Privacy impact assessment for websites and web applications: the OAIC's ten steps, and where test evidence fits
For owners, managers and technical leads planning a new form, client portal, chatbot or hosting move, or asked for a PIA by a customer or an agency.
This guide sets out the OAIC's method step by step, who has to carry out an assessment and who is encouraged to, and a worked data-flow canvas that traces an invented enquiry and client account record to show where a targeted test can inform a step.
In this guide
Worked website PIA data-flow canvas: one enquiry and one client account record
A worked example, not a template. It follows two records through an invented home renovation business, an enquiry from someone who has not hired it and the portal account record of a client who has, from collection to deletion for the enquiry and to account closure for the client record, a closure row D2 tries on a second invented client's account.
How to use itHide the instructions
The business, its systems, its settings and the one finding in row R1 are made up to show the method, and none of them describes a real business. Read each row left to right: the stage, what happens there in the example, the step of the OAIC's Guide to undertaking privacy impact assessments (publication date 2 September 2021) it feeds, privacy risk questions in the guide's words or a short form of them plus one this page adds, a targeted test where technical evidence helps, the limits on the row, and what a result can inform and cannot decide. To use it, replace the example column with your own project and keep the rest. Start with row P0, because later rows draw on its system list and the vendor terms it records. Rows P0, E2, E3, A5 and R1 read settings, lists, documents and logs and send no test requests. Rows E1, E4, E5, A1, A2, A3, A4, D1 and D2 send requests, sign in or change test records, and each one states its own limits: the business's own systems or ones it has written permission to test, the staging copy where there is one, invented people and test accounts for every test sign-in, request and change (row D1's read-only search, like row E2, uses staff accounts the business already has), a fixed count of requests with no load, messages routed to an inbox or number the business holds, and secrets and links that work for whoever holds them kept out of the record. Any of those nine rows that turns up a real person's information halts there, with the request and the response noted. Each result speaks for the build, accounts and records it used; it never settles that a risk is acceptable or that the PIA is finished. The PIA and every conclusion in it belong to the entity carrying out the project, working with the adviser it chooses.
14 rows
Flow and stage
In the invented example
P0. Before any test: every system the two records pass through
- In the invented example
- The invented business runs a website with an enquiry form from a form service, a customer relationship tool (CRM), an automation service that copies each enquiry into the CRM, an email delivery service, an analytics tag and an advertising tag, a chat assistant from an AI chat vendor, and a client portal its developer built, hosted with a cloud provider in a region outside Australia.
- PIA step it feeds
- Step 3, describe the project, and step 5, map information flows.
- Privacy risk questions to ask
- From step 5: describe the current personal information environment and how the project will affect it. From step 1, for an entity bound by the Privacy Act: the entity is responsible and accountable for the personal information it collects, even when external service providers or contractors in Australia or overseas hold it. This page adds: who runs each system, under which contract, and what does each vendor's documentation say it does with what it receives?
- Targeted test, where technical evidence helps
- No test. Build the list from invoices, subscriptions, the developer's list of integrations and the hosting account. Then read what each vendor's terms say about security testing and about data handling, and send the vendor a written question on any point they leave open.
- Limits on this row
- Read-only: reads settings, lists, documents or logs and sends no test requests. Any sign-in is a staff account the business already uses, used as it normally is. Open no screen or file that displays a password, key or other secret, and copy none. A vendor's written answer covers only what it says, and the later rows stay inside it.
- What the result can inform, and what it cannot decide
- A dated system list with each vendor's terms and answers, which later rows draw on. It cannot show what a vendor does inside its own systems. Whether a vendor acts on the business's behalf is for the business and its adviser.
E1. Enquiry: the form that collects it
- In the invented example
- Fields for name, email, phone, suburb, budget range and a message, an optional photo of the room, and a hidden field that records the advertising campaign the visitor arrived from. A link to the privacy policy sits under the submit button.
- PIA step it feeds
- Step 5, collection; step 6, the APP 3 and APP 5 questions.
- Privacy risk questions to ask
- From step 5, Collection: why the personal information, including the particular items and kinds of information, is necessary for the project; whether individuals can choose not to provide some or all of it; whether unsolicited personal information may be used in the project. From step 6, APP 5: if the information is collected directly from the individual, whether notice will be given, such as by displaying it on a form, providing a link on a web page or advising the individual over the phone, and the individual asked to confirm they have been notified of the APP 5 matters before providing their personal information. This page adds: does a photo need to be collected before a quote is agreed, and what else in a home could it show?
- Targeted test, where technical evidence helps
- Submit the form once as an invented person, with the marker PIA-CANVAS-7Q in the name and message fields, a small invented image, and an email address the business controls. In a fresh browser profile that is not signed in to any of the business's systems, with the browser's developer tools recording network requests, note every field the submission sent, hidden ones included. Then ask the developer for the stored record of that one submission and compare it with what the screen showed.
- Limits on this row
- One submission. Request row. Run it on a system your business owns or holds written permission to test, on the staging copy if there is one, and only with the invented clients, enquirers and staff accounts the canvas sets up: no real client, staff member or member of the public. Make the requests this row names and no others, with no load. Point every email or text it sets off at an inbox or number the business holds. Inside a vendor's service, stay within what its terms permit (row P0). Passwords, keys, session tokens and any link that works for whoever holds it stay out of the record, and no screen or file that shows a secret is opened. If a response carries a real person's information, halt the row, note the request and the response, look at nothing more, and hand the matter to the person who looks after privacy in the business.
- What the result can inform, and what it cannot decide
- A field list as built and as stored for one invented submission, the campaign value the hidden field kept, and the notice as a visitor saw it, all dated. It cannot decide whether a field is reasonably necessary or a notice adequate: that is for the business and its adviser.
E2. Enquiry: copies in the CRM and the automation service
- In the invented example
- The automation service copies each new entry into the CRM and tags it with the campaign. The form service keeps its own copy, and the automation service keeps a run history that shows the fields it moved.
