AI Risk Assessment for Accountants: Vetting a Tool
Before you approve an AI tool, check the provider. Run an AI risk assessment covering training, retention, storage, access and exit, and keep the record.
- Published
- Updated
An AI risk assessment checks whether a tool suits a specific practice task. Record its data handling, the risks that remain and who approves the use.
When to Run an AI Risk Assessment
Assess a new AI tool before staff use it for client work. Repeat the assessment when the intended use changes or the provider changes terms that affect your information.
Approval belongs to a particular use and account arrangement. A tool approved for public research needs another assessment before staff connect client files or use it to prepare advice.
Reassess when an existing product adds an AI feature, changes its model provider or requests wider file access. A new subscription name alone tells you little about what happens to the information.
The Tax Practitioners Board’s AI guidance asks practitioners to consider the task, storage and intended reliance on the output. Practitioners remain responsible for the tax agent services they provide.
Before the vendor review, write one sentence describing the proposed task. Add the information it needs and the person who checks its output. This defines what you are deciding to approve.
For example, assessing a tool to summarise public tax guidance is different from assessing it to summarise a client meeting. The second use adds client information and recording arrangements to the assessment.
Vendor Security Assessment Questions
Ask the provider about the exact product and account type you intend to use. Keep written answers and the documents supporting them, including the applicable contract and privacy terms.
The Office of the Australian Information Commissioner’s guidance on commercially available AI products covers intended use and who receives information. It also addresses security, training uses and continued monitoring.
Use these questions as a vendor security assessment. Record an unanswered question as unknown, with an owner to resolve it before the proposed use starts.
| Area | Questions to put to the provider | Evidence to retain |
|---|---|---|
| Training | Are prompts, files or outputs used to train or improve models? Does the answer cover every model provider? Can an administrator enforce the setting? | Contract clauses and the account setting |
| Retention | How long are prompts, uploads, outputs and security logs kept? What changes when a chat or account is deleted? | Retention schedule and stated exceptions |
| Storage location | Where does processing occur? Where are primary copies, backups and support access located? | Written location details for the selected service |
| Subprocessors | Which other businesses process the information? What does each receive? How are changes to that list notified? | Subprocessor list and change terms |
| Certifications | What independent security assurance covers this service? What are its scope, period and exclusions? | Certificate or assurance report and its scope |
| Access | Can staff use practice-managed accounts and two-step sign-in? Who at the provider can view content? What account and access records are available? | Access documentation and configured permissions |
| Deletion and exit | Can the practice export what it needs? What gets deleted on exit, including at subprocessors? What remains in backups and for how long? | Export method, deletion terms and confirmation arrangements |
| Incidents | How does the provider report a breach? Who receives the notice, within what period and with what information? | Contract terms and the practice contact |
A statement that data is excluded from training does not answer the retention question. Ask about files and logs separately, including any exceptions for abuse monitoring or legal retention.
A security certificate also needs context. Match the organisation and service in its scope to the service you plan to buy. A logo on a sales page does not establish what the assessment covered.
Read the data processing agreement alongside the provider’s answers. Record any difference between the contract and the settings available in your account.
The TPB explains that entering client information into an AI tool can disclose it to a third party, depending on configuration and use. Address required client permission before that disclosure occurs.
Permission and security are separate questions. Record both the basis for using client information and the safeguards for handling it. Where the Privacy Act applies, assess the relevant privacy obligations too.
Scoring and Recording the Result
Rate each material risk before and after the proposed safeguards. Use the remaining risk to make the approval decision, and keep the reason beside the rating.
The following is a simple internal model your practice can adopt. It is not a TPB or OAIC scoring system.
Give likelihood a score from 1 to 3. Use 1 for unlikely, 2 for possible and 3 for likely. Explain that choice using the provider evidence and proposed task.
Score impact on the same scale. Use 1 for a limited internal inconvenience, 2 for a material service error or restricted disclosure, and 3 for serious client harm or a confidentiality breach.
Multiply likelihood by impact for each risk. Keep a separate line for disclosure and for incorrect output, so a useful drafting feature cannot offset a client-data risk.
| Remaining score | Rating | Suggested decision |
|---|---|---|
| 1 or 2 | Low | Approve the defined use if its other requirements are met. Record who checks the output. |
| 3 or 4 | Medium | Require written conditions and principal sign-off before use. Confirm the conditions are in place. |
| 6 or 9 | High | Hold the proposed use until safeguards reduce the risk or choose another approach. |
Do not average scores across risks. Use the highest remaining rating, with unresolved requirements recorded separately. A low numerical score cannot replace required permission or resolve unknown data handling.
For a hypothetical meeting-summary tool, the provider has not explained transcript retention. Record that answer as unknown and hold client recordings. A trial using fictional material can test the summary without authorising client uploads.
Keep the evidence and decision in your AI governance register. Record the allowed task, account type and data limits so another staff member can apply the approval without repeating the assessment.
AI Risk Assessment Template
Copy this one-page assessment record for each proposed use. Keep detailed provider documents as attachments and reference them in the record. Complete every field, using unknown where an answer remains unresolved.
| Field | What to record |
|---|---|
| Assessment identity | Record number, assessor, assessment date and practice owner |
| Tool and account | Provider, product, account type, model or service version where available |
| Proposed task | The specific staff task and intended use of the output |
| Information allowed | Data categories, file types and connected systems |
| Information prohibited | Client details or other information excluded from this approval |
| Provider evidence | Links or saved copies of terms, security material and written answers, with their versions |
| Training and retention | Whether training is permitted, the enforced settings and retention periods |
| Locations and other providers | Processing and storage locations, support access and subprocessors |
| Access and incidents | Sign-in requirements, permissions, logs and breach notification arrangements |
| Exit arrangements | Export method, deletion scope, backup exceptions and confirmation |
| Permission and privacy | Required client permission, its record and relevant privacy assessment |
| Human review | Who checks outputs, what they check and who can stop use |
| Risk results | Each risk, likelihood, impact, safeguard and remaining score |
| Decision | Approve, approve with conditions or hold, with the reason and approver |
| Conditions and review | Conditions, action owners, due dates, next review and earlier reassessment triggers |
Before enabling the tool, have a second staff member read the record and describe the permitted task. Confirm the account settings match the approval and that any conditions are complete.
If the record says public information only, client uploads remain prohibited. A later request to add them starts a new assessment of that use.
One client form. One finished return.
Your client completes one guided form. Tax Genie checks it, chases mistakes and sends the details to your engagement letter and tax software.
Made for Australian accounting firms