Back to the reading room
Business insights

How to check staff keep personal information out of public AI

A practical input-check workflow for Australian teams: identify personal information, stop unsafe public AI use, record decisions and handle mistakes.

Stanic & Co.

Perspectives for business leaders

11 min read

An Australian business owner can check daily AI use by requiring a brief input check before staff paste, upload or connect information: identify personal information, confirm the approved tool and account, and stop if either is uncertain. For public generative AI, personal information means stop and use the manual process. Support that check with practical assessments and targeted reviews—not a promise that staff declarations guarantee compliance.

This is a proposed operating procedure for implementing the OAIC’s best-practice recommendation against entering personal information into publicly available generative AI tools. It is not an OAIC-prescribed checklist or a legal compliance certificate. The narrow question is how to make that boundary observable at the moment an employee uses AI.

Checking personal information before staff use public AI

Start with a business rule employees can apply without interpreting supplier contracts: no personal information in public generative AI tools. If a worker cannot establish whether the input contains personal information, or whether the particular tool and account are approved, they stop. The business—not the employee under deadline pressure—must resolve the uncertainty.

The OAIC recommendation is a best-practice position, not a statement that every public AI interaction is unlawful. For organisations subject to the Australian Privacy Principles, the guidance also explains that privacy obligations apply to personal information entered into AI and to generated outputs containing personal information. Ask a qualified privacy adviser about your business’s obligations rather than treating this workflow as a legal determination.

The useful unit of review is the complete material being submitted. A name field may be obvious, but a paragraph, spreadsheet attachment or connected document can also contain records linked to an identifiable person. Reviewing individual fields is helpful; checking only the highlighted text while ignoring the attached workbook is not.

Build a short approved-use register before asking staff to classify tools. Record the product, permitted account or workspace, allowed task, allowed information and decision owner. An entry should distinguish a business-approved workspace from a personal account using the same product name. If the entry is missing or unclear, staff should not infer approval from branding.

This article addresses the frontline check, not the entire training programme or supplier assessment. Use the existing AI staff training plan for broader role-based practice. Here, success means employees can demonstrate what they checked, why they stopped or proceeded, and which fallback they used.

The reusable Personal Info Red-Light Decision Tool

The following is an illustrative business procedure, not a test prescribed by the OAIC. Copy it into a task form or checklist. Assess the whole submission, including attachments and information the tool will retrieve. A pass authorises only the specified task; it does not make the tool suitable for every purpose.

Record these fields before submission:

  • Task and employee: what are you trying to produce?
  • Input: which text, fields, files or connected sources will be supplied?
  • Personal information: yes, no or uncertain, with a brief reason.
  • Tool route: public, approved restricted workspace or uncertain.
  • Approval reference: the relevant register entry and permitted data category.
  • Decision: proceed, stop or refer, plus the manual fallback.

Record categories and reasons, not copies of the personal information you are trying to protect.

Apply the stop conditions:

  • Personal information plus public generative AI: STOP. Use the non-AI process.
  • Uncertain personal information or uncertain tool/account status: STOP and refer to the named owner.
  • Personal information plus an approved restricted workspace: REFER to the documented permission and privacy impact assessment. Proceed only if this exact use and information category are authorised.
  • No personal information: proceed only if the tool, task and other confidentiality requirements are approved.

A restricted or enterprise-labelled product is not an automatic pass.

Example entry—stop: “Draft outreach; customer email addresses and purchase notes; personal information: yes; public chatbot; decision: stop; fallback: approved local mail-merge template.” Example entry—proceed: “Draft generic opening paragraph; invented product description, no customer records; personal information: no; approved tool/account; decision: proceed for generic copy only.” These are illustrative decisions, not findings about any named product.

Example entry—refer: “Summarise a complaint; name removed but detailed transaction history remains; personal information: uncertain; restricted workspace; decision: stop pending privacy-owner review.” Removing a name does not resolve whether someone remains reasonably identifiable. The OAIC guidance covers information about identified or reasonably identifiable individuals, including inferred or inaccurate AI-generated information.

Teach recognition, then verify decisions rather than attendance

Train staff on the materials they actually handle: customer messages, sales exports, complaint notes and screenshots. Use invented examples without real personal information. Ask each employee to identify the complete proposed input, select the decision and explain the fallback. A worker who can recite “do not share customer data” but overlooks an attachment has not demonstrated the required skill.

An illustrative assessment can include a generic product brief, a client email list, a complaint with the name removed, and an apparently harmless screenshot containing contact details. Make uncertainty an acceptable answer when followed by stopping and referral. Do not reward confident guesses or treat completing more AI tasks as evidence of responsible use.

For the initial assessment, a proposed pass condition is that the employee stops every public-tool case containing personal information and escalates every unresolved classification. Adapt the examples to the role and risk. This is a business-designed assessment threshold, not an OAIC requirement. Someone who fails should practise the missed cases before using AI independently for that workflow.

During normal work, retain a lightweight decision record and review selected records against the underlying task in an authorised environment. Check whether staff identified all inputs, used the registered account and followed stop decisions. Avoid copying sensitive source material into a new monitoring repository merely to prove that a check occurred.

The OAIC calls for regular reviews of product performance, staff training and monitoring throughout the AI product lifecycle in its commercial AI guidance. That supports ongoing verification, but does not prescribe this particular assessment. Reviews provide evidence about observed behaviour; they cannot establish that every unobserved interaction was safe.

A hypothetical failure: the marketing list that should never be pasted

Consider a hypothetical marketing coordinator preparing personalised outreach. They paste customer email addresses and purchase notes into a public chatbot and ask for individual messages. They believe the chatbot “forgets everything” when the conversation closes. This is an invented failure scenario, not a Stanic & Co. client experience or a claim about a particular supplier.