- PIA step it feeds
- Step 5, use, disclosure, and retention and destruction.
- Privacy risk questions to ask
- From step 5, Use: all the planned uses of the personal information, including infrequent uses, and how they relate to the purpose of collection. From step 5, Disclosure: to whom, how and why it will be disclosed, and whether it will have the same privacy protections after it is disclosed. This page adds: how many copies does one enquiry leave, and who can read each?
- Targeted test, where technical evidence helps
- No new request. Sign in to each system on the P0 list that the business uses and search for the E1 marker. For each copy, record the system, who holds accounts that can read it, and the retention setting as shown.
- Limits on this row
- Read-only: reads settings, lists, documents or logs and sends no test requests. Any sign-in is a staff account the business already uses, used as it normally is. Open no screen or file that displays a password, key or other secret, and copy none. Search only for the marker, never for a real enquirer.
- What the result can inform, and what it cannot decide
- A count of the copies one enquiry produced on the systems searched, with who can read each and how long each is kept as the settings read. It cannot decide whether a use relates to the purpose of collection, or whether a disclosure is permitted.
E3. Enquiry: the emails it triggers
- In the invented example
- A notification to a shared office inbox repeats every field, photo included, and an automatic reply to the enquirer confirms receipt. Both leave through the email delivery service.
- PIA step it feeds
- Step 5, security.
- Privacy risk questions to ask
- From step 5, Security: how information will be transferred between sites; who will have access; how personal information will be protected if it will be managed by someone else. This page adds: does the notification need to carry the whole message and photo, or would a link that needs a sign-in carry less?
- Targeted test, where technical evidence helps
- No new request. Read the two emails E1 produced, in the shared inbox and in the test address, and record which fields each repeats and who can read the shared inbox. Read the delivery service's setting for how long it keeps message content or logs.
- Limits on this row
- Read-only: reads settings, lists, documents or logs and sends no test requests. Any sign-in is a staff account the business already uses, used as it normally is. Open no screen or file that displays a password, key or other secret, and copy none.
- What the result can inform, and what it cannot decide
- Which fields leave in email for one enquiry, who receives them, and how long the delivery service's setting says it keeps them. It cannot show how a recipient's own mail provider stores a message.
E4. Enquiry: analytics and advertising tags
- In the invented example
- Both tags load on every page, including the thank-you page shown after the form is sent.
- PIA step it feeds
- Step 5, disclosure; step 6, privacy impact analysis.
- Privacy risk questions to ask
- From step 5, Disclosure: to whom, how and why the personal information will be disclosed. From step 6, among its key questions: do individuals have to give up control of their personal information? This page adds: does any value a visitor typed, or the thank-you page address, reach a tag's host?
- Targeted test, where technical evidence helps
- In a fresh browser profile, with network recording on, load the form page, submit once with the E1 invented details and marker, and list every request to a host other than the site's own. Search those requests, and the page addresses they carry, for the marker. Where visitors can accept or refuse cookies or tracking, repeat the row once per option.
- Limits on this row
- One page load and one submission per cookie or tracking option the site offers. Request row. Run it on a system your business owns or holds written permission to test, on the staging copy if there is one, and only with the invented clients, enquirers and staff accounts the canvas sets up: no real client, staff member or member of the public. Make the requests this row names and no others, with no load. Point every email or text it sets off at an inbox or number the business holds. Inside a vendor's service, stay within what its terms permit (row P0). Passwords, keys, session tokens and any link that works for whoever holds it stay out of the record, and no screen or file that shows a secret is opened. If a response carries a real person's information, halt the row, note the request and the response, look at nothing more, and hand the matter to the person who looks after privacy in the business.
- What the result can inform, and what it cannot decide
- The outside hosts each page contacted and any request that carried the marker, on the build and date tested. It cannot decide whether an identifier in such a request is personal information, or whether sending it is permitted.
E5. Enquiry: the AI chat assistant
- In the invented example
- The assistant answers questions about services, asks for a name and phone number to book a measure-up, and sends the conversation to the CRM. In this invented example, the chat vendor's terms say it passes messages to an AI model provider.
- PIA step it feeds
- Step 5, collection, use and disclosure; step 6, privacy impact analysis.
- Privacy risk questions to ask
- From step 5, Collection: whether unsolicited personal information may be used in the project. From step 6, among its key questions: whether the project recognises the risk of function creep, such as an interest in using the information collected for other purposes in future. This page adds: what does each vendor's documentation say about keeping conversations, and about the businesses it passes them to?
- Targeted test, where technical evidence helps
- Send one chat message containing the E1 marker and no real person's details, with network recording on, and list the hosts it reaches. Then, read-only, find the transcript in the CRM and read the chat vendor's terms on retention and onward providers. Trying to talk the assistant into revealing data or taking actions is a different exercise, which the AI chatbot security guide covers.
- Limits on this row
- One chat message, only within the chat vendor's terms (row P0), and the test conversation deleted afterwards where the tool allows. Request row. Run it on a system your business owns or holds written permission to test, on the staging copy if there is one, and only with the invented clients, enquirers and staff accounts the canvas sets up: no real client, staff member or member of the public. Make the requests this row names and no others, with no load. Point every email or text it sets off at an inbox or number the business holds. Inside a vendor's service, stay within what its terms permit (row P0). Passwords, keys, session tokens and any link that works for whoever holds it stay out of the record, and no screen or file that shows a secret is opened. If a response carries a real person's information, halt the row, note the request and the response, look at nothing more, and hand the matter to the person who looks after privacy in the business.
- What the result can inform, and what it cannot decide
- The hosts one conversation reached and where its transcript landed. It cannot show what the vendor, or a provider it names, does with a conversation after it arrives.
A1. Account record: setting up a client's portal account
- In the invented example
- When a quote is signed, staff create a portal account from the CRM record and the portal emails the client a link to set a password.
