ChatPDF is often easier for checking a supplier contract clause against the original page, because its answers link to the PDF beside the chat. ChatGPT is more useful when you must compare clauses, amendments and your own checklist in a shared Project. Neither should approve a contract: choose the tool your confidentiality rules permit, then verify every material answer against the signed document.
The page-limit comparison needs care. ChatPDF's public API documentation gives a 2,000-page or 32 MB limit per PDF for that API; it does not establish the same cap for every website account. OpenAI's upload help states 512 MB per file and a two-million-token cap for text documents in ChatGPT, with other storage and plan limits. Test the actual contract in the intended account before buying either product for size alone.
Make a clause register before asking either tool
A supplier agreement is not a normal reading-comprehension exercise. A single answer can depend on a definition, a schedule and an amendment signed months later. Start with a register of the terms that affect money or service: price and indexation, minimum order, delivery, defects, returns, cancellation, renewal, liability, confidentiality and the order of precedence between documents. Put “source page”, “answer”, “exception” and “reviewer” beside each item. The register makes missing clauses visible.
For example, a print shop reviewing a paper supplier contract should not ask only “What is the delivery time?” It needs to know whether the three-day dispatch statement is a binding commitment or an estimate, whether a separate schedule changes it for special paper, and whether delay gives a right to cancel. A useful AI question names the clause and demands the page: “Find every provision that changes the delivery timetable or the remedy for delay. Quote short fragments, identify the document and page, and say when the answer is uncertain.” The human then reads those pages.
In a five-person wedding planning office, an owner may hold a venue agreement, an amendment and an email that appears to waive a fee. The email may not satisfy the agreement's formal variation clause. AI can find the apparent conflict, but a qualified adviser should resolve legal effect when it matters. Do not turn a summary into legal advice. The office's practical job is to build a complete evidence packet for the decision-maker and ask the right follow-up questions before signing.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowWhat ChatPDF handles well in a contract pass
ChatPDF's official site advertises side-by-side reading with citations that jump to original PDF content. It also advertises multi-file chats using folders, accepts PDF, Word, PowerPoint, Markdown and text files, and can be tried without an account, though its FAQ says saved history and multi-document chats need a free account. Its main site says a free user can analyse two documents daily; Plus is described as providing unlimited document analysis. Those statements describe document count, not a verified answer to every page, file-size or question allowance for a particular account. Check the current plan screen for those details.
A reviewer can work clause by clause. Upload an approved copy, ask for the termination provision, click the cited page and note whether the page number matches the printed document. Then ask for a contradictory or missing provision, such as an exception in an annex. ChatPDF's direct page navigation is useful because the owner can keep the question and document visible. It does not eliminate the need to read surrounding sentences; a citation to the correct page can still support a wrong interpretation.
The API documentation offers one verified technical ceiling: PDFs submitted through its backend API are limited to 2,000 pages or 32 MB each. That does not prove the website UI accepts a contract of the same size or that a free account can process it. The prudent test is to upload a harmless sample with the same page count and scan quality. If the real agreement is 400 pages but most pages are appendices, ask whether the tool indexed the definitions and amendment as well as the opening pages. Split a document only when doing so does not hide the relationships between sections.
ChatPDF states that documents are private by default and may be shared through links that it says you can revoke; its main site says files are encrypted in transfer and at rest with a SOC 2 Type II certified storage provider, and can be deleted at any time. Its FAQ also says queries are routed between two OpenAI models, so a third party processes the text. Those statements do not answer every procurement question about retention, subprocessors or model training. The privacy policy linked from the site footer is a hosted page last updated 5 January 2023; the part readable in this check covers account details such as names, billing and email addresses and gives no retention period for uploaded documents. For a confidential supplier contract, ask ChatPDF for applicable written terms before upload, or use a cleared redacted test document. A claim of “secure” on a home page is insufficient evidence for a sensitive contract.
What ChatGPT adds with a Project
ChatGPT can read uploaded documents and combine the contract with the firm's own review checklist. Its Projects feature allows a defined collection of files and instructions, and OpenAI's help page lists 40 files per Project for Business and Enterprise users, with a maximum of ten files uploaded at one time. A Project can hold the current agreement, amendments, price schedule and a short policy for how staff should present findings. That does not make all Project members legal reviewers. Set membership deliberately.
OpenAI's file guidance says uploads to conversations have a hard 512 MB per-file maximum and text and document files are capped at two million tokens each. A scanned PDF that lacks usable text can still be a practical problem well below those limits. If a table is embedded as an image, test whether the actual account extracts it correctly. An answer about a unit price should be checked against the source table, not a model's fluent reconstruction of it.
