How to filter personal information before an AI knowledge assistant
A practical workflow for filtering personal information before AI ingestion, with release checks, clear ownership and a simpler curated-FAQ alternative.
Stanic & Co.
Perspectives for business leaders
Filter personal information before your AI knowledge assistant ingests it: select only necessary sources, remove unnecessary fields, sanitise the remaining text, then test the prepared files and answers before release. Quarantine uncertain material rather than uploading it. Automated masking should support—not replace—source selection and review. For businesses covered by the Privacy Act, privacy obligations extend to personal information in both AI inputs and outputs, as explained in the OAIC’s guidance on commercially available AI products.
The practical decision is not simply which redaction tool to buy. It is whether your business can maintain an approved, useful knowledge collection without exposing customer or staff records. This article proposes an illustrative pre-ingestion workflow for Australian businesses, based on the supplied evidence reviewed for 11 October 2026. It is educational guidance, not a legal determination or a guarantee of de-identification.
Filter personal information at the source, not just in answers
Start with the assistant’s permitted purpose. “Help staff explain our returns process” requires different material from “summarise individual customer disputes”. For the first purpose, approved policies and general troubleshooting notes may be sufficient. A complete customer-support export should not be the default input. Write down what questions the assistant should answer, what information it needs, and what it must leave to an authorised person.
A retrieval-augmented generation, or RAG, assistant retrieves material to help generate answers; this is distinct from training a model on those records. Microsoft’s documentation for its classic Azure OpenAI On Your Data service describes ingestion that cracks, chunks and embeds files into an Azure AI Search index. That is a vendor-specific example of why filtering must happen before indexing, rather than only after an answer is drafted. The Microsoft documentation also states that this particular service is approaching retirement on 14 October 2026; it should not be treated as a recommendation for a new deployment.
Use three proposed source categories: approved general knowledge; records requiring transformation; and excluded records. A published product manual might fit the first. Support notes may need a generalised summary. Medical attachments, identity documents and detailed staff case files should remain excluded unless a separately assessed purpose requires them. These are conservative design choices, not categories prescribed by the sources.
Ask a qualified privacy adviser whether the intended use is permitted and whether your organisation is covered by the relevant obligations. Removing names is not a substitute for that assessment. The OAIC advises businesses to consider whether using personal information in an AI system is necessary and the best solution, and to examine access, oversight and security during product selection. See its AI product selection and privacy guidance.
Build a sanitisation process that preserves useful knowledge
The following process is illustrative. Create a restricted preparation area between the original systems and the assistant’s ingestion service. Copy only approved source material into it. Keep originals under their existing access controls; do not edit the business record merely to make an AI-ready copy. Record each copy’s source, version, preparation date and responsible owner so you can identify what entered a release.
Remove unnecessary structured fields before processing narrative text. For example, exclude customer names, email addresses, telephone numbers and account identifiers from an export when they serve no knowledge purpose. Then inspect free text, attachments and document metadata. A clean customer-name column does not make a complaint description safe: the narrative may repeat the name or describe an identifiable person’s circumstances.
Configure automated detection around your actual sources. Proposed checks might include email and telephone patterns, customer identifiers, names, addresses and references to health or financial circumstances. Where the document’s useful content survives, replace the detail with a neutral label such as “[customer]”. Where several details together identify someone, rewrite the passage as general procedural knowledge or exclude it. A stable placeholder such as “Customer 472” should not automatically be labelled anonymous: ask whether it remains linkable to a person.
Test usefulness as well as removal. “Customer called about a damaged item; staff should follow the damage-assessment procedure” may preserve knowledge. A heavily masked transcript may not. If sanitisation removes the reason an answer is valid, create a reviewed FAQ instead. The OAIC’s guidance for AI developers warns that steps to remove or de-identify personal information may not always be effective. Treat masking as a control to evaluate, not proof that the resulting dataset is anonymous.
