Which internal documents are safe for AI ingestion?
Classify internal documents before AI ingestion with a reusable triage matrix, clear approval owners, exclusion rules and a practical non-AI alternative.
Stanic & Co.
Perspectives for business leaders
Classify documents before connecting an AI knowledge assistant: Green for approved public or non-sensitive material, Amber for restricted material needing an approved, sanitised copy, and Red for personal, sensitive or unresolved material kept out of the pilot. Apply these labels to individual documents, not whole folders, and isolate approved copies in a separate ingestion library. This is a proposed conservative operating policy—not a legal classification or proof that a document is safe.
The practical decision is not simply whether a file contains personal information. It is whether that particular version can be made available to the assistant’s intended users, for the approved purpose, with restrictions preserved. The workflow below helps an Australian business make that decision before switching on a connector.
Classify documents for AI ingestion before connecting the library
Start with a defined audience and task. For example: help all staff find approved operating procedures, without answering questions about individual customers or employees. A document suitable for a finance-only assistant is not automatically suitable for an assistant available to every employee. Treat classification as a decision about a document, an audience and a use—not a permanent property of the file.
The OAIC’s guidance on commercially available AI products calls for due diligence covering intended uses, human oversight, privacy and security risks, and who can access information entered or generated. It also recommends considering whether using personal information with AI is necessary and the best solution in the circumstances.
Use Green to mean eligible for direct ingestion after approval, not automatically approved. Public product instructions may qualify, but an obsolete version or a document with unapproved embedded material should still be withheld. Use Amber for internal or restricted documents that might yield a useful approved extract. Red means manual reference only within this proposed pilot, including anything whose ownership, audience or restrictions remain unresolved.
Personal information is Red by default under this conservative approach. That is not a claim that Australian privacy law prohibits every organisational AI use involving personal information. The OAIC specifically recommends against entering personal information, particularly sensitive information, into publicly available generative AI tools. Other arrangements still require assessment; an enterprise subscription is not an ingestion approval.
This article addresses the operational sorting of an existing library. Use the separate privacy-first due diligence audit to assess the proposed product and processing arrangements before releasing any document set.
Give one owner authority to release the approved collection
Appoint a business knowledge owner who can approve the assistant’s scope and accept the operational consequences of exclusions. Department document owners should confirm what their files contain and who may use them. A systems administrator should implement the approved boundary, rather than decide which business information is appropriate to disclose.
Nominate a privacy or risk lead to resolve uncertain cases, with qualified legal advice where needed. Questions for that adviser include whether the organisation’s privacy obligations apply, whether the proposed use fits the collection purpose, what contractual confidentiality restrictions exist, and whether a proposed sanitised extract remains personal information. Do not ask a reviewer to make unsupported legal conclusions from a traffic-light label.
Create an inventory before enabling automatic synchronisation. Record the document identifier, location, owner, version, intended audience and proposed purpose. Include attachments, spreadsheets, scanned files and linked material in the inspection scope. Where these cannot be examined reliably, mark the item unresolved and keep it out rather than assuming that the visible first page represents the whole file.
Keep the original repository separate from the proposed AI ingestion library. The latter should contain only specifically approved copies or extracts. Give the integration access to that collection, not a broad parent folder, where the selected system supports that configuration. Ask the implementer to demonstrate the actual scope; a folder name such as “AI safe” is not access control.
This separation is a proposed privacy-by-design procedure. The joint statement published by the OAIC on trustworthy data governance supports embedding privacy safeguards from initial planning and maintaining technical and procedural controls throughout an AI system’s lifecycle. It does not prescribe this particular folder structure.
Use the Knowledge Triage Matrix
Copy the following record into a spreadsheet or document register, with one entry per document version. This matrix and its pass/fail conditions are illustrative operating procedures, not rules prescribed by the sources. Use “personal information” rather than relying only on the narrower shorthand “PII”.
