1. What contract review software does
A contract review tool reads an agreement and turns it into structured findings: the clauses it contains, what each one commits you to, which of them carry risk, and what has to happen by when. The useful output is not a summary — a summary is the part a careful reader can do themselves in ten minutes. It is the part that does not fit in a person’s head: every renewal date across forty agreements, every liability cap that is lower than the contract value, every notice window that closes next month.
PactSage produces four things from an upload. A clause inventory, covering the nine to twelve clause types that belong to that kind of agreement and explicitly recording the ones that are absent. A risk rating per clause on four levels, each carrying the verbatim text it was drawn from. An obligation list — who owes what, in which direction, by which date, on what recurrence. And a set of dated events: renewals, expiries and the notice windows that precede them.
2. Why careful readers still get caught
Contracts rarely fail at signature. They fail eighteen months later, on a renewal nobody diarised, or during a dispute where the liability cap turns out to be a fraction of the exposure. Three things cause most of it.
Length. A master services agreement with schedules runs to forty pages of which perhaps two matter to you, and which two depends on what goes wrong. Nobody reads all of it twice.
Absence. The dangerous term is often the one that was never drafted. A vendor agreement with no liability cap reads perfectly smoothly, because there is nothing on the page to catch your eye.
Time. A ninety-day notice period on an auto-renewing contract is a date you have to act on, not a fact you have to know. Filed away, it is functionally invisible until it has passed.
3. How PactSage reads a contract
Extraction runs in a fixed order, and each step can refuse rather than guess.
Read the document. A digital PDF is read from its text layer. A scan is detected from how much of each page its images cover — not from how much text it appears to carry — and re-read page by page with optical character recognition. That distinction matters more than it sounds: most scanners embed their own OCR, and those engines are Latin-script only, so a Devanagari document scanned that way arrives carrying long, confident Latin nonsense that passes every length check.
Classify before extracting. The document type is identified first, and if it is not one of the supported types the run stops there and says so. Forcing a sale deed through a services-agreement taxonomy produces findings that look authoritative and describe nothing.
Walk that type’s clause set. Each contract type has its own list rather than one flat checklist, so a lease is never asked about intellectual property ownership and a statement of work is never asked about the things a settlement agreement contains.
Verify every quote against the source. Each extracted clause is re-found in the document text. A quote that cannot be located is kept and flagged rather than discarded — a plausible finding with a bad citation is worth showing with a warning, and must never be allowed to claim a citation it does not have.
4. The clause that is not there
Missing-clause detection is the part that is hard to do by reading. The extraction walks the full clause set for the document’s type and records an explicit finding when a clause is absent, with its own risk level — because the risk being described is the risk of the absence.
Why absence needs a rating of its own. A vendor agreement with no limitation of liability is not a neutral document with one fewer paragraph. It is an uncapped one. Reporting it as “not found” without a severity understates it in exactly the direction that costs money.
5. Citations, not assertions
Every claim PactSage makes about your document carries a numbered marker that resolves to the passage it came from, and you can open the passage. Markers that resolve to nothing are stripped before you see them, and the remainder renumbered — a citation that cannot be opened is a claim of evidence that does not exist, which is worse than no citation at all.
Two consequences follow that are worth knowing about, because both are deliberate. Silence in a contract is never read as a rule: asked whether a landlord may raise the rent mid-term when the agreement says nothing about it, the honest answer is that the document does not address it, not that it is prohibited. And when a question is about general law rather than about your file, the answer is labelled as general and cites public sources you can open, rather than borrowing the authority of a document citation.
Extraction accuracy is measured per clause type against reviewed contracts, with a precision floor of 0.85 that a build has to clear. A quote that cannot be verified against the source counts as wrong, and a missing-clause call is scored in both directions.
6. Contract types and clauses supported
Eight document types are supported end to end: master services agreements, statements of work, non-disclosure agreements, vendor agreements, employment contracts, leases and residential rent agreements, sale deeds, and settlement agreements. Affidavits, powers of attorney and loan agreements are recognised by the classifier but declined rather than approximated.
A master services agreement, for example, is checked for:
- Payment terms
- Intellectual property ownership
- Limitation of liability
- Indemnity
- Termination rights
- Confidentiality
- Non-compete
- Automatic renewal
- Dispute resolution
- Warranty
- Force majeure
- Assignment
7. Indian contracts and Indian law
Most contract tooling is built where the training data is, which means it answers an unqualified question with at-will employment, US state statutes and dollar thresholds. On an Indian rent agreement that is not a stylistic slip — it is the wrong law delivered with the confidence of the right law.
PactSage resolves jurisdiction in a fixed order. A governing-law or dispute-resolution clause in the document wins outright, including when it points somewhere other than India; applying Indian law to a contract expressly under Singapore law is the same mistake in the other direction. Where the document is silent, jurisdiction is inferred from what the document shows — currency, addresses, stamp paper and sub-registrar entries, GSTIN, PAN, CIN and IFSC numbers, PIN codes, the statutes cited, the script. Only then is India assumed, and the assumption is stated rather than hidden. Both the clause found and the conclusion drawn are stored, so a reviewer can disagree with either.
Documents in Hindi and Marathi are read and answered in their own language rather than translated, and quoted clauses stay verbatim in the document’s script — a translated quote can never be re-found in the source, which would mean every clause in a Hindi agreement arriving marked unverified.
8. What this does not do
PactSage is not a law firm and its output is not legal advice. It does not tell you whether to sign, it does not negotiate, and it does not replace a lawyer on anything consequential. What it does is make sure that when you do sit down with one, you already know which three clauses to spend the hour on.
It also declines rather than improvises. A document type outside the supported set is refused. A question its retrieval cannot ground in your documents is answered from general knowledge and labelled as such, with anything it could not verify marked in the text. Those refusals are the reason the rest is worth reading.
Questions people ask about AI contract review
10. Reviewing your first contract
The free tier needs no card. Upload one agreement you already know well — a signed contract whose terms you can check the output against is a far better test than one you have never read, because you can catch it being wrong.
Related reading: what to check in a rent agreement before you sign. The Android app is on the download page, and how uploaded documents are handled is set out in the privacy policy.