File size is not the same as how much the model reads at once. OpenAI's Business comparison lists the fast Instant setting at an input maximum of about 40 pages and the reasoning setting at about 320 pages; Enterprise raises Instant to about 250. An 80-page supplier packet therefore sits well above the Instant figure. Run the clause check with a reasoning setting, and test whether answers cite late pages such as amendments and schedules, not only the opening clauses. OpenAI's upload help also caps uploads at 80 files every three hours, which matters only if someone loads a whole archive.
Business and Enterprise plans do not use business content for model training by default according to OpenAI's pricing page. Consumer ChatGPT has different data settings, so an employee's personal account is not an approved substitute for a company workspace. OpenAI's Projects help says shared Projects inherit their workspace's data controls and restrictions. It also says deleting a project removes its chats, instructions and project-only files from OpenAI's systems within 30 days unless OpenAI must keep them for legal or security reasons, while files saved separately in the Library follow their own rules. That is useful, but the office still has to decide who may upload supplier documents, who may join the Project and what happens when the contract is replaced. The customer-data tutorial provides the broader policy check.
A Project instruction can force a useful format: “For each answer, identify the source file, page or section, exact contractual wording, competing wording and an uncertainty flag. Never invent a clause.” Yet an instruction is only a request to the AI, not a guarantee that a page reference is correct. If the model cannot locate a page, a human should search the PDF or mark the answer unverified. For a general primer on where ChatGPT file reading breaks, use that separate tutorial; the concern here is supplier-contract evidence.
A 72-page supplier agreement, checked twice
Consider an illustrative four-person print shop. It receives a 72-page paper supply agreement, a six-page price schedule and a two-page amendment. The draft says standard paper arrives within five working days, but the amendment changes the delivery expectation for two specialist lines. The owner wants to know which line has the longer lead time, what notice is required to cancel an order and whether a price increase applies to already accepted orders. The total packet is 80 pages, though each tool's acceptance still needs a trial.
The owner manually identifies three expected source locations before asking AI: clause 4.2 on page 11 for general delivery, the price schedule's table on page 3, and amendment paragraph 2 on page 1. Those are illustrative page references, not statements about a real agreement. In ChatPDF, the owner loads the PDFs into a folder, asks each question and clicks cited pages. In ChatGPT Business, the owner creates a restricted Project with the files and the review register, then asks for a clause matrix. Both produce draft findings. The reviewer checks each against the three marked locations and searches for exceptions.
Suppose ChatPDF answers “five working days” and cites page 11 but misses the amendment on the specialist line. Suppose ChatGPT identifies the amendment and the seven-day period but incorrectly applies it to all paper. Both results are partially useful and materially wrong. The corrected register should say: “Standard line: five working days, clause 4.2; named specialist lines: seven working days, amendment paragraph 2; verify schedule item codes before order.” The correction is the real deliverable, not the AI output. A strong workflow records the error so staff do not repeat it on the next contract.
For an illustrative cost comparison, say the owner spends 35 minutes marking expected clauses and another 25 minutes checking the two AI outputs. That is 60 minutes of human review. If a manual first read would have taken 90 minutes for this packet, the AI-assisted process might save 30 minutes on this occasion, but only if it catches the amendment and the team does not spend extra time resolving new errors. The figures are an example of how to measure, not a performance claim. Subscription price should be compared with reviewed packets per month, not answers generated.
Use a prompt that demands evidence and exclusions
Paste a short prompt into the approved tool after the source packet is ready. Keep it specific to supplier terms and tell the model what to do with missing information:
Review the attached supplier agreement and signed amendments.
Find the delivery period, cancellation notice and price-change rule.
For each finding give: source file, page or clause, short quotation,
any exception or later amendment, and an uncertainty flag.
Do not infer a term that the documents do not state.
If two sources conflict, show both and ask for human review.An illustrative output might say: “Cancellation notice: 30 days before renewal, agreement page 19, clause 8.1. Amendment page 2 changes notice to 45 days for the next renewal. Confirm the amendment's effective date.” The reviewer should improve it by adding the actual date, whether notice must be in writing, the recipient address and how the deadline is calculated. The example shows why a short answer can be inadequate even when the main number is right.
Another example is a price schedule stating “prices may be revised on 60 days' notice”. The AI may report that the supplier can raise prices at any time after 60 days. The missing issue is whether the revised price applies to accepted purchase orders or only future orders. Ask a second question about that specific scope and search the order-acceptance provision. If the contract remains unclear, flag it for negotiation or qualified advice. Do not let AI supply a commercially convenient interpretation.
Page references need their own quality check
PDFs can have a cover page and printed page numbers that differ from the viewer's page count. A citation to viewer page 18 may correspond to printed page 16. Put both in the register when they differ. For a scanned contract, inspect whether the text layer preserves footnotes, tables and handwritten changes. In a print shop example, a scanned price table may misread “0.06” as “0.60”; a single decimal error changes a 10,000-sheet order from $600 to $6,000. Compare figures visually against the original before using them in a purchase decision.