Reusable data sanitisation release checklist
Copy the following checklist into each ingestion approval. It is a proposed evaluation protocol, not a legal compliance test. For every gate, record: evidence location; accountable owner; reviewer; pass, fail or unresolved; and corrective action. An unresolved gate means hold the affected material outside ingestion. The business owner should approve the scope, the information owner should approve its content, and the technical lead should provide evidence of what was actually processed.
Gate 1 — Audit source data. Input: source inventory and assistant purpose. Example entry: “Returns policy approved; support transcripts require transformation; customer attachments excluded.” Pass when every included source has an owner, a stated knowledge purpose and a defined extraction boundary. Fail when an export contains unknown attachments, uncontrolled folders or sources nobody can authorise.
Gate 2 — Categorise personal information. Input: field list and representative narrative examples. Example entry: “Names and contact fields removed; health references and identifying case descriptions quarantined.” Pass when structured fields and free-text risks have documented treatment. Fail when the team cannot explain how unusual formats or sensitive passages will be handled.
Gate 3 — Apply and verify sanitisation. Input: preparation configuration, transformed files and exception report. Example entry: “Contact-pattern checks completed; unresolved passages held for review.” Pass when the intended controls ran, exceptions are resolved or excluded, and the prepared copy remains useful. Fail if the process silently skips a file or leaves detectable personal details.
Gate 4 — Check authority and consent records. Input: documented purpose assessment, relevant notices and any consent or opt-out records. Example entry: “General policy material only; customer case reuse not approved.” Pass when the information owner and appropriate adviser have resolved the basis for inclusion and any applicable exclusions. Fail when permission is assumed merely because the business already holds the data.
Gate 5 — Validate residual risk. Input: prepared-file review and assistant test results. Include questions seeking names, contact details, individual complaints and combinations of identifying circumstances. Pass only when the release meets the business’s documented acceptance criteria, with no known unresolved disclosure in the tested material. Stop if any prohibited personal information appears in retrieved passages, citations or answers. Record the release decision, excluded sources and recovery procedure. A clean sample is evidence about that sample—not assurance that every possible prompt or document is safe.
Test the prepared data and the assistant separately
Pre-ingestion inspection and answer testing serve different purposes. Inspect prepared files to check whether the sanitisation process removed what it was meant to remove. Then test the assistant to examine what it retrieves and discloses. Reviewing only polished answers can miss personal information still present in the underlying collection. Reviewing only files can miss an answer that combines separate passages into an identifying account.
Use synthetic test records with known markers before handling real records. For example, create an invented complaint containing a name, contact address, account reference and health detail, then check whether each expected control responds. Include variants such as a detail repeated in a signature or placed in an attachment. These are proposed test cases, not claims about a particular detection tool. Record expected results so a reviewer does not decide success simply because an output looks plausible.
For real prepared material, prioritise higher-risk sources and inspect uncertain passages in context. Test direct requests—“Give me the complainant’s contact details”—and apparently ordinary requests—“Summarise recent complaints”. Also check accessible citations and retrieved excerpts. Define the sample and repeat-test policy according to the volume, variability and consequences of disclosure. No universal sample size is supported by the supplied evidence, and no pass rate should be presented as a compliance threshold.
Human review has limits: a reviewer may miss context, overlook indirect identification or approve a familiar-looking document too quickly. Use explicit criteria, independent checks for higher-risk material and escalation for uncertainty, rather than a general “reviewed” tick. The OAIC calls for ongoing review, staff training and monitoring rather than set-and-forget adoption in its commercial AI guidance. Once the filtering boundary is established, the complementary RAG readiness guide covers broader questions such as citations, freshness and missing answers.
Hypothetical failure: a support export exposes a customer
Consider a hypothetical retailer building an internal assistant to explain complaint-handling procedures. Its team uploads a full support-log export after removing the dedicated name and email columns. The narrative still contains a customer’s full name, telephone number and medical circumstances explaining why a delivery mattered. Nobody inspects attachments or tests broad summary questions. This is an invented example, not a Stanic & Co. client experience.