- PIA step it feeds
- Step 5, necessity of identity verification.
- Privacy risk questions to ask
- From step 5, Necessity of identity verification: whether it is necessary to verify identity, and the degree of confidence needed; how identity will be verified. This page adds: could the set-up link be used by someone other than the client, or used again?
- Targeted test, where technical evidence helps
- Through the business's normal process, create two invented clients, A and B, each with an email address the business controls. Record which details the set-up email repeats, leaving the link itself out of the evidence. Use A's link once to set a test password, then open it once more and record whether it was refused. If the developer says links expire, try B's once after that time.
- Limits on this row
- Two test accounts created through the normal process, then three uses of set-up links at most: two of A's, and one of B's only if the developer says links expire, with each link kept out of the evidence. Request row. Run it on a system your business owns or holds written permission to test, on the staging copy if there is one, and only with the invented clients, enquirers and staff accounts the canvas sets up: no real client, staff member or member of the public. Make the requests this row names and no others, with no load. Point every email or text it sets off at an inbox or number the business holds. Inside a vendor's service, stay within what its terms permit (row P0). Passwords, keys, session tokens and any link that works for whoever holds it stay out of the record, and no screen or file that shows a secret is opened. If a response carries a real person's information, halt the row, note the request and the response, look at nothing more, and hand the matter to the person who looks after privacy in the business.
- What the result can inform, and what it cannot decide
- How account set-up behaved for two invented clients on the build tested. It cannot decide what degree of confidence in a client's identity the business needs.
A2. Account record: what one client can reach
- In the invented example
- Each client sees their contract, plans, progress photos of their home, invoices and messages. Files sit with the cloud provider and open from a link in the portal.
- PIA step it feeds
- Step 5, security; step 6, the APP 11 questions; step 7, technical controls.
- Privacy risk questions to ask
- From step 5, Security: who will have access, and who will authorise access. From step 6, APP 11: whether control procedures will be in place requiring authorisation before personal information is added, changed or deleted. This page adds: could client B, or a browser that is not signed in, open client A's photos, plans or invoices?
- Targeted test, where technical evidence helps
- Sign in once as B. From B's session, and again from a browser that is not signed in, request A's project page, one of A's files by its link and one of A's invoices, and record each response with the file link itself left out. Then, from B's session, send A's update request for one detail with A's record identifier and record whether the server refused it.
- Limits on this row
- Invented clients A and B and their test records only: one sign-in as B, then seven requests (three from B's session, three from the signed-out browser and one update attempt from B's session). Request row. Run it on a system your business owns or holds written permission to test, on the staging copy if there is one, and only with the invented clients, enquirers and staff accounts the canvas sets up: no real client, staff member or member of the public. Make the requests this row names and no others, with no load. Point every email or text it sets off at an inbox or number the business holds. Inside a vendor's service, stay within what its terms permit (row P0). Passwords, keys, session tokens and any link that works for whoever holds it stay out of the record, and no screen or file that shows a secret is opened. If a response carries a real person's information, halt the row, note the request and the response, look at nothing more, and hand the matter to the person who looks after privacy in the business.
- What the result can inform, and what it cannot decide
- Request and response pairs showing whether the portal refused each attempt, for the accounts and records tried on the build tested; a refusal is a result for those attempts only. It cannot decide whether the steps taken are reasonable under APP 11, which the APP 11 guide covers.
A3. Account record: staff access and the audit trail
- In the invented example
- Office staff and the site supervisor have staff accounts. An administrator can export every client record.
- PIA step it feeds
- Step 5, security; step 6, the APP 11 questions.
- Privacy risk questions to ask
- From step 5, Security: the systems that will prevent and detect misuse or inappropriate access; what action will be taken if there is a data breach. From step 6, APP 11: whether audit mechanisms will identify inappropriate system access. This page adds: does each staff role need everything it can reach?
- Targeted test, where technical evidence helps
- Read the role list and record which roles can open and export client records. Then, with a test staff account, open invented client A's record once and export it once, and ask the developer to show what the application's logs recorded for those two actions, with passwords and session tokens removed, and who can read those logs.
- Limits on this row
- One sign-in with a test staff account, then one view and one export of a test record only, with the export saved where the business controls it and deleted afterwards. Request row. Run it on a system your business owns or holds written permission to test, on the staging copy if there is one, and only with the invented clients, enquirers and staff accounts the canvas sets up: no real client, staff member or member of the public. Make the requests this row names and no others, with no load. Point every email or text it sets off at an inbox or number the business holds. Inside a vendor's service, stay within what its terms permit (row P0). Passwords, keys, session tokens and any link that works for whoever holds it stay out of the record, and no screen or file that shows a secret is opened. If a response carries a real person's information, halt the row, note the request and the response, look at nothing more, and hand the matter to the person who looks after privacy in the business.
- What the result can inform, and what it cannot decide
- Which roles can export, and whether two staff actions left log entries, on the build tested. It cannot decide whether monitoring is adequate or what the business would do after a breach; the notifiable data breaches guide covers the scheme.
A4. Account record: access and correction by the client
- In the invented example
- A client can update their phone number and address in the portal. Other corrections go by message to the office.
- PIA step it feeds
- Step 5, access and correction; step 6, the APP 12 and APP 13 questions.
- Privacy risk questions to ask
- From step 5, Access and correction: how individuals can access their personal information, including any costs; how they can have it corrected, or annotations made, if necessary; how decisions will be made about requests for access or correction. This page adds: does a change made in the portal reach every other copy?
- Targeted test, where technical evidence helps
- Sign in as invented client A and change the phone number once in the portal. Then have the developer check, without changing anything, whether the new number reached the CRM and the email delivery service's contact list, and when.
