Skip to content
New: the feature-by-feature comparison tool covers all 18 platforms.Compare features
Take the quiz
AI capabilities

Preparing a Document Set for AI Review

AI review is only as good as the documents it reads. This guide covers the preparation that makes AI search, summaries and extraction work: text layers, formats, naming, bundles and exclusions.

By Data Rooms Software editorsPublished 8 min read

When a deal team switches on AI features in a data room, expectations are high: instant summaries, natural-language search, extracted key terms across hundreds of contracts. The first results are often uneven. One contract is summarized perfectly; the next returns nothing useful; a third is summarized as if it were two different documents.

The difference is rarely the AI. It is the documents. Models work from the text they can extract, and deal document sets are full of scans, merged bundles, odd formats and cryptic file names. This guide covers the preparation that makes AI review work, and the decisions about what AI should not see at all.

Why does preparation change AI results?

Most AI features in a data room follow the same pattern. The platform extracts text from each file, splits it into passages, indexes those passages and retrieves the relevant ones when a user asks a question or requests a summary. A language model then writes the answer from what was retrieved.

Every step depends on the text. If extraction fails, nothing is indexed. If one PDF contains twenty separate contracts, retrieval mixes passages from all of them. If a file is called “Scan_0043.pdf”, the model has no context about what it is reading. Good preparation removes these problems before they reach the model.

Which file formats work best?

FormatHow AI reads itPreparation
Native PDF exported from WordClean text, structure mostly keptReady as is
Word documentClean text and headingsAccept or reject tracked changes; remove comments
Scanned PDF with OCR text layerReadable, quality depends on the scanCheck a page by searching for a word you can see
Scanned PDF without OCRNothing readable unless the platform runs OCRRun OCR before upload, or confirm the platform does
Excel workbookCell values read; layout, formulas and hidden sheets varyRemove hidden sheets and external links; add a short description tab
Image files (JPG, PNG)Only if the platform runs OCR on imagesConvert to PDF with OCR
Email files (MSG, EML)Often poorly supportedExport to PDF with attachments as separate files
Password-protected or encrypted filesUnreadableRemove protection on the disclosure copy
Zip archivesMust be unpacked firstUnpack and upload the contents

A practical test before upload: open a scanned PDF, use the find function to search for a word you can see on the page and check it is found. If it is not, the file needs OCR before AI can read it.

How should bundled documents be handled?

Deal document sets often contain bundles: a single PDF holding a master agreement, five amendments and three side letters, or a year of board minutes scanned in one file. For people, bundles are an annoyance. For AI, they are a source of error.

When a summary is requested on a bundle, the model may blend terms from different documents, describe the original agreement as if the amendments did not exist, or simply run out of room to cover everything. Extraction across contracts suffers too: a key-terms table built from a bundle has one row where there should be nine.

Split bundles into separate files before upload, one per legal document, and name each one clearly. Keep amendments next to the agreement they amend, numbered so they sort in order. If splitting is not possible in time, at least add a cover page listing the documents in the bundle and their page ranges.

Why do file names matter to AI?

File names are context. A model asked to summarize “Scan_0043.pdf” has to infer what the document is from its content alone. A model asked to summarize “3.2.4 Supply Agreement - Northwind Ltd - 2023-04-01.pdf” starts with the document type, the counterparty and the date, which improves both summaries and retrieval.

A good convention includes the index number, document type, counterparty or subject, and date. Apply it consistently, avoid abbreviations only your team understands, and remove version labels such as “FINAL v3 (2)” from disclosure copies. The same names help human reviewers and keyword search.

What should AI not process?

Before switching AI features on, decide which documents should be excluded from AI processing or handled under separate rules. This is a legal and policy decision, not a technical one.

CategoryConcernTypical approach
Legal advice and privileged materialPrivilege may be put at risk by wider processing or disclosureKeep out of the room, or in a restricted folder excluded from AI features
Clean team materialCompetitively sensitive data limited to named peopleRestricted folder; confirm AI answers respect folder permissions
Personal data (HR files, customer records)Privacy law, data processing terms, cross-border transfersRedact or minimize first; check the vendor’s processing terms
Third-party confidential documentsContractual limits on who may process themCheck the confidentiality terms before upload
Export-controlled technical dataRules on where data may be processedSeek advice; may need to stay out of cloud AI entirely

Two questions to ask every vendor apply here. Does AI processing respect the same folder permissions as viewing, so that a user cannot get a summary of a document they cannot open? And where is document content processed when AI features run, including which model providers are involved and whether content is retained or used for training?

How do platform types compare on AI and automation?

Our AI and automation criterion reflects whether a platform lists AI assistance, plus bulk import, a developer API and a Q&A workflow. Averaged by software category, it shows where AI features are most common in our facts file.

Average AI and automation score by software category

Virtual data room12 platforms, 4 with AI listed 6.3
File-sharing platform3 platforms, 1 with AI listed 6.0
Capital markets room1 platform, 0 with AI listed 4.5
Document tracking tool2 platforms, 0 with AI listed 3.0
Five of the 18 platforms list AI assistance: four virtual data rooms and one file-sharing platform.
dataroomssoftware.info
Average of the AI and automation criterion (out of 10) per category, from our scoring model and facts file. Category labels are our own.

The averages hide wide spreads inside each category, so they are a guide to where to look rather than a ranking. For platform-level detail, the AI data room software hub lists the AI capabilities recorded for each platform.

What is the preparation routine?

Preparing a deal document set for AI review

  1. Agree exclusions

    With counsel, list the categories of documents that must not be processed by AI or must sit in restricted folders.

  2. Check text layers

    Test a sample of scanned PDFs for searchable text. Run OCR on any that fail, or confirm the platform runs OCR on upload.

  3. Convert awkward formats

    Export emails to PDF, convert images to searchable PDFs, unpack zip archives and remove passwords from disclosure copies.

  4. Split bundles

    Break multi-document PDFs into one file per legal document, keeping amendments next to the agreements they change.

  5. Rename consistently

    Apply a naming convention with index number, document type, counterparty and date.

  6. Clean spreadsheets

    Remove hidden sheets and external links, and add a short description tab explaining what the workbook contains.

  7. Test on a sample

    Upload twenty representative files, run summaries and searches, and check the results before uploading the full set.

The general clean-up that comes before any upload, such as removing duplicates, drafts and hidden metadata, is covered in our guide to cleaning up documents before a data room upload.

How do you know preparation worked?

Run a short check after the upload. Pick ten documents of different types and ask the AI feature to summarize each one, then compare the summaries with the documents. Search in plain language for three facts you know are in the set. Ask for a list of every contract with a particular clause and count whether the list is complete.

If results are weak for one type of document, such as older scanned leases, the problem is usually text quality. Re-scan or re-OCR those files rather than relying on the AI to cope. And whatever the results, AI output still needs checking against the source before it reaches a diligence report.

Frequently asked questions

Can data room AI read scanned documents?

Only if the scan has a text layer from OCR, either created before upload or by the platform. Test a sample of scanned files before relying on AI summaries or search.

Should privileged documents go into a data room with AI features?

Usually privileged advice should not be in the disclosure room at all. If it must be, keep it in a restricted folder and confirm with the vendor that AI processing respects folder permissions.

Does AI work on Excel files in a data room?

Many AI features read cell values, but layout, formulas and hidden sheets are handled inconsistently. Clean workbooks and add a description tab to help.

How long does document preparation take?

For a few hundred documents, one to three days of work, mostly on OCR checks, splitting bundles and renaming. It is far quicker than correcting AI output afterwards.