An employee without permission to see that case asks for a summary of recent complaints. The assistant retrieves the unsanitised passage and includes those details in its response. The failure began before generation: the team equated column removal with whole-record sanitisation and used transactional records where general knowledge would have served the purpose. An instruction telling the assistant not to reveal personal information would not establish that the source collection was safe.
Under the proposed response procedure, the operator stops access to the affected collection, preserves appropriate incident evidence under restricted access and alerts the business’s incident and privacy owners. They identify the source release, remove affected material from the prepared collection and deployed index, and examine any retained copies, caches or conversation histories within their actual system. Rebuild and retest before restoring access. Do not assume deleting the original uploaded file reverses every downstream copy.
The legal response depends on the facts. The OAIC’s notifiable data breaches guidance says that organisations covered by the Privacy Act must notify affected individuals and the OAIC when a personal-information breach is likely to result in serious harm. This hypothetical disclosure calls for prompt assessment, not an automatic conclusion that notification is mandatory. Qualified advisers should help assess coverage, the information involved, who received it and the likely harm. Technical containment and legal assessment should proceed together.
Choose a curated knowledge repository when filtering is too uncertain
The simpler alternative is an access-controlled, knowledge-only repository of approved FAQs, procedures and troubleshooting summaries. Staff search or browse those documents without a generative assistant. An authorised person handles requests for individual records through the existing business system. This separates general guidance from case access and avoids making raw customer histories the foundation of everyday knowledge search.
For example, replace dozens of complaint transcripts with an approved note explaining how to handle delayed deliveries, which exceptions need manager approval and where staff can find the current procedure. Do not include colourful customer anecdotes merely to make the note readable. Where a specific record is genuinely needed, require an authorised person to verify the requester’s need before providing access; do not embed a broadly accessible case summary in the FAQ.
This approach trades flexible questioning for editorial effort. Someone must maintain the summaries, resolve conflicting procedures and retire obsolete material. Staff may need more clicks to find an answer. An AI assistant also requires maintenance, plus sanitisation, ingestion checks and answer evaluation. Compare the total operating burden against the actual problem: if staff chiefly need a small set of stable procedures, ordinary search may be sufficient.
Do not proceed with transcript-based ingestion when you lack a source owner, cannot inspect exports, cannot maintain exclusion records or cannot remove a faulty release. Also pause when the intended answers depend on sensitive individual details and the proposed general-purpose assistant has not been separately assessed for that use. A curated repository is not automatically compliant either; it still needs appropriate content and access controls. Its advantage here is a smaller, more inspectable information boundary—not a guarantee.
Assign ownership and make every release reversible
The business sponsor owns the decision to use AI rather than the simpler alternative. The information owner decides which sources are appropriate and whether sanitised summaries retain their meaning. The technical lead implements preparation, exception handling, indexing and removal. The privacy lead or qualified adviser addresses coverage, permitted use, notices and incident assessment. Name these roles even if one person holds several; the tool supplier should not implicitly become the decision-maker.
Maintain a release record linking approved sources to prepared copies and the deployed collection. Keep the sanitisation configuration, unresolved exclusions, test evidence and approval decision together. Limit access to records that contain removed content: a redaction report that reproduces every sensitive passage needs protection too. This is a proposed operating design; confirm how your chosen platform stores ingestion artefacts, logs and histories rather than assuming they disappear after processing.
Use clear stop conditions: newly discovered personal information, a changed source format that defeats extraction, an unreviewed source entering the pipeline, or inability to trace deployed material back to approval. Disable the affected collection or route staff to the curated repository while investigating. The complementary guide on pausing a knowledge assistant after a policy change addresses release control when the business information itself changes.
Approve the project only when you can answer four questions with evidence: Why is this information necessary? What was removed or excluded? How was the remaining risk tested? How will we withdraw a faulty release? Filtering personal information is an ongoing information-management task, not a one-off redaction exercise. Start with the smallest useful approved collection, expand only when the preparation and recovery process is workable, and retain human escalation for requests that properly belong in individual records.
Sources
General information only; verify sources before making business decisions.