- Limits on this row
- One sign-in as invented client A and one change on that test account. Request row. Run it on a system your business owns or holds written permission to test, on the staging copy if there is one, and only with the invented clients, enquirers and staff accounts the canvas sets up: no real client, staff member or member of the public. Make the requests this row names and no others, with no load. Point every email or text it sets off at an inbox or number the business holds. Inside a vendor's service, stay within what its terms permit (row P0). Passwords, keys, session tokens and any link that works for whoever holds it stay out of the record, and no screen or file that shows a secret is opened. If a response carries a real person's information, halt the row, note the request and the response, look at nothing more, and hand the matter to the person who looks after privacy in the business.
- What the result can inform, and what it cannot decide
- Whether one correction reached each connected copy on the build tested, and when. It cannot decide whether a request was handled within a reasonable period, or how a refusal should be handled.
A5. Account record: hosting in a region outside Australia
- In the invented example
- The portal, its database, uploaded files and backups run in a cloud region outside Australia, as the hosting contract names it.
- PIA step it feeds
- Step 5, disclosure and security.
- Privacy risk questions to ask
- From step 5, Disclosure: whether the personal information will be disclosed to overseas recipients. From step 5, Security: how personal information will be protected if it will be managed by someone else. This page adds: where does each part sit, according to the documents, and who at the provider can reach it?
- Targeted test, where technical evidence helps
- No test. Read the hosting account's region settings and the hosting contract, and record where the application, database, files and each backup copy sit as those documents state, with the date read. Ask the provider in writing about anything they leave open.
- Limits on this row
- Read-only: reads settings, lists, documents or logs and sends no test requests. Any sign-in is a staff account the business already uses, used as it normally is. Open no screen or file that displays a password, key or other secret, and copy none.
- What the result can inform, and what it cannot decide
- The locations the documents give, dated. It cannot settle whether the arrangement is a use or a disclosure, or what steps are reasonable; the APP 8 cross-border disclosure guide and its register cover those questions.
D1. Enquiry: deletion when it does not become a job
- In the invented example
- The invented business has decided to delete enquiries that do not become jobs after a set period, and staff do it in the CRM.
- PIA step it feeds
- Step 5, retention and destruction.
- Privacy risk questions to ask
- From step 5, Retention and destruction: when personal information will be de-identified or destroyed; how this will be done securely; whether an information retention policy and destruction schedule is in place. This page adds: when the CRM record goes, which copies stay?
- Targeted test, where technical evidence helps
- Delete the E1 test enquiry through the business's normal process. Then, read-only, search the form service, the CRM, the automation service's run history, the shared inbox, the chat transcript store and the analytics reports for the marker, and record which still hold it.
- Limits on this row
- One deletion of the test enquiry only, where the rules for the test allow it, and nothing that existed before the test, under the request-row limits below. The search for the marker that follows is read-only: one search for the marker in each system named, and for nothing else, signed in with a staff account the business already uses, used as it normally is. Request row. Run it on a system your business owns or holds written permission to test, on the staging copy if there is one, and only with the invented clients, enquirers and staff accounts the canvas sets up: no real client, staff member or member of the public. Make the requests this row names and no others, with no load. Point every email or text it sets off at an inbox or number the business holds. Inside a vendor's service, stay within what its terms permit (row P0). Passwords, keys, session tokens and any link that works for whoever holds it stay out of the record, and no screen or file that shows a secret is opened. If a response carries a real person's information, halt the row, note the request and the response, look at nothing more, and hand the matter to the person who looks after privacy in the business.
- What the result can inform, and what it cannot decide
- Which copies outlived the deletion of the original, on the systems and the date searched. It cannot decide how long any copy should be kept, or whether a law requires one to be kept: that is for the business and its adviser.
D2. Account record: closing the account when the project ends
- In the invented example
- At handover, staff close the client's portal account. The invented business keeps the signed contract and invoices and has not yet decided about the progress photos.
- PIA step it feeds
- Step 5, retention and destruction; step 6, the APP 11 questions.
- Privacy risk questions to ask
- From step 5, Retention and destruction: how compliance with the retention policy and any relevant legislation about record destruction will be assessed. From step 6, APP 11: whether reasonable steps will be taken to destroy or de-identify personal information which is no longer needed for any authorised purpose. This page adds: does closing the account stop every route to the files?
- Targeted test, where technical evidence helps
- Close invented client B's account through the normal process. Then try once to sign in as B, and once to open one of B's file links from a browser that is not signed in, and record both responses, leaving the file link itself out of the record. Read, without changing, the backup retention setting.
- Limits on this row
- One closure of test account B through the normal process, then two requests: one sign-in attempt and one file link. Request row. Run it on a system your business owns or holds written permission to test, on the staging copy if there is one, and only with the invented clients, enquirers and staff accounts the canvas sets up: no real client, staff member or member of the public. Make the requests this row names and no others, with no load. Point every email or text it sets off at an inbox or number the business holds. Inside a vendor's service, stay within what its terms permit (row P0). Passwords, keys, session tokens and any link that works for whoever holds it stay out of the record, and no screen or file that shows a secret is opened. If a response carries a real person's information, halt the row, note the request and the response, look at nothing more, and hand the matter to the person who looks after privacy in the business.
- What the result can inform, and what it cannot decide
- Whether sign-in and a file link stopped working after closure on the build tested, and how long backups are kept as the setting reads. It cannot decide what the business should keep, or for how long.
R1. From findings to the PIA report and its review
- In the invented example
- Invented finding: row A2 showed that one of A's progress photos opened from its link in a browser that was not signed in. It is made up to show the paperwork.
- PIA step it feeds
- Step 7, privacy management; step 8, recommendations; step 9, report; step 10, respond and review.
- Privacy risk questions to ask
- From step 8: who each recommendation is addressed to, and the timeframe for implementation. From step 10: which recommendations the entity intends to implement or has implemented, which it does not, and the rationale. This page adds: which build was retested, and on what date?
- Targeted test, where technical evidence helps
- No new test of its own. A retest is the row that found the issue run again on an identified build after the fix.
