What AI contract extraction gives you and what it doesn't

When you upload a contract to a review tool, a few fields fill in by themselves. Start date, end date, the notice you have to give to get out. That happens before anyone opens the file, and if you run legal ops for a few hundred vendor agreements a month it looks like the renewal calendar nobody had time to build.

It is worth knowing exactly what those fields hold before you build anything on them.

Extraction finds the words

Contract extraction is software reading a document and pulling out the pieces a person would otherwise search for: parties, dates, notice language, the governing law line. The research benchmark for this kind of work, the Contract Understanding Atticus Dataset, describes the task as highlighting the parts of a contract a human needs to review. Its labeling handbook keeps effective date, expiration date, renewal term and the notice period to terminate renewal as separate items, each one a passage the annotator marks in the text.

Extraction finds where the thing is said and shows you the sentence. Turning that sentence into a number your spreadsheet can sort is a separate step, and extraction on its own does not take it.

What the notice column actually holds

In goHeather, uploaded contracts land in a list with columns for Name, Start Date, End Date, Notice to Terminate Date and Date created. Those fields fill at upload with no action from you. Look down the notice column on a real set of contracts and you see the contract's own language: "no less than 7 days' written notice", "7 days' written notice", "thirty (30) days' written notice", "30 days".

That is four contracts and four different ways of writing roughly two ideas. A reviewer can read each one and see what it says. What you cannot do is sort the list by notice period or filter for everything under 30 days, because the column is text, not a number. There is a reason for showing the raw wording. A field that just said "30" would hide the "no less than" and the "written" that a reviewer needs to see. But it means the portfolio-wide renewal view you want is still a step away, and any tool that shows you a clean date column is doing a translation somewhere that you should ask about.

Why the translation is the risky part

Contract dates often depend on each other. The notice deadline is usually counted from the end of the current term, and in an auto-renewing contract the current term itself depends on what happened at the last renewal. The document cannot tell you what happened, so a field labelled "notice date" is a calculation built on assumptions about events outside the file.

Here is where it goes wrong. An amendment extends the term by a year and sits in a different folder from the master. The extracted end date on the master is now wrong, and every notice date computed from it is wrong with it. The wording on the page is still accurate, because the page has not changed, but the number computed from it is wrong. Keeping the words in front of you means that when the calculation goes stale, you can still see what it was built from.

Where the full extraction shows up

The list view is the short version. The fuller extraction comes with a review, in a memo called the Deal Abstract that pulls every fact of the deal into sections: parties, agreement type, effective date, term and renewal, fees, payment terms, termination summary, governing law and venue, notice requirements, and the operational obligations. On a real consulting agreement it ran to about ninety bullets, down to the hourly rate, the 21-day payment term and the named person who approves invoices. It also noted that venue was not specified and that the notice attention lines were blank, which is the kind of gap a date field would silently hide.

There are two limits. The abstract reads what you gave it, so an amendment you did not upload is not in it, though you can add it as a reference document when you run the review. And if the document is edited after the review, the memo carries a banner saying it is stale and the review needs re-running. A spreadsheet column does not tell you that.

The same file problem shows up earlier in the pipeline. A scanned exhibit with no text layer gives extraction nothing to read at all.

What to expect from it

Extraction is useful because it cuts the searching. A reviewer who wants the notice language gets it in one line instead of paging through a sixty-page file. A team building a renewal calendar gets the raw material for one, in the contract's own words, with the judgment about what each phrase means and which amendment governs left to a person.

Run a batch of your own vendor agreements through the AI contract review workflow and look at what the list view shows before you decide what to build on top of it.

This is legal information, not legal advice; consult a lawyer for legal advice.