Document title and identifier: Source location and version: Business owner: Proposed assistant purpose and audience: Data sensitivity level: Public / Internal / Restricted / Sensitive / Unknown Personal information present: Yes / No / Uncertain Specific legal, privacy or contractual restrictions: Yes / No / Uncertain; describe evidence Attachments, comments and embedded content inspected: Yes / No Proposed classification: Green / Amber / Red Action: Ingest directly / Redact and ingest / Manual reference only Approved copy location and permitted audience: Approver, approval date and review trigger:
Green — Ingest directly. Pass only when the owner confirms the purpose and audience, inspection finds no personal or sensitive content, restrictions permit the proposed use, and the exact version is approved. Example entry: “Published equipment care guide; Public; personal information No; restrictions reviewed and compatible; current version; owner approved.” A missing owner, unknown restriction or uninspected attachment fails the release check.
Amber — Redact and ingest. Use this route where a useful, restricted document can become a separately approved extract. Example entry: “Internal service procedure; Restricted; personal information Yes in worked examples; confidentiality review required.” Keep the original out. Release the derivative only after its content, remaining contextual identifiers, audience and restrictions pass review. Removing names alone is not a sufficient pass condition.
Red — Manual reference only. Example entry: “Customer hardship applications; Sensitive; personal information Yes; customer-specific financial and health details; exclude from pilot.” Apply the same stop condition to unknown ownership, uncertain restrictions, failed sanitisation or material outside the approved purpose. Any “Uncertain” answer blocks release until resolved. Where signals conflict, use the more restrictive outcome and record why.
Inspect the document, then isolate the approved version
Review content in context, not just by searching for names or email addresses. Ask whether an example, combination of facts or small group description could identify someone. Also inspect comments, revision history, spreadsheet tabs, attachments and embedded objects using suitable tools. These are recommended inspection targets; whether a particular connector processes them depends on its configuration and supported formats.
For Amber documents, create a derivative instead of altering the only business record. Replace customer-specific examples with genuinely fictional examples, remove material irrelevant to the assistant’s task, and preserve the original in its existing authorised location. Record what was removed and why without copying the excluded personal details into a broadly accessible triage register.
Do not describe a derivative as de-identified solely because obvious identifiers were removed. A reviewer should assess whether the remaining information could identify a person in its intended environment. The OAIC’s commercial AI guidance makes clear that personal information can include inferred or incorrect information about an identified or reasonably identifiable individual.
Before release, have the document owner check that sanitisation has not changed the business instruction. Removing an exception or qualifying condition may make a procedure misleading. Have a separate reviewer challenge the privacy and audience decision for higher-risk extracts. Human review can miss contextual identifiers, misunderstand restrictions or approve the wrong version; it reduces uncertainty but does not guarantee correctness.
Publish only the approved derivative and retain a traceable link to its approval record. Confirm that excluded originals are outside the integration’s reach. For the detailed content-cleaning stage, the complementary guide on filtering personal information before ingestion provides a narrower workflow. Classification decides the route; filtering does not replace that decision.
Test the boundary with a hypothetical hardship-folder failure
Consider a hypothetical business preparing an all-staff knowledge assistant. Its operations folder contains a blank hardship form, an approved hardship policy and completed customer applications. Someone labels the whole folder “customer service procedures” and connects it. During testing, an employee asks for examples of hardship decisions and receives a named customer’s financial circumstances despite having no authority to see that application.
This is a hypothetical disclosure failure, not a reported Stanic & Co. client incident. It is also distinct from the OAIC’s workplace GenAI case study, which describes a fictional insurer uploading a hardship application to a public tool and relying on an inadequate summary. Both scenarios show why the intended purpose does not make all supporting records suitable inputs.
Under the proposed matrix, the approved policy could be Green after inspection. The blank form might be Green or Amber depending on its content and restrictions. Completed applications would be Red for this pilot. If staff need examples, the document owner could prepare fictional cases that teach the process without drawing on identifiable customer circumstances.
