Analyzing a document with AI: contracts, reports, PDFs

What does an AI actually do when you give it a document?

AI document analysis is a machine learning technique that examines a document and automatically extracts its key information [11]. This definition comes from a software vendor, as do most of the sources available on the subject. The goal is to turn static documents, often PDFs or images [12], into usable data [10].

The technical chain, in three steps

1. Text extraction. The system first extracts the readable text from the PDF [25].
2. OCR if necessary. If the document contains scanned pages rather than selectable text, OCR (optical character recognition) may be needed [25]. It converts the image of the pages into text [10].
3. Segmentation and indexing. The extracted content is then divided and indexed [25].

Formats read

Document tools advertise that they read PDFs, Word files, images, scanned forms, emails and reports [15]. Some also process digitized documents and handwritten notes [34].

Simple OCR or contextual understanding

Several vendors distinguish simple OCR from contextual understanding. According to Kairntech, simple OCR only digitizes, whereas document AI adds contextual and business understanding: extraction, classification, semantic search [10]. iWeaver likewise claims that its tool, unlike basic OCR, understands the semantic relationships between concepts [34]. Both claims come from commercial players and describe their own products. They do not constitute an independent measure of the quality of the analysis.

How do you analyze a contract with AI, step by step?

The method comes down to four steps: define a grid, request citations, compare each point against the full contract, then have the result validated by a qualified human [14].

1. Define an analysis grid. Before launching the AI, list the points to examine: parties, term, amounts, liability, termination, governing law. The AI can then flag any deviation from this grid [14].
2. Request citations. For each answer, require the passage of the contract that supports it [14]. This lets you find the source in the text.
3. Compare each point against the full contract. An isolated answer is not enough: check it against the entire document [14].
4. Have a qualified professional validate it. The AI's output does not constitute legal advice [14].

What AI can do with a contract

• Segment the document into sections (preamble, definitions, clauses, annexes) and extract the parties and dates [12].
• Recognize clauses such as limitation of liability or indemnification, and extract parties, dates, amounts and governing law into a structured summary [39].
• Compare the text against a reference contract, identify deviations and prepare questions to examine [35].
• In due diligence, spot non-standard clauses or risks across hundreds of agreements in a few minutes [30].

These uses are described by solution vendors and training organizations, not by institutional sources. They indicate what these tools claim, not a guaranteed level of reliability. The grid, the citations and human validation exist precisely to check what the AI produces.

How do you summarize a PDF or long report and extract data from it?

Summarizing a long PDF and extracting data from it involve three distinct uses, each requiring a different kind of verification.

• Summarization: producing a short version of one or more long documents. Current models handle PDFs of many pages, including tables and appendices: financial reports, technical studies, audits [10]. This capability is presented by tool vendors.
• Data extraction: transforming a document into usable information, whether it is a PDF, a contract or an invoice [10]. AI document analysis relies on machine learning to automatically extract key information [11]. For example, one tutorial shows extraction into Google Sheets with Claude-3 [18].
• Question answering: querying the document directly rather than reading it in full [22]. Some tools display the chat alongside the PDF, with clickable citations that point to the exact location in the source [20].

Checking a dense report

On a dense report, check every extracted figure against the original table. Also verify that the appendices were taken into account and that page references are correct. Clickable citations make this check easier, and the method "request citations, compare each point against the full document" is recommended for contracts [14]. Adobe states that its assistant preserves references to pages, tables and sections [17].

Finance and audit

Academic research places the use in accounting and auditing: AI can read contracts or other documents there [1]. A practical book describes automating the analysis of contracts and PDF invoices [8].

Promises of time savings or return on investment come from vendors. For instance, one solution claims to accelerate ROI for SMEs [19], and a legal services provider speaks of reviewing hundreds of agreements "in a few minutes" [30]. None of the sources consulted measures these gains independently: test on your own documents.

How do you compare two documents or two versions of a contract?

Comparing two documents means checking a text against a reference contract, or two versions of the same contract, to identify deviations and prepare questions to examine. AI can extract the structure of the document, compare the text against this reference contract and identify differences [35]. It flags the discrepancies, but it is up to you to interpret them.

A two-step method

1. Send the documents one after the other, announcing the comparison. Specify which one serves as the reference and which one must be examined [23].
2. Request a table of deviations with a column "what is not comparable." This column is essential: two quotes, for example, often differ in more than just their price [23].

For a contract, you can structure the table as follows: clause, reference version, version under review, nature of the deviation, question to ask. This structure is a suggested presentation, not a requirement from the sources.

Each row of the table must still be checked against both texts. The tool prepares the review; it does not replace it.

Change tracking, a commercial example

Adobe, a software vendor, has presented a feature of this type for contract management. According to LeMagIT, this new capability relies on generative AI. It makes it possible to summarize contracts in PDF format and detect changes made during review cycles [40]. This is a commercial feature, described here as the specialized press reports it.

A precaution specific to contracts

Comparing two versions means submitting documents that are often sensitive. Some vendors respond with local execution. Jaimes AI, for example, claims to analyze documents on a device installed on the premises, without passing through an external cloud service [24]. This is a commercial argument, to be verified against your confidentiality requirements.

Where does Oupi fit into this analysis work?

Oupi reads the documents you provide and answers based on them. This is an operating principle: answers are tied to the content of the document you submit, and this link is what makes verification possible.

This principle aligns with the verification method described for contract analysis, from a training organization (a commercial source). It comes down to four actions: define an analysis grid, request citations, compare each point against the full contract, then have the result validated by a human [14]. With Oupi, this looks like:

1. Frame the request. List the points to examine (term, termination, liability, for example) rather than asking for a general opinion.
2. Request the passages. For each point raised, require the excerpt from the document on which the answer rests.
3. Go back to the text. Reread the excerpt in the full document. An isolated quotation may be accurate yet lose its meaning out of context, for example if an exception appears in another clause.
4. Have it validated. The same source states that the AI's output does not constitute legal advice and must be validated by a qualified professional [14].

An answer anchored in the document is faster to verify, but it does not replace careful reading. What remains the responsibility of the lawyer, the financial analyst or the assistant is the judgment on compliance, risk and interpretation [14].

To consult your own documents in this context, see chat with documents.

What precautions should you take: confidentiality, scans, responsibility?

Six points of vigilance frame the use of AI on documents: human judgment, document length, scan quality, confidentiality, the legal framework, and the absence of a reliability figure.

• Judgment remains human. AI does not alone determine compliance, risk, legal interpretation or the strategy to follow, and its output does not constitute legal advice [14]. The recommended method is to request citations, compare each point against the full contract, then have the analysis validated by a qualified professional [14].
• Long documents require precautions. Work by INSEE on company agreements highlights the strong potential of generative AI for analyzing long documents, while noting that its use calls for caution [3].
• Scan quality affects the result. The system first extracts the readable text from the PDF, and scanned pages may require OCR [25]. An approximate text recognition therefore feeds into all the analysis that follows. It is best to reread critical passages in the original.
• Contract confidentiality raises the cloud question. Some vendors offer local execution: Jaimes AI claims that documents never transit through an external cloud service [24]. Others emphasize the confidentiality of your documents [34]. These are commercial claims, to be verified in the contractual and technical terms of each offer.
• The legal framework is still being developed. Research is looking at the autonomy of AI systems in contract law [4] and at the diligence of parties in using AI tools, including the documentation relating to the tool used [2].

The available sources do not make it possible to quantify the reliability of contract or report analysis. No error rate can therefore be cited, which justifies systematic verification through citations and human validation.