- Limits on this row
- Read-only: reads settings, lists, documents or logs and sends no test requests. Any sign-in is a staff account the business already uses, used as it normally is. Open no screen or file that displays a password, key or other secret, and copy none. Any retest is the original row run again, with that row's limits.
- What the result can inform, and what it cannot decide
- For each finding: the risk it raises for step 6, the recommendation with its owner and date for step 8, and its retest status for step 10. It cannot decide whether the project's privacy outcomes are acceptable: the guide gives that to the privacy impact analysis, which belongs to the entity.
What a privacy impact assessment is, and where the method comes from
The Office of the Australian Information Commissioner (OAIC) describes a privacy impact assessment (PIA) as a systematic assessment of a project that identifies potential privacy impacts and recommendations to manage, minimise or eliminate them.
Its method is set out in the Guide to undertaking privacy impact assessments, which has a publication date of 2 September 2021 and which the OAIC says is intended to provide guidance to all Australian Privacy Principle (APP) entities. A shorter page, 10 steps to undertaking a privacy impact assessment, has a publication date of 29 April 2021 and comes with a downloadable poster, and the OAIC also publishes a PIA tool and a free eLearning course on conducting a PIA.
The Privacy Act 1988 carries its own description. Section 33D(3), in Compilation No. 104 of the Act (compilation date 4 June 2026), says a privacy impact assessment is a written assessment of an activity or function that identifies the impact that the activity or function might have on the privacy of individuals, and sets out recommendations for managing, minimising or eliminating that impact.
Section 33D(4) says subsection (3) does not limit the matters the assessment may deal with. Both sit in the section that lets the Commissioner direct an agency to give one, covered below.
The guide is plain that a PIA reaches beyond compliance. It calls a PIA much more than a simple compliance check, one that should tell the full story of a project from a privacy perspective, going beyond compliance to also consider the broader privacy implications and risks, including whether the planned uses of personal information in the project will be acceptable to the community, and a footnote adds that a PIA may also consider other types of personal privacy, such as bodily, behavioural and communications privacy.
Information security is one strand of that story.
This page therefore treats a security test as one input to a PIA, and never as the PIA or as approval of one.
This guide is general information about the OAIC's published method and the text of the Act and the agencies' code, applied to websites and web applications. It is not legal advice and not an audit, and nothing on it says that an assessment is complete, adequate or approved.
Who has to carry out a PIA, and who the OAIC encourages to
For Australian Government agencies, a PIA can be a requirement.
The Australian Government Agencies Privacy Code, the name by which section 1(2) of the 2017 governance APP code says it may also be cited, was developed under section 26G of the Privacy Act 1988, as its section 3 says, and section 2 says it comes into force on 1 July 2018.
Section 7 says it is binding on all agencies as section 5 defines them, which takes the meaning of agency in section 6(1) of the Act but does not include a Minister. Section 12(1) says an agency must conduct a PIA for all high privacy risk projects, and section 12(2) says a project may be a high privacy risk project if the agency reasonably considers that the project involves any new or changed ways of handling personal information that are likely to have a significant impact on the privacy of individuals.
The code goes on to the paperwork. Section 15(1) says an agency must maintain a register of the PIAs it conducts and must publish the register, or a version of the register, on its website, and section 10(5)(c) lists assisting with the preparation of PIAs conducted under section 12 among the Privacy Officer functions an agency must ensure are carried out.
The Federal Register of Legislation's details page, read on 15 November 2026, lists the code as in force, as a principal instrument as made, registered on 27 October 2017, and says it is due to sunset on 1 April 2028.
The Act adds a direction power in section 33D(1). If an agency proposes to engage in an activity or function involving the handling of personal information about individuals, and the Commissioner considers that the activity or function might have a significant impact on the privacy of individuals, the Commissioner may, in writing, direct the agency to give the Commissioner, within a specified period, a privacy impact assessment about the activity or function.
Under section 33D(6), if an agency does not comply with a direction, the Commissioner must advise the Minister of the failure and, if another Minister is responsible for the agency, that other Minister.
For a business that is an APP entity but not an agency, the OAIC presents a PIA as something it recommends and encourages. Its guide says the OAIC's power to direct agencies to undertake a PIA does not apply to private sector organisations, and that the OAIC encourages organisations to undertake PIAs for projects that involve handling of personal information, and to share their findings publicly.
The ten-step page says the OAIC recommends that organisations conduct PIAs as part of their risk management and planning processes. Paragraph 1.7 of chapter 1 of the OAIC's APP guidelines (version 2.0, updated 30 September 2026) lists examples of practices, procedures and systems that an APP entity should consider implementing.
One of them is a commitment to conducting a PIA for new projects that will handle personal information, or when someone proposes a change to information handling practices, and the paragraph adds that whether a PIA is appropriate will depend on a project's size, complexity and scope, and the extent to which personal information will be collected, used or disclosed.
Whether your business is an APP entity at all, whether it is an agency or works for one, and what any contract, funding agreement or other law asks of it are for your business to settle with its own adviser; the Australian Privacy Principles guide explains who the principles apply to.
Where a request for a PIA comes from a contract or a law other than the Privacy Act, its own wording says what is asked, and this page does not summarise it.
When a website or web application project raises the question
The guide uses the word project loosely, for the full range of activities and initiatives that may have privacy implications, and its list includes new or amended programs, activities, systems or databases, new methods or procedures for service delivery or information handling, and changes to how information is stored.
Reading that list against a website, this page would put a new enquiry or booking form, a customer account area or client portal, a chatbot or other AI feature, a new customer relationship tool or automation fed by the site, and a move of hosting, files or backups to another provider or country through the threshold step. That reading is this page's, not the OAIC's.
Step 1 is a threshold assessment, and the guide says not every project will need a PIA. Its first question is whether any personal information will be collected, stored, used or disclosed in the project. The guide says that, generally, if personal information is involved, some form of PIA may be necessary, and the ten-step page says that if the answer is yes, a PIA is usually necessary.