Test this boundary before launch using synthetic documents and test accounts representing the intended roles. Place a fictional restricted application outside the approved collection and verify that neither its content nor restricted metadata appears in answers, search results or citations. Ask both direct questions and indirect requests such as “give me a real example”. A passed test is evidence about those cases, not proof that every boundary holds.
If a restricted item appears, stop access to the affected assistant or collection, preserve necessary incident evidence with restricted access, and involve the privacy and security leads. Investigate connector scope, permissions and copied content. Do not assume deleting the source removes indexed copies, conversation history or logs. Establish the product-specific removal process, assess any incident obligations with qualified advice, and retest before reopening.
Choose a human-managed FAQ when the library is not ready
A structured, human-managed FAQ on a secure non-AI intranet is a credible alternative. Publish approved answers, nominate an owner for each topic, provide links to authorised source documents, and give staff read-only access appropriate to their role. Keep individual customer and employee records in their existing controlled systems.
Prefer this option when the useful knowledge set is small, most source documents are restricted, ownership is unclear, or staff mainly need a stable set of procedural answers. It is also the safer interim choice when the proposed integration cannot demonstrate a boundary between approved derivatives and excluded originals. Do not connect the whole library merely to avoid the classification work.
The FAQ still needs content review and access controls. A mistaken answer can be published manually, and a poorly configured intranet can expose information too. Its practical advantage for this decision is that the business explicitly publishes each answer rather than approving a wider collection for generated responses. Staff can see the approved wording and escalation route.
The trade-off is ongoing editorial work and less flexibility for unexpected questions. An AI assistant may offer a more convenient way to navigate a larger approved collection, but classification, derivative creation, exception handling and monitoring remain operating costs. Compare both approaches on the same set of staff questions, including the effort to maintain answers—not just the speed of producing them.
Government support for governance should not be mistaken for a recommendation to deploy AI in every case. The business.gov.au guidance announcement describes practices and policy and register templates to support responsible adoption. Use those tools to record a decision to defer AI just as clearly as a decision to proceed.
Approve a bounded collection, not an entire knowledge estate
Make the release decision explicit: approve only Green documents and approved Amber derivatives for the named audience and purpose. Keep Red and unresolved items excluded. The business knowledge owner signs off the scope, document owners approve content, and the systems administrator supplies evidence that the implemented connection matches the approved inventory.
Before launch, reconcile the ingestion inventory against the release register. Investigate every unexpected item. Check example answers and citations with authorised and unauthorised test roles, and confirm that users have a clear route to report a suspect answer. If the product cannot expose enough information to verify what has been ingested, narrow the pilot or retain the FAQ alternative.
Reopen classification when a document changes, its audience changes, a new attachment appears, or the assistant’s purpose expands. These are proposed review triggers, not statutory intervals. The OAIC’s commercial AI guidance says due diligence should not be set-and-forget and calls for ongoing review, staff training and monitoring throughout the product lifecycle.
Maintain a reversal record: affected document identifiers, approved-copy locations, ingestion status, removal steps, retest evidence and the person authorised to restore access. Preserve excluded business originals according to the organisation’s applicable recordkeeping requirements rather than deleting them simply to simplify an AI pilot. Seek advice where retention and removal requirements are uncertain.
The decision is ready when the business can explain which exact documents are included, why each is suitable, who may receive their content, and how a mistaken inclusion will be contained. If those answers depend on assumptions about folder names, reviewer vigilance or subscription labels, the collection is not ready. Start with a smaller approved set—or publish the FAQ first.
Sources
- OAIC: Guidance on privacy and the use of commercially available AI products
- OAIC: GenAI tools in the workplace—balancing protection of personal information and business efficiency
- OAIC: Joint statement on trustworthy data governance frameworks for AI
- business.gov.au: New guidance helps Australian businesses adopt AI safely and responsibly
General information only; verify sources before making business decisions.