AI for documents: send a contract, get the risky clauses, then check the invoice against it

The short answer

Send the contract to the agent and say whose side you are on. Ask for the risky clauses with their numbers and the questions to send back. Then ask for a file you can forward: a table with the clause, the current wording, your proposed wording and a reason.

When an invoice arrives, send it to the same chat and ask whether every line matches the contract. The agent has both documents in front of it and checks line by line.

A summary of a document tells you what it says. This guide is about what to do with it. For summaries, quotes and pulling figures out of a PDF, see Twelve things to ask an AI about a PDF.

The test: a bakery, a web studio and two documents

We wrote a sample contract ourselves. A small bakery orders a website from a web studio: eight pages, online ordering, USD 6,400. The contract has eleven sections on two pages, and several clauses lean towards the studio, the way real first drafts often do. A week later the studio sends the first invoice.

The companies are made up. The answers are real: everything below was sent to our test agent on 29 October 2026, on the default model, in one chat.

Step 1. Send the contract and say whose side you are on

“Check this contract” gets a general overview. “We are the Client, what is risky for us” gets a review from your side of the table. Clause numbers make the answer easy to check against the document.

The first reply came in 36 seconds, the questions ten seconds later. The agent ranked the clauses by what they could cost, read them in pairs where they work together (a USD 500 cap plus no liability for delay), and pointed at three things the contract leaves out: the domain, customer data and a bug-fix period after launch. It also caught the smaller items: rates that can change “at any time”, an uncapped hourly rate for extras, a late fee of 0.5% a day and disputes only in the studio's courts.

There were twelve questions in total. Each one names a clause and suggests an answer the studio can accept, which makes a reply more likely than a list of objections.

Step 2. Turn the questions into a file you can send

Questions in a chat are for you. The studio needs a document: what the contract says now, what you propose and why, side by side. Ask for exactly that.

The file took 78 seconds. It is three pages: a short letter, a table of fourteen rows and a closing paragraph that names the two points that matter most. Here is the top of the first page.

Two details make the file usable as it is. The current wording is quoted word for word, so the studio can find each clause in seconds. And the details only you can supply, the jurisdiction and your contacts, are marked in brackets for you to fill in.

Step 3. Check the invoice against the contract

A week later the first invoice arrives. This is a small, common job: compare every line with what you agreed. The agent already has the contract in the chat, so the request takes one sentence.

The answer took 18 seconds. Every mismatch comes with the clause that settles it, and the sum adds up: 900 + 140 + 848 = 1,888. Then the agent went past the question. Our sample had placeholder bank details, and it treated them as a payment risk and told us to confirm the account by phone. It also pointed out that paying the advance on the original text would read as accepting the clauses still under discussion.

Step 4. Write the reply to accounts

The last step is the message itself. Ask for the length and the tone you want.

Sixteen seconds. The email asks for one thing, a corrected invoice, and gives three reasons with clause numbers. The agent also explained what it left out and why, so you decide whether to add it.

With Gmail connected, the same chat can send the email from your own mailbox. The agent shows the draft first and sends after your “yes”.

Where a lawyer comes in

Your chat history, with the documents you send, stays in your agent's isolated container. To answer, the agent sends requests to the model: on the default model they go through DeepSeek's infrastructure in the People's Republic of China, which you confirm during onboarding, and on Pro you can attach your own Claude or Codex, so requests go to that provider under your own account. If a contract has a confidentiality clause, this is what to match it against. How the agent handles data and access is covered in Is an AI agent safe? 8 things to check before you let one into your chats.

What it costs

We read the balance in the agent's status line before and after each message. The contract review cost about 1 , the Word file about 5 , the invoice check about 6  and the email about 1 . That is about 13  for the whole case.

Text work on the default model costs 1  per 20,000 tokens, and a photo or screenshot of a page costs 1.5  to read. The Free plan gives 200  in the first month and 100  every month after, with no card. Pro costs $13.90 a month for 400 , or $1.39 for the first month if you subscribe within 7 days of signing up. All rates are on the pricing page.

Messages worth copying

Create your agent, send it the next contract before you sign, and keep the invoice for the same chat.