AI document analysis
AI document analysis that keeps every finding tied to the source.
Bring the documents and the question. ZeroTwo helps extract the relevant passages, compare terms or claims, and prepare a reviewable findings brief without pretending the model replaces a qualified reviewer.
Anatomy of a finding
- Statement
- One sentence on what differs or what the document says.
- Anchor
- File, page, section and the quoted text, per document.
- Basis
- Direct evidence
- Interpretation
- Needs a reviewer
Four jobs
What document analysis extracts or compares
Pick the job first. It decides which files you supply, what comes back, and what you must check before the result is used.
| Job | You provide | You get back | Check before relying |
|---|---|---|---|
| Extract | One or several documents and the specific things to pull out: clauses, defined terms, dates, amounts, obligations, table rows. | A list or table of what was found, each item with its file, page or section anchor and the quoted text. | Open each anchor and confirm the quote is on that page. Look for items the document contains that the list omits. |
| Compare versions | Two versions of the same document, such as an agreement and its amendment, or a policy before and after a rewrite. | A comparison table of what changed, what stayed and what one version does not mention at all. | Read both anchors side by side. Treat “not mentioned” rows as questions for a person, not as proof the term is unchanged. |
| Reconcile claims | Several reports or letters that describe the same event, figure or commitment, plus the claim to test. | Where the sources agree, where they conflict, and which statements only one source makes. | Confirm the sources are the same period and the same definition before treating a mismatch as a conflict. |
| Check against a standard | A document and the checklist, policy or requirement list it should satisfy. | Requirement by requirement: where the document addresses it, where it does not, and where it is ambiguous. | Spot-check the “not addressed” rows by searching the file yourself; absence is the easiest thing to get wrong. |
This is a description of how the work is organized, not a benchmark of accuracy or speed.
Illustration
Two documents, three findings, one review queue
A made-up agreement and amendment, used only to show the shape of the output. The text below is invented sample content. It is not output from a ZeroTwo run and not a measured result.
Services Agreement, signed
- §4.2, p.3 Invoices are payable net 30 days from receipt.
- §9.1, p.7 Liability is capped by reference to fees paid in the prior period.
- §11.3, p.8 Either party may terminate on written notice stated in days.
Amendment No. 2, draft
- §2, p.1 Section 4.2 is replaced: invoices are payable net 45 days.
- No reference to liability or the cap.
- Schedule C, p.4 A scanned image of a notice-period table.
| Finding | Document A anchor | Document B anchor | Basis |
|---|---|---|---|
| F1 · Payment terms | Services Agreement §4.2, p.3: “Net 30” | Amendment §2, p.1: “Net 45” | Direct evidence |
| F2 · Liability cap | Services Agreement §9.1, p.7: cap set by fees paid in the prior period | Amendment: no mention | Interpretation |
| F3 · Termination notice | Services Agreement §11.3, p.8: notice period stated in days | Amendment Schedule C, p.4: scanned table | Needs a reviewer |
One extraction failure, kept out of the findings
Suppose the Schedule C table in the invented amendment is a low-quality scan, and a digit in the notice period could be read two ways. F3 stays in the review queue with the page image attached. It is not reported as a difference until a person has read the page. That is the review boundary: the model surfaces the candidate, a person decides.
What comes back
- Cited findings
- Comparison table
- Review queue
Process
Move from source material to a reviewed output
Keep the evidence, the intermediate decisions and the finished file connected instead of scattering them across tabs.
Inventory the files
Identify versions, missing documents, extraction limits and the scope of review before asking any question.
Find the relevant evidence
Locate the passages and tables that address the question rather than summarizing everything.
Compare and qualify
Separate direct evidence, interpretation, conflicts and unanswered questions.
Prepare the findings brief
Organize conclusions, citations, limitations and actions for a human reviewer.
Before you upload
Start with the context analysts and client-service teams already have
The analysis is only as good as the file inventory behind it. Settle these three things first.
Authorized documents
Upload the relevant PDFs or supported business files and leave out sensitive material the question does not need.
An analysis question
Name the clauses, claims, differences, risks or decisions to investigate. “What changed in payment and termination terms” beats “analyze these”.
A citation standard
Ask for page, section or quoted-text anchors on every finding that affects a decision, and say what to do when an anchor cannot be found.
