored to the source clause so a reviewer who does not read Chinese can still see what triggered it.
Is this legal advice?
No. It is a structured first pass that tells you where to look and why. Anything you are about to sign, ship or terminate should still go to qualified counsel — the value is that you arrive with the specific clauses already identified.
How is this different from pasting a contract into a chatbot?
Three things: findings are anchored to specific clauses in the source rather than generated as free text; the cross-border mode checks against a real regulation corpus (219 instruments, 22,671 provisions) rather than recollection; and the document stays in a workspace where you can annotate and export the marked-up file instead of copying text back and forth.
What happens to my contract afterwards?
It is not retained. Review findings are produced from the document you uploaded and the file is not kept after the job completes.
Try it on your own file
Upload a document and see what comes back. Free to start, nothing to install.
Review a contract now — free to startKeep going
Sources and further reading: ISO 32000 (PDF specification); Adobe PDF Reference; Schema.org structured data; Aggarwal et al., GEO: Generative Engine Optimization (arXiv:2311.09735). This page is a technical guide, not legal advice.