The guide also says a PIA may not be necessary if the project does not propose any changes to existing information handling practices, if the privacy implications of these practices have been assessed previously and controls are current and working well. Whichever way the answer falls, it says you should keep a record of the threshold assessment.
Some passages in the guide bear directly on websites. One says that under the Privacy Act, information does not always have to include details such as an individual's name to qualify as personal information, and may include other information that can identify an individual or allow their identity to be determined.
Another, written for an entity bound by the Privacy Act that is developing a project involving personal information, says the entity is responsible and accountable for the personal information it collects, even when the information is held by external service providers or contractors operating in Australia or overseas.
On this page's reading, a site assembled from a form service, a mailing platform, an analytics tag and a hosted portal needs each of those services in its map.
The guide says the size or budget for a project is not a useful indicator of its likely privacy impact, and that even a small-scale project may have significant privacy implications. Its examples of circumstances in which a project's privacy scope can increase include personal information handling that will be or has been outsourced, new legislation or new technology needed for handling or storing the information, aggregation in databases, entirely new collections, and a new method of using or disclosing personal information.
On timing, it says a PIA should be undertaken early enough in the development of a project that it is still possible to influence the project design or, if there are significant negative privacy impacts, reconsider proceeding with the project.
The ten steps, read for a website or application
The guide sets out a suggested ten step process, which it says can be used alongside existing project management and risk management methodologies or as a process in its own right.
It allows that different entities might use different processes, and says that ideally these processes will address each of the steps in some way. The subsections keep the guide's step numbers and names; what each asks of a website project, and where test evidence can inform it, is this page's reading.
Step 1 is the threshold record described above. For step 2, the guide's planning elements include how detailed the PIA needs to be, who will conduct it, the timeframe, the budget, the extent and timing of stakeholder and public consultation, and the steps that will follow it.
It says that generally whoever is managing the project would be responsible for ensuring the PIA is carried out, that a PIA is unlikely to be effective if done by a staff member working in isolation, and that a range of expertise may be required, including information security, technology, risk management, law, ethics, operational procedures and industry-specific knowledge.
Where a project will have substantially more privacy impact than others, it says a robust and independent PIA conducted by external assessors may be preferable.
On a website project, this page would plan for the developer, whoever runs the hosting and each vendor's documentation as inputs to steps 3 and 5, and for a tester only once a built system exists to test. Step 3 asks for a brief, big-picture description of the project, which the guide says should be detailed enough for external stakeholders to understand the project and written in plain English.
For step 4, the guide says an entity will have internal and external stakeholders, including regulatory authorities, clients, advocacy organisations, service providers, industry experts, academics and others, and that while it may not be necessary to consult all of those identified, depending on the scale and likely privacy impacts of the project, some form of consultation should occur as part of the PIA.
Step 5 is where a website project gets concrete. The guide says detailed mapping should include whether identity verification will be necessary, what personal information will be collected and how, its use and disclosure, the processes for ensuring information quality, the security safeguards that are or will be in place, and the ability individuals have to access and correct their personal information, and that mapping should also describe the current personal information environment and how the project will affect it.
It suggests, if appropriate, diagrams of the flow of information or tables setting out the key information for different types of personal information. The worked canvas on this page is one such table.
Its prompts under Security include the security measures that will protect the information, including for contracted service providers; how information will be transferred between sites; who will have access, and who will authorise access; the systems that will prevent and detect misuse or inappropriate access; and what action will be taken if there is a data breach.
On this page's reading, a test of a built system can speak to some of those, such as who can reach a record and whether access leaves a trace, and to none of the breach response.
The prompts under Collection, Use, Disclosure, Retention and destruction, and Access and correction are answered in the canvas from documents, settings and decisions, with a targeted observation added where one helps.
Step 6 identifies and critically analyses how the project affects privacy, both positively and negatively. The guide says the analysis should ultimately attempt to determine whether the project has acceptable privacy outcomes, or unacceptable privacy impacts, and its key questions include whether individuals have to give up control of their personal information, whether the project recognises the risk of function creep, and how valuable the information would be to unauthorised users.
For the compliance check, the guide says you will need to consider whether your project complies with each of the APPs. Its questions for APP 11 include whether control procedures will be in place requiring authorisation before personal information is added, changed or deleted, and whether audit mechanisms will identify inappropriate system access. Those two can be informed by testing the built system; whether an outcome is acceptable cannot.
Step 7 considers what options may remove, minimise or mitigate the negative privacy impacts the analysis found; the guide heads it addressing risks, and the ten-step page considering risks. The guide says strategies may include technical controls, such as access control mechanisms, encryption and design changes; operational controls, such as policies or procedures, staff training, and oversight and accountability measures; and communication strategies, such as privacy notices.
Step 8 turns those into recommendations, and the guide says it should be clear who they are addressed to and that they should include a timeframe for implementation.
Step 9 is the report. The guide's key elements for it include a project description, the PIA methodology, a description of information flows, the outcome of the privacy impact analysis and compliance checks, recommendations to avoid or mitigate privacy risks, and a description of any privacy risks that cannot be mitigated, the likely community response to them and whether they are outweighed by the public benefit of the project, with appendices if necessary.
Step 10 keeps the PIA open. The guide says the project manager and the entity may decide not to implement all the recommendations but should document which they intend to implement, or have already implemented, which they do not, and the rationale, and that the PIA should be revisited and updated or revised if developments in the design or implementation create new privacy impacts that were not previously considered. A retest record belongs here.
What a security test can supply to a PIA, and what it cannot
A PIA is the entity's own assessment, and a security test is one input to it.
In the canvas that input sits at the step 5 security and access prompts, the APP 11 questions in step 6, the technical controls weighed in step 7, and the review in step 10.
