How ContractExtract Works

What the tool does

ContractExtract reads an uploaded contract and returns a structured report identifying the contracting parties, key dates, payment terms, obligations, termination clauses, and any provisions the model flags as potentially risky or unusual. The output is a plain-English reference designed to help you spot what to look at before a detailed review, not a substitute for that review.

The intended use is pre-review triage: a freelancer scanning a new client agreement, a tenant reviewing a lease, an employee reading an offer letter, or a small business owner working through a vendor contract. The goal is to surface the questions worth asking a qualified attorney, not to answer them.

What model powers extractions

Extractions are produced by Anthropic's Claude API. The model identifier currently in use is claude-sonnet-4-20250514, Claude Sonnet 4. PDFs are sent directly to the model using Anthropic's native document support, image uploads use the same model's vision input, and DOCX files have their text extracted server-side before being sent. The system prompt instructs the model to return a single JSON object containing the extracted fields and nothing else.

The model identifier above is the value hard-coded in the extraction route at the time of this writing. If you are reading this page after a model upgrade, the deployed code is authoritative. See Anthropic's model documentation for current capabilities.

Fields the extractor identifies

For every uploaded contract, the model is asked to return the following structured fields:

• Summary. A short plain-English description of what the contract is and what it does.
• Parties. The named entities or individuals bound by the agreement, with their roles where stated.
• Key dates. Effective date, expiration date, renewal triggers, notice deadlines, and any other dates with contractual significance.
• Obligations. What each party is required to do, deliver, refrain from, or maintain under the agreement.
• Payment terms. Amounts, currencies, schedules, late fees, invoicing rules, and any conditional payments.
• Termination clauses. Conditions under which either party can end the agreement, required notice periods, and consequences of termination.
• Risky clauses and flags. Provisions the model identifies as unusual, one-sided, or warranting closer review, with brief reasoning for each flag.

Field names match what is described above. See the clauses glossary for definitions of common contract clause types referenced in extraction output.

How uploaded documents are handled

• Uploads are processed in memory on a serverless function and are not written to long-term storage.
• The contract content is sent to Anthropic's API solely to produce the extraction. It is not used to train the model and is not retained as part of any training corpus.
• The structured extraction result is returned to your browser and held in your session storage for the lifetime of that browser tab. Closing the tab clears it.
• No contract content is sent to advertising systems, analytics vendors, or any third party other than the Claude API.
• The Anthropic API key is read from a server-side environment variable and is never exposed to the browser.
• For details on cookies and analytics, see the Privacy Policy.

Honest limitations

AI extraction is not legal interpretation. The output of this tool should be treated as a structured reading of the document, not a legal opinion about it. Specific limits to keep in mind:

• The model can miss provisions, especially in long or oddly formatted contracts. A clause not flagged is not the same as a clause not present.
• Jurisdiction matters. Contract enforceability depends on the governing law and the venue, and the tool does not analyze how a specific clause would hold up in a specific jurisdiction.
• Clauses are often ambiguous. Two reasonable readers, including two attorneys, can read the same provision differently. The tool returns one reading.
• Risk flags are heuristic. The model uses general knowledge of what tends to be unusual or one-sided. It does not know your negotiating leverage, your industry norms, or your prior relationship with the counterparty.
• OCR and parsing errors can degrade results on low-quality scans or images. If text is unreadable to a human eye, the model is guessing too.
• The tool does not draft, redline, or sign anything on your behalf. It does not file or transmit the contract to any party.

Why this kind of tool matters

Contracts are dense by design. A typical commercial agreement runs ten to forty pages and uses defined terms that pull meaning from other parts of the document. People who do not read contracts for a living often skim, miss material terms, or sign without identifying the provisions worth questioning. Lawyers cost money, and not every agreement justifies a billable review.

Pre-review tools narrow the gap. They give a non-lawyer enough structure to know which clauses to ask about, and they give a paying client a faster on-ramp into a billable conversation. The value is in surfacing questions, not answering them.

Sources for clause definitions

Clause definitions and standards used in our content are grounded in widely recognized legal references:

• American Bar Association, Business Law Section for general commercial contract concepts and best-practice guidance.
• Uniform Commercial Code (Cornell Legal Information Institute) for sale-of-goods, warranty, and merchant rules adopted in substantially the same form across U.S. states.
• Cornell Wex Legal Encyclopedia for plain-English definitions of contract terms cited in our clauses page.
• Federal Rules of Civil Procedure (Cornell LII) for procedural context behind dispute-resolution clauses.