Summarising Contracts with AI: What to Check Before You Trust It

You can use an AI assistant to summarise a contract, but you should treat the result as a map of the document, not as a reading of it. The safe routine is to check the tool's data handling first, force the model to quote the clauses it relies on, and then verify the high-risk terms against the original text. This article gives you that routine, with a prompt and a checklist.

Who this is for and what you will have

This is for operations managers, procurement staff, finance teams, founders and administrators who receive supplier agreements, customer terms, NDAs and service contracts and need a quick first read. By the end you will know what to confirm before uploading a document, have a prompt that produces a checkable summary, and have a clause checklist for verification. This article is general information about using AI tools. It is not legal advice, and it does not replace review by a qualified professional.

Step 1: Check where the contract text goes

Before you paste or upload a contract, answer three questions.

1. Are you allowed to share it with this tool? Many contracts include confidentiality clauses or customer data. Check your own agreement and your organisation's AI policy first.
2. Is the content used for training? Vendor statements differ by product:
   • OpenAI's API documentation states: "As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models", unless you opt in. It also says abuse monitoring logs are retained for up to 30 days by default, and that Zero Data Retention is an approval-based control.
   • Anthropic's privacy centre states that, by default, it will not use inputs or outputs from its commercial products for training, with exceptions such as feedback you explicitly submit. Consumer plans are covered in a separate article, so read that one if you use a personal account.
   • Microsoft's documentation states that prompts, responses and data accessed through Microsoft Graph are not used to train the foundation models behind Microsoft Copilot, and that Copilot only surfaces organisational data the user already has permission to see.
3. How long is it kept, and who can see it? Look for retention periods, admin controls and audit options for the plan you use. For Microsoft 365, prompts and responses are stored with your other Microsoft 365 content and can be managed with Microsoft Purview.

Vendor statements describe defaults and plan types, and they change. Re-read the linked pages in the References section for the plan you actually use, and ask your administrator which plan and settings your organisation has enabled. A statement that applies to the API or to a business plan does not automatically apply to a personal account. Anthropic's page also notes that feedback you submit, for example with a thumbs up or down button, can be used for training, so avoid rating chats that contain confidential contract text.

If you cannot answer these questions for a tool, do not upload the contract. Use an approved tool, or redact the parties and sensitive figures first.

Step 2: Prepare the document

Quality in gives quality out.

• Use the final or current version, and say so in the prompt.
• Prefer a text-based file over a scan. If you must use a scanned PDF, check the extracted text for missing pages and garbled numbers, because errors there carry straight into the summary.
• Include schedules, order forms and annexes. Key commercial terms, service levels and data terms often sit outside the main body.
• Remove anything the task does not need, such as unrelated personal data.
• Note the page count so you can tell whether the model saw the whole document.

Step 3: Ask for a summary that can be checked

A plain "summarise this contract" invites a smooth paragraph that is hard to verify. Ask for structured output tied to the source. Anthropic's guidance for reducing hallucinations recommends having the model extract word-for-word quotes before analysing, and having it verify each claim with a supporting quote and retract any it cannot support.

You are helping a non-lawyer understand a contract. Use only the contract text below.
Do not use outside knowledge and do not give legal advice.
 
<contract>
{{PASTE OR ATTACH CONTRACT TEXT}}
</contract>
 
Step 1: Extract exact quotes, with clause or section numbers, for each topic below.
If a topic is not addressed, write "Not found in the document."
 
Topics: parties; effective date; term and renewal; fees and payment terms;
price changes; termination (by each party); liability cap and exclusions;
indemnities; confidentiality; data protection; intellectual property;
service levels; assignment; governing law and dispute resolution.
 
Step 2: Using only those quotes, write a plain-English summary table with the columns
Topic | What it says | Clause | Exact quote.
 
Step 3: List anything unusual, one-sided or ambiguous, each with a clause reference.
Do not guess. If you are unsure, say "I don't have enough information."

The "Not found" rule matters. Anthropic notes that explicitly allowing a model to say it does not know can drastically reduce false information, and it keeps the model from inventing a clause to fill a blank.

Step 4: Verify against the original

Open the contract next to the summary. For every row in the table, find the clause and compare. Pay particular attention to the following.

Then look for the places where summaries tend to fail.

• Dropped conditions. Words such as "unless", "provided that", "subject to" and "except" change meaning. A summary that says "either party may terminate on 30 days' notice" may have lost a clause saying that this applies only after the first year.
• Cross-references. A clause may be modified by a definition, schedule or amendment elsewhere. Search for the defined terms the summary relies on.
• Defined terms. A capitalised term such as "Fees" or "Confidential Information" may be defined more narrowly or broadly than the plain-English meaning.
• Numbers and dates. Re-read every figure. Check that the summary did not mix up days and business days, or months and years.
• Missing sections. If the model summarised only the main body, the schedules were probably not read. Ask it to list the headings it saw and compare them with the table of contents.

Step 5: Decide what needs a human expert

Use the summary to decide where to spend expert time, not to avoid it. Send the contract to a qualified legal or finance professional when it involves any of the following:

• High value, long duration or automatic renewal.
• Unlimited or uncapped liability, or unusual indemnities.
• Personal data, regulated data or cross-border transfers.
• Intellectual property ownership or licensing.
• Exclusivity, non-compete or minimum-purchase commitments.
• Anything the summary flags as unclear.

Keep a record of which tool and prompt were used, the version of the contract, who checked the summary and when. For related guidance on checking AI output more generally, see AI hallucinations at work: a verification checklist.

Troubleshooting

The summary cites a clause number that does not exist. The model has miscounted or invented a reference. Rerun with the instruction to quote exactly, and treat any row without a verifiable quote as unsupported.

The summary leaves out whole sections. The document may have exceeded what the tool processed, or the text extraction failed. Split the contract into parts and summarise each, then compare the headings with the original.

Two runs give different answers. Treat the differences as the places to read yourself. Anthropic describes running the same prompt several times and comparing results as a way to spot hallucinations.

The model gives legal advice or a confident verdict such as "this is safe to sign". Discard that part. The prompt should ask for what the document says, not for whether you should sign it.

You cannot upload the contract at all. Use an approved internal tool, such as Microsoft Copilot working on files you already have permission to open, or summarise redacted extracts. Microsoft states that Copilot respects the user's existing permissions, so confirm the file's sharing settings are correct before using it. See Microsoft 365 Copilot readiness and oversharing.

Checklist

• Confirm you may share the contract and that the tool's data handling fits your obligations.
• Use the final version, with all schedules and annexes.
• Prompt for exact quotes and clause numbers, with "Not found" for missing topics.
• Compare every row of the summary with the original text.
• Re-read all parties, dates, amounts, notice periods and defined terms.
• Escalate high-value, high-risk or unclear contracts to a qualified professional.
• Record the tool, prompt, version and reviewer.

References

• Anthropic, Reduce hallucinations: https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations
• Anthropic Privacy Center, Is my data used for model training?: https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training
• OpenAI, Data controls in the OpenAI platform: https://developers.openai.com/api/docs/guides/your-data
• Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-privacy