File inventory checklist
- Which version is current, and are older versions in the set?
- Are any pages, exhibits or schedules missing?
- Which files are scans rather than text?
- Who approves the conclusion?
This page is for a bounded set of files and one question. If you need to answer questions across a maintained organizational corpus, that is a different job with a persistent knowledge base.
Operating limits
Where extraction goes wrong, and how to catch it
OCR, scans, tables, handwriting, missing pages and ambiguous language can all cause errors. Legal, medical, financial and compliance conclusions require qualified review.
Scans and OCR
A scanned page is read as an image. Digits and punctuation are where it slips: one figure can come back as another.
To catch it: Compare every number in a finding with the page image.
Tables
Merged cells, rotated headers and tables split across pages can attach a value to the wrong row.
To catch it: Check the row and column labels, not only the value.
Handwriting and stamps
Margin notes, signatures and stamps are often skipped or misread.
To catch it: Ask what the file contains that is not typed text.
Missing pages
A comparison can only be as complete as the files supplied, and it can report an absence that is really a missing exhibit.
To catch it: Confirm page counts and attachments before you start.
Ambiguous language
“Reasonable efforts” or “material change” have no single reading. A model can pick one confidently.
To catch it: Send those items to a qualified reviewer.
Before approval
Three checks before anyone relies on the brief
- 01
Human approval gates
Review facts, calculations, external actions and client-facing language before approval or publication.
- 02
Claim-to-source checks
Require citations or calculation anchors for statements that affect a decision, and reject any finding without one.
- 03
Output inspection
Open the finished file, confirm the required sections exist, and test any links, formulas or exports before use.
Plans
Try a small comparison before you commit
Start on the Free plan with two short files. Model access, usage and file limits depend on your plan and the providers used during the run, so the pricing page is the source of truth.
Free
$0/mo
100 free credits to start. 1 file upload per message.
Plus
$14.99/mo
See the pricing page for usage and file limits.
Pro
$29.99/mo
See the pricing page for usage and file limits.
Choose the right workflow
Document analysis is not always the job
Pick by the assignment, not by the file type. Several of these start from the same PDF.
- Compare or extractStay on this pageYou need differences, obligations or evidence across documents, with an anchor for each finding.
- One document, shorterAI PDF summarizerYou need a bounded summary of a single file and no comparison.
- Quick questionsChat with PDFsYou want to ask a file questions in conversation without producing a findings brief.
- Tables into numbersAI agent for data analysisThe evidence is already a table and the next step is checked calculation.
What people ask
What does ZeroTwo extract or compare in document analysis?
Extraction means pulling named items out of a document: clauses, defined terms, dates, amounts, obligations and table rows, each with its location. Comparison means setting two or more documents against each other, such as an agreement and its amendment, to show what changed, what stayed the same and what only one of them says. Either way the output is a set of findings you can check against the source.
How do I check a finding against the source?
Ask for the file, page or section and the quoted text with every finding, then open that location and read the passage. For numbers and dates, compare against the page image rather than the extracted text. If a finding has no anchor you can locate, treat it as unsupported.
What goes wrong with scans, tables and handwriting?
Scanned pages, complex tables and handwriting are the most likely places for extraction errors, including a wrong digit or a value attached to the wrong row. Missing pages and ambiguous wording cause a different kind of error: a confident conclusion that the source does not support. Put these items in a review queue for a person instead of reporting them as findings.
Does ZeroTwo replace a qualified reviewer?
No. ZeroTwo prepares and organizes the evidence. Legal, medical, financial and compliance conclusions need a qualified person to approve them before use.
Can I compare two documents on the Free plan?
The Free plan lists "100 free credits to start" and "1 file upload per message", with no card required. That is enough to try a small comparison, but with one upload per message you should confirm that both documents are being read before you rely on the result. The pricing page shows what each paid plan allows.
Is it safe to upload confidential documents?
Only upload material you are authorized to use, and remove fields the question does not need. For how ZeroTwo handles your data, read the privacy policy before uploading anything sensitive.
Explore related workflows
- AI for analysts
The same evidence-first habits applied to analyst work.
- AI for consultants
Reviewable findings for client-service teams.
- AI report generator
Turn checked findings into a finished report.