Plan for a timing gap. The guide wants a PIA early enough to shape the design, and a test needs something built to run against. One way to bridge it is to complete the PIA on the design, carry each security question forward as an open item with an owner, test a staging build before launch, and fold the results into the update the guide describes.
It says a PIA may be useful more than once during a project's development and implementation, and should be revisited and updated when changes to the project are considered.
What a targeted test can show is narrow and dated: the way one control responded, on one build, to the accounts and records used. For the canvas, that covers whether one invented client could open another client's record or file, whether a browser that was not signed in could, whether a value typed into a form reached an analytics or advertising host, how many copies one enquiry left and whether deleting the original removed them, and whether a staff action left an entry in the application's logs.
A refusal or a clean search is a result for the requests made, and says nothing about requests nobody tried.
A test cannot answer the questions that turn on judgement. It cannot say whether a field is necessary, whether a use is one people would expect, whether consultation was enough, whether the project fits community expectations, or whether an exception to a principle applies. Staff training and written procedures are invisible to it, as is anything outside the agreed scope, and it reaches into a vendor's platform no further than that vendor's terms permit.
Nor can it decide whether a privacy outcome is acceptable: the guide gives that question to the privacy impact analysis, which belongs to the entity.
Security, overseas recipients and breaches: where the detail lives
A website PIA runs into three subjects that have guides of their own here, so each gets a sentence and a link.
- Security of personal information: APP 11.1 in Schedule 1 to the Privacy Act 1988 says that if an APP entity holds personal information, it must take such steps as are reasonable in the circumstances to protect the information from misuse, interference and loss, and from unauthorised access, modification or disclosure, and the APP 11 guide covers what test evidence can support towards it.
- Overseas recipients: APP 8.1 says that before an APP entity discloses personal information about an individual to an overseas recipient, it must take such steps as are reasonable in the circumstances to ensure that the recipient does not breach the principles, other than APP 1, in relation to the information, and APP 8.2 sets out cases where APP 8.1 does not apply; the APP 8 cross-border disclosure guide has a register for the overseas flows that row A5 raises.
- Data breaches: the OAIC explains that under the Notifiable Data Breaches scheme, an organisation or agency covered by the Privacy Act 1988 must notify affected individuals and the OAIC when a data breach is likely to result in serious harm to an individual whose personal information is involved, and the notifiable data breaches guide covers it, while step 5 of the PIA guide asks what action will be taken if there is a data breach.
How to work through the canvas on your own project
The canvas follows two records through an invented home renovation business: an enquiry from a person who has not hired it, and the portal account record of a client who has.
They raise different questions. An enquiry comes from someone who may never become a client, which puts the weight on how many copies it leaves and what happens to them. An account record grows over a project, with a contract, plans, photos of a home and invoices, which puts the weight on who can reach it.
The business, its systems, its settings and the one finding in row R1 are invented to show the method. This is a worked example, not a template, and it describes no real business.
Each row reads left to right: the stage, what happens there in the example, the PIA step it feeds, the privacy risk questions, a targeted test where technical evidence helps, the limits on the row, and what a result can inform and cannot decide. Questions credited to the guide give its words or a short form of them under the guide's own heading, and each row adds one question of this page's own.
To use it, replace the example column with what your project does, keep the questions, and drop any test your project has no use for.
- Begin with row P0, whose system list the later rows draw on. Where a vendor's terms say nothing clear about security testing or data handling, put the question to the vendor in writing, and leave that vendor's service untouched until its reply is filed.
- The read-only rows are P0, E2, E3, A5 and R1: they read settings, lists, documents and logs, and send no test requests. E2 and E3 read what row E1 produced, so they follow it, and R1 comes last.
- Rows E1, E4, E5, A1, A2, A3, A4, D1 and D2 send requests, sign in or change test records. They belong on systems your business owns or holds written permission to test, on the staging copy if there is one, and every test sign-in, request and change uses invented enquirers, clients and test staff accounts, never a real client, staff member or member of the public. The one exception is row D1's read-only search, which, like row E2, uses staff accounts the business already has.
- Carry one unique marker, such as PIA-CANVAS-7Q, in the invented details so every copy can be traced and removed, and give the test an email address and phone number that the business holds.
- Make only the requests a row lists, sign-ins included. No row calls for load.
- Passwords, keys, session tokens and any link that works for whoever holds it, such as a set-up or file link, stay out of the evidence, and no screen or file that shows a secret gets opened.
- If a real person's information appears, halt there. Note the request and the response, look at nothing more, and pass the matter to the person who looks after privacy in the business.
- Run the deletion and closure checks in rows D1 and D2 only on the test enquiry and test account the canvas created, then delete or list everything else the test created.
- File each result in the PIA under the step its row names, with the build and the date, and hand it to whoever is responsible for the PIA.
Who decides what the PIA concludes
On this page's reading of the guide, the entity running the project does.
The guide says that, generally, whoever is managing the project would be responsible for ensuring the PIA is carried out, and in step 10 it is the project manager and the entity that may decide not to implement all the recommendations, and that should document which they intend to implement and which they do not, with the rationale.
A privacy consultancy or law firm, or the independent review the guide says has benefits, adds expertise and a second view without changing whose decisions those are.
For an agency, the Privacy Officer functions under section 10(5)(c) of the code include assisting with the preparation of PIAs conducted under section 12.
The OAIC does not sign off an organisation's PIA. Its guide says the OAIC does not have a role in the development, review, endorsement or approval of the PIA process for private sector organisations. For agencies, it says the OAIC has no formal role in the development, endorsement or approval of PIAs it has not directed, and may, subject to available resources, be able to assist with advice during the PIA process; where it directs an agency to give it a PIA under section 33D, it says it will generally review the PIA report to ensure that it sufficiently identifies privacy impacts and sets out appropriate recommendations.
Publication is the entity's call too. The guide says the OAIC strongly encourages the publication of PIA reports, acknowledges there may be circumstances when full release is not appropriate, for example if the project is still in its very early stages, or if there are security or commercial reasons for not releasing the full report, and encourages a summary or edited version where there are difficulties making the full PIA available. A PIA that quotes test findings in detail is one place that choice arises.
