How to Review a Contract for Termination and Renewal Clauses with AI

AI contract clause extraction is at its most useful when it is aimed at a narrow, high-consequence target — and there is no narrower, higher-consequence target than the termination and renewal clauses buried in an agreement you signed three years ago. Miss a 90-day notice window by a week and you have bought yourself another full year of a vendor you already decided to leave.

This is a practical walkthrough: what to look for, how to ask for it, how to turn clause language into actual calendar dates, and — the part most guides skip — how to verify every extracted term against the source text before you rely on it.

The clause family that quietly costs the most money

Most contract review effort goes into negotiation: liability caps, indemnities, SLAs, payment terms. Those matter at signing. But the clauses that cost money after signing are the boring ones about how the agreement ends and how it continues.

The classic failure mode is the evergreen renewal. An agreement auto-renews for successive twelve-month terms unless either party gives written notice at least ninety days before the end of the then-current term. Nobody diarises the date. The ninety-day window closes. The contract renews. You are now locked in for another year at a price you did not want to pay, and the only lever you have left is goodwill.

The second failure mode is the mismatch between what people believe the contract says and what it actually says. "We can leave with thirty days' notice" is a thing that gets repeated in meetings for years. Very often the real clause says thirty days' notice for termination for cause, after a thirty-day cure period, and termination for convenience is not available at all until the initial term expires. Those are wildly different positions, and nobody notices until someone actually reads the clause.

Both failures are extraction problems. The information is in the document. It is just spread across four sections, written in cross-referencing legalese, and possibly amended by a later document nobody filed in the same folder.

What you are actually looking for

Before you ask an AI anything, know your checklist. Extraction is only as good as the list of things you asked for.

Renewal mechanics:

• The initial term — start date, end date, or duration plus a trigger (execution date? go-live? first invoice?).
• Whether renewal is automatic, by mutual written agreement, or absent entirely.
• The length of each renewal term.
• Whether renewals are capped ("up to two successive renewal terms") or perpetual.
• The notice period required to prevent renewal, and who can give it.
• Any price escalation that applies on renewal — a fixed percentage, a CPI link, or "then-current list price," which is effectively uncapped.

Termination rights:

• Termination for convenience: does it exist, for whom, after what point, with how much notice, and does it carry a fee?
• Termination for cause: what counts as cause, and is there a cure period?
• Termination for insolvency, change of control, or breach of a specific obligation.
• What survives termination — confidentiality, IP licences, payment of accrued fees, data-retention or data-deletion duties.
• Wind-down obligations: transition assistance, data export windows, final invoicing.

The mechanics that determine whether you can actually use any of it:

• How notice must be delivered. "Written notice" sometimes means registered post to a named legal address, and an email to your account manager does not count.
• How days are counted — calendar or business, and from the date of sending or the date of receipt.
• Which document controls. An order form, a master services agreement, and two amendments can each say something different about the same term.

That last one is why extraction needs to run across the whole agreement rather than a single PDF.

Step 1: Assemble the complete agreement

Get every document that forms part of the contract into one workspace: the master agreement, the order form or SOW, every amendment, every signed addendum, and any incorporated policy documents the contract references by name.

This matters more than any prompt you will write. An AI that has been given only the master agreement will confidently tell you the notice period is ninety days, and it will be right about the document it can see and wrong about the contract you are actually party to — because Amendment 2 changed it to sixty. Grounded answers are only as complete as the grounding.

If you are working through a stack of agreements for the first time, the general workflow in How to Analyze a Contract with AI (and Verify Every Clause) covers the setup end to end; this post is the deep dive on one clause family.

Step 2: Ask an inventory question first

Do not open with the hard question. Open with one that tells you whether the model can see what you think it can see:

List every section in these documents that deals with the term of the agreement, renewal, or termination. Give the section number and heading for each, and say which document it comes from.

You get back a map. It takes ten seconds to read, and it tells you two important things: whether all your documents actually made it into the workspace, and where the relevant language lives. If the map shows a master agreement and one amendment when you uploaded two amendments, stop and fix that before going further.

This is also the point where an honest tool earns its keep. If the documents genuinely do not contain a termination-for-convenience clause, you want to hear "these documents contain no termination for convenience provision" — not a plausible invented one. A tool that says I don't know when the answer isn't there is doing the single most valuable thing it can do in a contract review.

Step 3: Extract the renewal mechanics

Now ask for the specifics, one clause family at a time. Narrow questions produce answers you can check; sprawling ones produce essays.

What is the initial term of this agreement, when does it start, and when does it end? Quote the language that sets the start date.

Does this agreement renew automatically? If so: how long is each renewal term, how many renewals are permitted, and what notice is required to stop a renewal — how many days, from whom, and delivered how?

Does the price change on renewal? Quote the exact escalation language.

Ask the price question separately even when you think you know the answer. Escalation language often sits in a schedule or an order form rather than the renewal clause itself, and "then-current list price" hides in a subordinate sentence.

Step 4: Extract the termination rights

Same discipline, one right at a time:

Can either party terminate this agreement for convenience? For each party, state whether the right exists, when it becomes available, how much notice is required, and whether any termination fee applies.

What events allow termination for cause, and is there a cure period? Quote the cure language.

Which obligations survive termination of this agreement?

Then ask the question that catches the trap:

How must a notice of termination or non-renewal be delivered to be effective under this agreement? Include the required method, the address or recipient, and whether email is sufficient.

An otherwise perfect analysis is worthless if the notice gets emailed to a sales rep when the contract requires courier delivery to the General Counsel at a named address.

