CLAUSEWATCH
Questions
What does ClauseWatch actually do?
You upload your contracts. ClauseWatch reads each one, finds the term end, the notice period and the auto-renewal clause, and works out the last day you can give notice. Then it re-reads every one of them every night against that day's date, and tells you the morning something crosses a line.
What happens when it cannot find a renewal date?
It says so, and it names the reason — the clause is absent, the clause is there but states no date, the term runs from a date the document never states, two dates disagree, or the file is a scan with no text layer. It never fills the gap with a guess. A wrong renewal date is worse than no renewal date, because you would act on it.
Do I have to connect anything?
No. Version one is upload and nothing else — no OAuth, no mailbox, no admin consent, no security review. Drag a PDF in and it is read on the spot.
Will it email my vendor without me seeing it?
No. Nothing goes out until you approve it, and an approved notice then sits in the outbox behind a kill window with a Kill button on it. The queue is a list of prepared actions, not a list of things that already happened.
How do I know the date it shows me is really in my contract?
Every date it reports carries the paragraph it was read from and the page number, and you can read that paragraph next to the date. Internally the same thing is a hard check: a value whose own quoted span does not contain it is thrown away rather than shown.
How is this different from a contract management platform?
A CLM is a place to store contracts and route signatures, priced accordingly. ClauseWatch does one job — watch the dates in what you already have — and it works on the drawer of PDFs you have now rather than on the contracts you sign next year.
Does ClauseWatch use a language model to read my contracts?
No. The extraction is deterministic — pattern matching over the document's own text, with the offsets of every match kept. That is a deliberate choice: a model produces a plausible date for a contract that does not state one, and a plausible wrong renewal date is the failure this product is built to make impossible.
What exactly does it look for?
Five clauses: the effective date, the term end, the notice period, whether the agreement renews automatically, and any price uplift. Each one is located in the sentence that governs it, so a date in the payment-terms paragraph is never reported as the term end.
How does it decide which date in a sentence is the end date?
By what the sentence says next to it. A date belongs to the clause whose language it sits closest after — so “commences on January 1 and ends on December 31” yields two different clauses from one sentence. A date that sits in front of every cue in its sentence is not claimed at all, because it almost always belongs to the clause before it.
What happens when two dates disagree?
It reports ambiguity rather than choosing. An amended contract that leaves the old expiry date in the recitals is the common case, and picking one of them silently is how a tool tells you something false with total confidence.
Can it work out a term end from a duration?
Deliberately not. “Twelve months from the Effective Date” with no effective date stated is reported as not found, naming the missing anchor. A date this product prints is a date it can point at in your document — a computed one could not be.
Does it get better over time?
It gets more literal. Every clause pattern it learns is one it can quote; nothing is inferred from other customers' contracts, and nothing about your documents trains anything.
What does version one connect to?
Nothing. Upload is the entire system of record — no OAuth, no mailbox scope, no admin consent, no security review to schedule. Google Drive folder watching is the next thing on the roadmap, and it is opt-in when it lands.
Where do my contracts live?
In your workspace, as extracted text plus the page offsets every quote indexes into. The text is what the nightly pass re-reads; nothing is shared between workspaces and nothing leaves for a third-party model.
Can it send email as us?
Only a notice you approve, and only after the kill window has run out. Version one sends from the product's own address with your organisation named in the body; sending from your domain is a later, opt-in step.
What stops it emailing a counterparty by mistake?
Three things, all in code: nothing is queued without an explicit approval, an approved notice cannot be picked up by the dispatcher until its kill window has elapsed, and a notice with no source span is refused outright rather than sent.
What happens to a scanned contract?
It is stored and reported as having no text layer, for every clause. There is no OCR step, because an OCR misread produces exactly the kind of plausible wrong date this product refuses to show you.
Who can see the contracts in my workspace?
Only members of your organisation. Every query the app runs is scoped by the organisation derived from your own membership row, never from anything the browser sends.
Every answer above is already published on ClauseWatch itself, at https://clausewatch.thecompound.tech/. Nothing on this page is written for it: the answers are read out of this product's own question register by a script, and a build gate re-runs that script and fails the deploy if a word here has drifted from it. The register was last re-checked against its sources on 14 August 2026.