Used well, an AI pass gives you a map of the contract. One plain-English line per section, the obligations sorted by who owes them, every date, dollar figure and notice period pulled into a single list, and a note on what looks unusual for an agreement of that kind. The map is worth having before you read the document properly, and it is not a replacement for reading it. Every item has to be checked against the clause it came from, because the summary is the index and the contract is the text. Done that way it is one of the better things a small business can point a general assistant at.
Most small businesses sign things nobody read. A twenty-page service agreement arrives as a PDF attached to a friendly email, the work starts on Monday, and the honest options are skim it, pay someone to read it, or sign it. An AI pass adds a fourth: know what is in the document before Monday, then decide which of the other three you still need. Summarizing long documents is one of the everyday tasks AI can take over, and a contract is that task with consequences attached. None of this is legal advice.
What the summary is good for
It does four jobs. It maps the document, so you know what section 9 covers without reading section 9 yet. It sorts obligations by party, which is the question you have: what have you promised, and what have they. It extracts every date, amount, rate, notice period and renewal trigger into one list, which is dull mechanical work and the sort a machine is good at. And it compares the document loosely against the ordinary shape of that kind of agreement, which is how you end up with a written list of questions to put to the other side.
That last output is the one owners underrate. It gives you specific questions about clauses 7, 11 and 14 to put on the call.
It is not a lawyer. A lawyer tells you which of the unusual terms is dangerous in your state and your trade, and carries insurance for being wrong. A summary tells you the term is there. It is also not a guarantee that nothing was missed. A model that reads twenty pages and finds nothing odd looks the same as one that skipped clause 14.
Whose information is in the file
A contract carries the other party's information as well as yours. Their prices, their terms, and very often a confidentiality clause covering those. ABA Formal Opinion 512, issued on 29 July 2024, tells lawyers that before they put client information into a generative AI tool they must evaluate the risk that it will be disclosed to or accessed by others outside the firm, which is a useful reminder to everybody else that the file they are about to paste is somebody else's confidential document too. Lawyers are bound by this, and we go through their version of it in our post on lawyers and ChatGPT.
So read your own confidentiality clause first. Some prohibit disclosure to third parties without consent, and a cloud AI service is a third party until the vendor says otherwise in writing.
Choose the tier that does not train on what you upload
The tier matters more than the brand. The same company treats a business account and a free consumer account differently.
- Anthropic says that by default it will not use inputs or outputs from its commercial products, including Claude for Work and the Anthropic API, to train its models, as of September 2026.
- Data sent to the OpenAI API is not used to train or improve OpenAI models unless you explicitly opt in to share it, as of September 2026.
- On consumer Claude, chats are used for training only if you have chosen to allow it, through a control called Model Improvement under Privacy Settings, as of September 2026.
- Google says a subset of Gemini app chats are reviewed by human reviewers, and tells users not to enter confidential information they would not want a reviewer to see or Google to use to improve its services, as of September 2026.
Consumer plans differ from one another and the terms move, so open the tool's own data controls and read what they say about your account before a contract goes near it. If you have not settled which work can go into a cloud tool at all, which parts of your work can go in ChatGPT works through it.
Strip what you have to, and paste the rest
Replace the party names with roles, so the file reads as us, the client and the guarantor. Take out staff or customer personal details sitting in an annex. Leave the attachments that are not the agreement on your own disk: price books, account lists, personnel schedules.
Then stop. If you replace every amount with a placeholder you have deleted the thing you wanted extracted, and if you cut clauses to make the document shorter you get a confident summary of a document that does not exist. The numbers and the operative clauses go in whole.
Exhibit 1
Swapping the names for roles costs a minute and removes most of what you were worried about.
Four passes, in this order
Run them as four separate prompts in one conversation, so each pass still has the document in front of it, and read each output before you send the next.
Map it first. Summarize the agreement in one line per section, in the order the sections appear, with the section number on each line. You are after the shape of the document, and you will usually find a section you did not know was in there.
Then obligations, party by party. List every obligation, grouped by the party who owes it, with the clause number. The imbalance shows up here, because the two lists are rarely the same length.
Then every date and figure. List every date, deadline, dollar amount, rate, notice period and renewal trigger, each with its clause number. This is the pass a person does worst, because the numbers are scattered across twenty pages and three schedules.
Then what is unusual, and what is missing. Ask how the document compares with an ordinary agreement of the same type, then ask separately what would normally be present and is absent. The absences are what to watch for. No cap on liability, no termination for convenience, no notice before a price rise, no definition of what counts as acceptance.
Ask for clause numbers in every pass. A summary without them cannot be checked, and an unverifiable summary of a contract is worse than no summary, because you will act as though you have read the document.
Exhibit 2
Every pass produces claims, and the last rung is where you find out which of them are true.
Check every claim against the clause it cites
Verification decides whether the other four passes were worth running, and it is quick because you asked for clause numbers.
Open each cited clause and read it. Four kinds of error turn up. The clause says something slightly different: thirty days and thirty business days are not the same deadline, and may terminate is not may terminate for cause. The obligation is attached to the wrong party. A defined term is used loosely, so Services in the summary covers more than Services as schedule 1 defines it. Or the cited clause does not carry that meaning.
The fifth failure is silence, and you cannot catch it by checking what the tool wrote. That is what the fourth pass is for, and it is why some of the document still gets read in full: anything about money leaving your business, anything about ending the agreement, anything containing the word indemnify, and every clause the unusual pass flagged. On most agreements that is a small share of the pages.
One mechanical check before any of it. If your copy is a scan, confirm the text came through, because the tool will summarize only the pages it could read.
Three situations where a lawyer is the next step
The first is money. The number to look at is the exposure, the amount you could lose if the arrangement goes badly, which is often far larger than the contract value or the fee. It hides in the liability and payment clauses.
The second is a personal guarantee or an indemnity. A guarantee moves risk off the business and onto you and whatever you own. An indemnity makes you responsible for somebody else's losses, sometimes without a ceiling. Search the file for the words guarantee, indemnify and hold harmless before you do anything else.
The third is anything you cannot get out of. A multi-year term, an automatic renewal with a narrow notice window, an exclusivity clause that quietly closes off other customers. How long the agreement holds you decides what a bad clause inside it costs.
In all three the AI pass still earns its keep, because you walk in with the questions already written and pay for judgment instead of for somebody to read a PDF at their hourly rate. Contracts you send out are a separate job, and mostly a matter of confirming the scope and the exclusions in writing before the work starts.
Working out which tier the business sits on and writing the four prompts down somewhere anyone can run them takes an afternoon. Small setups of that kind are most of what our AI enablement work consists of.