← All posts

ai:documents

Using AI to review company documents: outdated policies and conflicts.

An AI document review reads the internal policies, procedures and knowledge documents on your shared drives and finds the ones that are outdated, duplicated or in conflict, quoting the passage behind each finding. Its scope is internal governance content; legal discovery, reviewing documents for litigation, is a different job with its own rules. Every finding is a candidate with its citation attached, and the decision belongs to the document's owner.

This guide is for the operations, compliance and knowledge-management leads who own those drives. It covers what to set up before the first run, the three checks worth running, and how findings reach the people who can act on them.

What an AI document review does.

The review takes the document set you scope, reads each document, and returns three kinds of finding: documents that are outdated, copies that duplicate or nearly duplicate another, and passages that conflict with a passage in another document about the same rule. Each finding names the documents it connects, quotes the passages and says where they sit.

The review does the reading: comparing every policy against every other policy across years of shared drives. It does not retire a policy, merge two templates or choose winning wording; it puts the candidates in front of the owner with the citations attached.

Start with an inventory: drives, owners and what current means.

Before any model reads a document, write down what the review covers: which drives and sites (the whole tenant, or one department first), which document types (policies, procedures, templates, knowledge articles), and who owns each document. Where the owner is unknown, finding one is the first deliverable of the review, ahead of any finding about content.

Define current in a way a system can check: each document carries a review date and an approver. A document past its review date is a finding whatever its content says, and a document with no date at all is a finding too.

The same inventory sets the removal rule. In Microsoft 365, Purview retention settings can retain content, delete it, or retain it and then delete it, and for files in SharePoint, OneDrive and Microsoft 365 Groups the retention period can start when the file was created or when it was last modified (checked 17 November 2026). Choose the start event with care: a policy still in force can go years without an edit.

Access first: the review reads only what its runner may read.

The review runs under an identity, and that identity's permissions are the ceiling on what it reads and what it can quote back. If it runs with broad admin rights, its findings can put restricted files in front of people who were never meant to see them. Run it under an account whose reach matches the audience of the report, and keep restricted libraries out of the findings even where the runner can read them.

The decision mirrors the tool ceilings in our guide to AI data classification: the review is one more tool reading company data, and its ceiling is the highest class it may receive.

If the review builds a retrieval index over the drives, treat that index as a security boundary of its own. The OWASP Top 10 for LLM Applications 2025 entry LLM08:2025, Vector and Embedding Weaknesses, warns that inadequate or misaligned access controls can lead to unauthorized access to embeddings containing sensitive information, and that the model could retrieve and disclose personal data, proprietary information or other sensitive content; its mitigations include fine-grained access controls, permission-aware vector and embedding stores, and strict partitioning between classes of users or groups.

Microsoft starts its Copilot deployment guidance with the same cleanup: remediating oversharing is the first of its three pillars, with recommended admin actions to identify high-risk sites and files, apply interim access restrictions if needed and fix access issues, using Microsoft Purview and SharePoint Advanced Management (checked 17 November 2026). A document review inherits the same exposure: an overshared file becomes a quotable finding.

Three checks: outdated, duplicates, conflicts.

With the inventory and the access rule set, the review runs three checks. Each produces findings with the passages attached.

  1. 01:

    Outdated.

    Two signals: the inventory's review date has passed, or the content gives itself away by citing a retired policy, a departed team, a superseded form number or a system the company no longer runs. Both are findings, with the passage quoted.

  2. 02:

    Duplicates and near-duplicates.

    The same policy in three drives, two versions of one template, a PDF and a Word copy that drifted apart. The review groups them and names the differences that matter: dates, owners, clauses, obligations.

  3. 03:

    Conflicts.

    Two documents that state different rules for the same thing: expense limits, approval thresholds, notice periods. The review quotes both passages side by side and names the documents. Which rule applies is the owner's call.

The review queue: findings go to owners, and a person decides.

Every finding lands in a queue with the cited passages, the documents' locations and their owners. Nothing is edited, merged or deleted automatically; the queue exists to move each decision to the person accountable for that document. Our guide to human-in-the-loop AI workflows covers the same pattern across other AI work.

Keep the human role explicit. NIST AI 600-1, the Generative AI Profile of the AI Risk Management Framework (July 2024), warns in its Human-AI Configuration section that people may over-rely on generative AI systems or perceive their content as higher quality than content from other sources, an example of what the profile calls automation bias, excessive deference to automated systems. Quoted passages let the owner check each finding against the source text before acting on it.

Write down who decides per document type before the first findings arrive. NIST AI 600-1's inventory guidance for generative AI systems (action GV-1.6-003) lists human oversight roles and responsibilities among the items to consider in each system's inventory entry.

Arabic and English versions of the same policy.

Organizations that hold each policy in two languages get a fourth check: compare the Arabic and English documents of the same policy against each other, as well as against the rest of the set. Where they differ on a rule, a threshold, an obligation or an effective date, flag the pair with both passages quoted, each in the language it is written in.

Which version prevails is a question for your policy framework and your counsel. If the framework is silent, that silence is a finding of its own: queue it for whoever owns policy standards, and settle the controlling version once, in writing.

From one-off audit to standing check.

Documents keep changing after the first report, so its findings go stale. Re-run the three checks on a schedule tied to how fast the library moves, and let each run compare against the inventory: review dates that came due, documents added since, conflicts the previous pass left open.

The standing version is software: a pipeline that reads new and changed documents, a queue that routes findings, and a re-run that measures what the last pass closed. For that build, out:grow embeds forward-deployed AI engineers inside enterprise and government teams to build AI agents, automations and integrations in your own environment, with your team able to run and extend the result after we leave.

Do it yourself, or bring us in.

Everything above can be run in-house by the leads who already own the drives: build the inventory with owners and review dates, set the access rule, run the three checks, and stand up the review queue. Scope the first pass to one department and a handful of document types, and widen it on what the first findings teach you.

Outside help fits when the drives span several systems, or when the first pass needs to land while the team keeps its normal work. A forward-deployed engineer from out:grow embeds with your team, learns your systems, and builds the thing: the review pipeline, the queue and the re-run, wired into your stack.

Book a 30-minute call

Not ready for a call? Take the 10-question AI readiness check

Frequently asked questions.

Can AI decide which version of a document is current?

No. The review surfaces the candidates with the passages quoted; the document's owner, or the approver named in the inventory, decides which version stands. NIST AI 600-1 names over-reliance on generative AI output as a risk, which is why findings carry citations instead of verdicts.

What about scanned PDFs?

A scan with no text layer is an image to the review; it has to pass through OCR before its text can be compared, and OCR output deserves a spot-check against the page images. Some scanned files already carry a hidden text layer, and those read as text straight away.

How do we keep restricted files out of the results?

Three layers: run the review under an identity whose permissions match the report's audience, keep restricted libraries out of the index scope, and if you build a retrieval index, make it permission-aware and partitioned, as OWASP's LLM08:2025 entry recommends.

How long does a first pass take?

It depends on the number of documents, the file types and how complete the inventory is; no figure holds across organizations, and we give none. Scope the first pass narrowly and let its findings tell you the shape of the next.

Who signs off a change?

The document's owner, and where the inventory names an approver, that approver. Have the queue record who closed each finding and when, so the sign-off trail stays with the documents.

Keep reading.

:

What data employees can put into AI tools: a classification guide.

:

Shadow AI: how to find the AI tools your staff already use.

:

How embedded AI engineers work inside enterprise teams.