A customer, agency buyer or insurer that asks to see your PIA, or asks questions one would answer, decides what it will accept for its own purposes, and the security questionnaires guide covers answering from evidence without overstating it. The canvas, a test report and a retest record feed those decisions without making any of them, and none of the three is a certificate.
Where a Clever Ops test fits in a PIA, and where it stops
Clever Ops tests what customers reach through a browser or an API, and several of its services meet the canvas at particular rows.
Behind the login, the web application and API penetration test works through a custom application's authentication and sessions, what each role can see and change, whether one customer can reach another customer's records, the API called directly as well as through the screens, and file uploads, among other areas; that overlaps rows A1 and A2 and the role questions in row A3.
From the outside, the website security check reviews what anyone on the internet can see without signing in, such as certificates, security headers, cookie settings, and exposed files or admin paths, and it stops at the login.
The AI chatbot and agent security test is aimed at the chatbot or agent, not the AI model behind it, and takes in data the bot can be made to reveal and every tool or integration it can call, which reaches well past row E5.
For a WordPress site, including a WooCommerce store or member area, the WordPress security audit follows each role that can log in.
None of those services is a privacy impact assessment, and a report neither completes nor approves one. A Clever Ops test report sets out the scope as tested and each finding with redacted evidence, and lists checks that came back clean as negative results, which speak only for the build and accounts tested.
After the one consolidated retest, each finding is marked fixed, partially fixed, still reproducible or not retested, and row R1 carries that status into step 10.
The record describes what was checked. It is not a certificate or a compliance opinion, and Clever Ops is not CREST accredited and is neither a PCI Approved Scanning Vendor nor a Qualified Security Assessor.
Destructive actions are off by default for real people's accounts and records that existed before the test, and reset, lockout and deletion checks run only on test accounts and test records where the rules of engagement allow them.
Evidence is redacted of session tokens, passwords, keys and personal details, stored encrypted in Australia, processed outside Australia only where the statement of work approves that, and deleted on the schedule in the data handling terms.
Clever Ops does not test a vendor's own product or platform, such as the hosted form, chat, booking or payment service a site relies on. A tender or contract that requires a CREST-accredited provider is referred on: Clever Ops is not CREST accredited.
On a site Clever Ops built or hosts, the same work is called a security review, not an independent test, and the report names who built the site and who hosts it.
The PIA itself stays with your business and the adviser it chooses; the OAIC's guide notes that a number of privacy consultancies and law firms offer PIAs as a service.
Frequently asked questions
For agencies bound by the Australian Government Agencies Privacy Code, section 12(1) says an agency must conduct a PIA for all high privacy risk projects, and section 12(2) says a project may be one if the agency reasonably considers that it involves any new or changed ways of handling personal information that are likely to have a significant impact on the privacy of individuals. Under section 33D(1) of the Privacy Act 1988, if an agency proposes to engage in an activity or function involving the handling of personal information about individuals and the Commissioner considers it might have a significant impact on the privacy of individuals, the Commissioner may, in writing, direct the agency to give the Commissioner a PIA about it within a specified period. For private sector organisations, the OAIC's guide says that power to direct does not apply, and that the OAIC encourages organisations to undertake PIAs for projects that involve handling of personal information. Whether your business is an agency, an organisation bound by the Act or neither, and what a contract, funding agreement or other law that asked for a PIA requires of it, are questions for your own adviser.
The OAIC's guide lists key elements: a project description; the PIA methodology; a description of information flows; the outcome of the privacy impact analysis and compliance checks, including positive privacy impacts, the privacy risks identified and strategies already in place; recommendations to avoid or mitigate privacy risks; and a description of any privacy risks that cannot be mitigated, the likely community response to them and whether they are outweighed by the public benefit of the project, with appendices if necessary. It also says there is no single way of doing a PIA or setting out a PIA report. On this page's approach, test evidence goes into the information flows and the analysis in summary, with the full test report kept where the business controls who reads it.
The OAIC publishes a PIA tool, which its guide says helps you conduct a PIA, report its findings and respond to recommendations, and which it encourages entities to adapt to the size, complexity and risk level of their project. Appendix A of the guide lists a small sample of PIA reports or summaries published by some government agencies and organisations, and says the OAIC does not endorse any of these reports or encourage a particular format. The canvas on this page is an invented worked example for a website project, built around steps 5 and 6: it shows a method, it is not a template, and it is not a complete PIA.
While the design can still change. The OAIC's guide says a PIA should be undertaken early enough in the development of a project that it is still possible to influence the project design or, if there are significant negative privacy impacts, reconsider proceeding with the project, and that it should be revisited and updated when changes to the project are considered. A test needs something built, so on this page's approach the PIA starts on the design and the results of a test on a staging build go into its update.
No. In the OAIC's words a PIA is much more than a simple compliance check, and questions such as whether a collection is necessary, how the use of personal information aligns with community expectations and whether the privacy outcomes are acceptable are judgements a test cannot make. A test can inform the security and access parts of steps 5 and 6, and a retest can feed the review in step 10. It records how controls behaved on the build and accounts tested, and it does not complete or approve the PIA.
Not for a private sector organisation: the OAIC's guide says it does not have a role in the development, review, endorsement or approval of the PIA process for private sector organisations. For an agency, it says it has no formal role in the development, endorsement or approval of PIAs it has not directed, though it may, subject to available resources, be able to assist with advice during the PIA process. Where it directs an agency to give it a PIA under section 33D of the Privacy Act, it says it will generally review the report to ensure that it sufficiently identifies privacy impacts and sets out appropriate recommendations.
Scope the right test for your application
Tell us what you have built, who uses it and when it needs to launch. You will get a plain-English recommendation: a website audit, a web application test or a referral to a specialist.