Step 5: Turn clauses into dates

Extraction gives you rules. What you need is a date in a calendar.

Work it backwards. If the initial term ends 31 March 2027, and non-renewal requires ninety days' prior written notice, your true deadline is 31 December 2026. That is the date that goes in the calendar — not the contract end date, which is the date everyone diarises and which is already too late by the time it arrives.

Then subtract the decision time. Someone has to decide to leave, get the decision approved, and get the notice drafted and delivered. If that realistically takes six weeks, the date that actually protects you is mid-November 2026.

Have the AI do the arithmetic, but treat the arithmetic as a draft:

The initial term ends on the date you identified. Based on the notice period in this agreement, what is the last day on which a notice of non-renewal would be effective? Show the calculation and quote the clause that defines how days are counted.

Ask it to show the calculation, because that is what makes the answer checkable. If the clause counts business days and the model counted calendar days, a shown calculation makes the error obvious in five seconds. An unshown one hides it until it matters.

Step 6: Verify every clause before you act

This is the step that separates a usable review from an expensive guess.

For each extracted term, open the citation and read the clause. Not skim — read. You are checking three things: that the quoted language actually says what the summary says it says; that it comes from the controlling document rather than a superseded one; and that no cross-reference in it points somewhere you have not looked. Clauses like "subject to Section 11.4" are load-bearing, and a summary that swallows the cross-reference has lost the qualifier that changes the meaning.

This is the whole argument for citations that open the exact source passage rather than a footnote-shaped gesture at a page number. Verification you can do in three seconds per clause is verification that actually gets done. Verification that means re-opening a 60-page PDF and hunting for a phrase does not get done, and an unverified extraction is just a rumour with better formatting. If the idea of grounding is new to you, what "grounded AI" actually means is the short version.

Verify everything you will act on. For everything else, verify the surprises — any answer that contradicts what you believed about the contract is either an important discovery or an extraction error, and both are worth thirty seconds.

Step 7: Roll it up across the portfolio

One contract is a task. Forty contracts is where this becomes leverage.

Once the per-contract questions are settled, ask for the same fields across every agreement in the workspace and put them in a table: counterparty, document, initial term end, renewal length, notice days, notice deadline, termination-for-convenience availability, escalation. That table is a renewal calendar, and it is the artefact that actually prevents the ninety-day miss. The mechanics of getting clean tabular output — and checking it — are covered in extracting data from PDFs into a table with AI.

Two rules for the roll-up. First, verify the deadline column for every row, even though you will not verify every cell — that column is the one you will act on. Second, keep a column for "not stated." A blank cell is ambiguous; an explicit not stated in these documents tells you the difference between a contract with no convenience right and a contract whose amendment you forgot to upload.

Where AI clause extraction goes wrong

Honest limits, because knowing them is what makes the tool safe to use.

Superseded language. The single most common error. The model finds a clause, quotes it correctly, and it is the version an amendment replaced. Uploading every amendment fixes most of this; checking the effective dates of the documents fixes the rest.

Defined terms. "Term," "Agreement," and "Services" are capitalised for a reason, and their definitions can be narrower than the everyday word. An extraction that treats a defined term as its ordinary meaning can invert a clause.

Day counting. Business days, calendar days, months rather than days, notice effective on receipt rather than dispatch, and the occasional clause that measures from the anniversary of the effective date rather than the term end. This is where showing the calculation pays for itself.

Interacting clauses. A termination right in Section 9 that is qualified by a minimum-commitment clause in a schedule. Extraction is good at finding sections and weaker at reasoning about how three of them combine. That combination is your job, and it is the part of contract review that is genuinely skilled work.

And the obvious one: this is not legal advice, and an extraction is not an opinion. It tells you what the documents say. Whether you can rely on it, and what to do about it, is a question for a lawyer — but a lawyer with an accurate, cited summary in front of them is a much cheaper lawyer than one starting from a box of PDFs.

Can you upload this contract at all?

Frequently the reason none of this gets done is that the agreements are confidential and nobody is willing to paste them into a consumer chatbot — which is a completely reasonable position. Before uploading anything, know where the files are stored, whether they are used to train models, whether deleting really deletes, and who else can see them. FileAI keeps uploaded files private and does not train on them unless you opt in; whatever tool you choose, check the same things before the first upload, and see how the whole workflow fits together for contract analysis.

A repeatable pass

For a single agreement, about forty minutes:

1. Upload the master agreement, order form, and every amendment.
2. Ask the inventory question. Confirm all documents are present.
3. Extract renewal mechanics: term, renewal, notice, escalation.
4. Extract termination rights: convenience, cause, cure, survival, notice method.
5. Compute the true deadline backwards from the term end, and make the model show the calculation.
6. Open the citation for every term you will act on and read the clause.
7. Put the deadline in a calendar with a reminder six weeks earlier.

For a portfolio, run steps 1–5 across the whole set, produce the table, verify the deadline column, and put the whole thing in a shared calendar.

Summary

Termination and renewal clauses are where contracts cost you money after signing, and they are an ideal target for AI clause extraction because the questions are narrow and the answers are checkable. Assemble the complete agreement including amendments, ask one clause family at a time, make the model show its date arithmetic, and open the citation for anything you will act on. Then turn the clauses into calendar entries, because a notice deadline you know about but did not diarise is functionally the same as one you missed.

If you want to try it on a real agreement, you can start free with one document — no card required — and see how it handles your own paperwork.