The red-light check should have stopped the task before submission: the list contains personal information and the destination is a public generative AI tool. The failure is the submission itself, not merely whether the supplier later uses the material to train a model. The OAIC’s recommendation addresses significant and complex privacy risks; it is not limited to model training.

A proposed detection route is a task review in which the decision record says “generic copy only” but the work product contains customer-specific messages. The reviewer asks what information was supplied and discovers the list. This discrepancy is a useful investigation signal, not proof by itself. Staff should be able to report an accidental submission promptly without first proving that harm occurred.

The immediate response should be to stop further use, prevent the generated messages from being distributed and notify the business’s privacy or incident owner. Record the tool, account, time, information categories and known recipients or access arrangements without unnecessarily duplicating the list. The owner then follows the established incident process and seeks qualified advice about containment and any applicable notification obligations.

Do not promise reversal by deleting the visible chat. Seek supplier-specific evidence about available deletion or containment options and document what remains unknown. Correct the workflow before restarting: generate generic copy from an invented brief, then personalise it through the approved non-AI process. Reassess the employee on the failed decision and check whether unclear instructions or an impractical fallback contributed.

Use the simpler alternative when personalisation needs customer data

For repeatable outreach, the simpler alternative is an approved document template or spreadsheet-based merge that operates within the business’s controlled storage environment. Staff select approved wording and combine it with customer fields locally. Configure the process so it does not send the list to an AI service or an unapproved external connector.

This alternative is particularly suitable when the task requires customer details but the writing is mostly standard. It also fits teams that cannot reliably distinguish personal accounts from approved workspaces. If the business cannot provide a clear tool register, an available escalation owner and a usable fallback, do not authorise public AI for workflows involving customer records.

The trade-off is less flexible wording and some template maintenance. Staff may also need to check spreadsheet columns, recipients and merge results. A non-AI process does not eliminate privacy or delivery mistakes; it avoids the particular disclosure involved in sending the customer list to a generative AI service. Test it using invented records before live use.

A split workflow can preserve limited AI assistance: ask an approved tool to draft a generic message from a fictional product brief, review the wording, then add customer-specific fields through the controlled merge. Keep real customer material out of both the prompt and any follow-up request. Do not paste a completed personalised message back into public AI for polishing.

Do not use this lightweight approach as the sole safeguard for AI decisions affecting individual rights, sensitive-information workflows or broad access to business records. The OAIC identifies decisions with legal or similarly significant effects on individuals as likely high privacy risk. Those uses need a separate suitability and privacy assessment; the frontline checklist cannot authorise them.

Assign ownership and avoid misleading assurances about tools

The business owner should approve the boundary and provide resources for the fallback. A named privacy lead should decide ambiguous information cases and maintain the relevant assessment. The IT or system owner should maintain the approved tool/account register. Team leaders should organise practice and review task records. Employees own the immediate stop decision, not the interpretation of complex supplier terms.

For personal information in a restricted workspace, the privacy lead should check whether the documented assessment covers the purpose, information category, access arrangements and human oversight. The OAIC recommends product due diligence and privacy by design, including a privacy impact assessment. Its guidance also explains purpose-related constraints on using or disclosing personal information under APP 6. A paid subscription does not settle those questions.

Avoid teaching “no model training” as shorthand for “safe to paste”. One supplied Microsoft-hosted Q&A about Copilot privacy settings contains an AI-generated answer and an explicit warning that it may be incorrect. It is insufficient evidence to approve a product or account. This article therefore makes no capability assurance based on that answer.

Similarly, do not justify submitting a customer record because some details are already online. The OAIC’s separate guidance on developing and training generative AI models says public availability does not by itself make personal information lawful to use for model training. That guidance concerns model development, not staff training; its relevance here is to reject an overly broad “public means unrestricted” assumption.

Where supplier evidence or permitted use is unresolved, keep the workflow stopped and use the manual alternative. For the wider selection process, use the privacy-first knowledge assistant audit. Do not make frontline staff compensate for an unfinished product assessment.

What an owner can conclude—and what remains unverified

A defensible operating decision is narrower than “our team complies”. You can conclude that the business has defined a boundary, trained employees to apply it, examined selected decisions and acted on observed failures. Whether those measures are legally sufficient depends on the circumstances and applicable obligations. This educational article does not determine that question.

Human review has limits. Reviewers can miss identifiers, misunderstand context or approve a familiar-looking task too quickly. Employees may forget an attachment or switch accounts after completing the check. A signed declaration is evidence of a decision, not proof of what was transmitted. Combine practice, clear tool configuration and targeted review rather than relying on acknowledgement forms alone.

Review outputs as well as inputs. The OAIC explains that inferred, incorrect or artificially generated information about an identified or reasonably identifiable person still constitutes personal information. If a generic draft unexpectedly includes claims about a real person, do not circulate it simply because no personal information was intentionally supplied. Refer it for assessment and accuracy checking.

Use review findings to distinguish individual mistakes from design failures. Repeated referrals about account status suggest a poor register. Repeated unsafe submissions under deadline pressure suggest an unusable fallback or weak supervision. Make the correction specific: clarify the permitted route, simplify the task, remove an unapproved connection or suspend the affected workflow until the owner resolves the problem.

Start with one everyday task rather than rolling out a declaration across the whole business. Establish its approved route, practise the red-light decisions and confirm that the non-AI fallback works. Expand only when staff can demonstrate the boundary. The aim is not to approve more AI use; it is to make “stop” a practical, supported decision before personal information leaves the business.

Sources

General information only; verify sources before making business decisions.

Turn an idea into a next step.

A conversation about what matters to your business.

Let's talk