Ask one question that the packet cannot answer, such as “What was agreed verbally at the last supplier meeting?” A safe response is “not in the supplied documents”. If either product invents a meeting record, mark the tool's output as unreliable for unsupported questions. A sixth concrete example is a missing signature page: AI may summarise an unsigned amendment as if it were effective. The reviewer must check execution status before treating the text as binding. These checks are specific to contract review, beyond a general chatbot quality test.
Resolve a conflict without making the AI the decision-maker
Give every conflicting term a status in the register: confirmed, conflicting, missing source or adviser review. Confirmed means the reviewer has read the cited passage and checked amendments. Conflicting means two passages appear to say different things, even if the AI offers a persuasive reconciliation. Missing source means the answer lacks a supporting page or the document itself is absent. Adviser review means the business consequence is significant enough that someone qualified should interpret or negotiate the language. These labels prevent an uncertain answer from being pasted into a purchasing summary as settled fact.
An illustrative supplier packet contains a general return period of 14 days, while a schedule states that made-to-order stock cannot be returned. ChatPDF might show the general clause first; ChatGPT might combine the two. The reviewer should record both with their clause references and ask which products the schedule defines as made-to-order. If the schedule has no clear product list, neither tool can supply it. The purchasing lead can ask the supplier for a written clarification before placing a large order. That is a concrete outcome from the AI review even when the contract answer remains unresolved.
Set an escalation threshold in advance. For example, the print shop might require owner review for any term that changes an order by more than $500, changes a delivery promise to a customer or creates an automatic renewal. The $500 is an illustrative internal threshold, not a legal standard. A routine wording issue can stay with the buyer; a disputed liability cap goes to a qualified adviser. Add the owner and due date to the register so uncertain clauses do not disappear after the chat session closes.
When the supplier sends a new version, compare the changed clauses rather than rerunning only the original questions. Save the old answer and corrected register as evidence of why a buying decision changed. Ask the tool to identify added, removed and altered wording, then verify the changes using a document comparison or manual page inspection. An AI summary may miss a single inserted “except” that reverses a duty. Treat the signed final copy as the source of truth and close the register only after the reviewer checks it.
Use version labels in filenames: “supplier-agreement-draft”, “amendment-signed”, “price-schedule-current”. Remove obsolete drafts from the live Project or folder after preserving them in the firm's proper records system. A folder containing three files called “final” invites the model to cite the wrong version and the reviewer to open the wrong page. The AI tool should support the records process, not become the sole archive. For a broader office document-tool comparison, consider the existing storage system before copying files elsewhere.
Privacy and purchase checks before uploading a real contract
Classify the packet first. Public product terms can be a safe demonstration; an agreement containing customer details, confidential pricing or supplier trade secrets needs approval under the firm's policy. Ask each vendor about model training, processing, deletion, sharing links, subprocessors and applicable account terms. Verify which ChatGPT workspace a staff member has selected before upload. Verify whether a ChatPDF link is private, who receives it and how it can be revoked. A “private by default” interface still permits deliberate sharing.
Keep the procurement calculation narrow. ChatPDF's free site allowance of two documents a day may be enough for occasional low-risk checks, but its Plus price and detailed site limits should be read from the current checkout before a cost claim is made. A ChatGPT Business seat may already be paid for internal work; if so, the incremental software cost of a Project test can be zero, though staff review time is not. Do not purchase a second product merely because its demo answered one question faster. Test the full clause register, access controls and deletion process.
Choose ChatPDF when the dominant job is page-by-page checking in a PDF and its written data terms fit the packet. Choose ChatGPT Business when the team needs a shared Project, a repeatable instruction set and cross-file comparison under its existing workspace controls. Use neither for unreviewed contract approval. Hand disputed clauses and legal interpretation to the appropriate qualified adviser. The success measure is a correct, source-backed register that a human can sign off, with every exception and amendment visible.
Further reads
- Acrobat AI Assistant for Teams: Plans, Pricing and Admin Controls — Explore cited PDF review in Acrobat.
- Box AI for Documents: Ask Questions of Contracts and Policies — Review contracts already stored in Box.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ChatPDF home page and FAQ (chatpdf.com); ChatPDF backend API documentation (chatpdf.com/docs/api/backend); ChatPDF privacy policy linked from the site footer (hosted page, last updated 5 January 2023); OpenAI Uploading files and audio to ChatGPT and Projects in ChatGPT (help.openai.com articles 8555545 and 10169521); OpenAI Business pricing and feature comparison (openai.com/business/pricing/). Checked 5 November 2026.