AI governance sets who approves an AI tool, which client information it can receive and who checks its work. For a small mortgage brokerage, a named owner and a written approval record can establish those controls without an enterprise programme. Approval needs to cover the way staff use the tool, including any access it has to client files.
What AI Governance Means in a Brokerage
AI governance is the set of decisions, responsibilities and records that controls AI use in your brokerage. It connects permission to use a tool with limits on its data access and checks on its output.
An AI use policy tells staff the rules. Governance establishes who makes those rules, approves exceptions and acts when something goes wrong. A tool can pass a technical review yet remain unsuitable for a particular client task.
For a practice of one to ten people, start with one accountable owner and one register of approved uses. A sole broker can hold that role. In a larger team, the owner can delegate checks while keeping responsibility for the decision. Credit representatives also need to follow their licensee’s approval requirements.
Consider a hypothetical assistant approved to draft generic appointment reminders. Giving that assistant access to every client’s financial documents changes its use and risk. The earlier approval does not cover the new access.
Australian Frameworks to Borrow From
Use Australia’s voluntary AI guidance to organise your controls, while treating privacy and credit licence obligations as binding requirements where they apply. Following a voluntary framework does not establish legal compliance.
The Department of Industry, Science and Resources names these eight principles:
- “Human, societal and environmental wellbeing”
- “Human-centred values”
- “Fairness”
- “Privacy protection and security”
- “Reliability and safety”
- “Transparency and explainability”
- “Contestability”
- “Accountability”
For a brokerage, apply those principles through concrete decisions. Name the person responsible for a tool’s output. Give clients a way to challenge an error and have a person resolve it. Explain AI use when it affects the service they receive.
The Voluntary AI Safety Standard sets out ten voluntary guardrails. The department now points readers to the updated and simplified Guidance for AI Adoption, which develops that earlier guidance into six practices.
Start with the foundations guidance for early or low-risk use. Its practices cover allocating responsibility, understanding impacts, managing risks, sharing information, testing systems and maintaining human control. The National AI Centre says to adapt the practices to the organisation’s size and risks.
A small brokerage can adopt its starting actions directly, including appointing an owner and maintaining an AI register. Use a short internal policy and a simple route for reporting problems. More complex or higher-risk uses need the Centre’s detailed implementation guidance, even when the brokerage has few staff.
Keep the legal duties beside this framework. The Office of the Australian Information Commissioner (OAIC) explains how privacy obligations apply to commercially available AI products. Credit licensees must also maintain arrangements to meet their licence obligations, including supervision and dispute resolution. The Australian Securities and Investments Commission (ASIC) explains its assessment in RG 205 Credit licensing: General conduct obligations.
Govern the Client Documents AI Tools Receive
Approve document categories for each tool and task before staff upload client material. A general approval for an AI assistant is too broad to explain what it can read.
Use the following categories in your approval record. These are suggested brokerage rules, not government classifications.
| Client material | Suggested approval rule |
|---|---|
| Public information or invented examples | Permit for approved tasks after checking that examples contain no client details. |
| Client correspondence and fact finds | Require approval for the exact task and the minimum fields needed. Check for financial and health information before use. |
| Payslips, tax records and bank statements | Restrict to a specifically approved processing task. Limit access to the relevant documents and outputs. |
| Identity documents and authentication details | Prohibit in general chat assistants. Escalate any proposed specialist identity workflow for separate approval. Never include passwords or recovery codes in prompts. |
Some financial records reveal health information or other sensitive information through their contents. Removing a client’s name does not establish that a document is anonymous.
For each proposed use, read the terms for the actual plan and inspect its settings. Record how long uploaded files and generated outputs remain available. Distinguish ordinary storage from backups or security logs. Identify whether the vendor can use inputs or outputs for model training and whether the setting applies to every user.
Record where processing and storage occur, including subprocessors that can receive the data. An Australian sales office does not establish Australian storage. A no-training term also leaves separate questions about retention and access.
The OAIC’s workplace guidance recommends actively managing organisational accounts’ privacy settings. It also identifies circumstances where organisations need to prohibit personal information in publicly available generative AI products.
A Statement Converter
In this hypothetical example, the brokerage wants a converter to turn a bank statement into a transaction table. The approval covers bank statements for that task, with staff checking the extracted amounts against the original before using them.
The reviewer records the plan’s retention terms and whether the uploaded statement reaches another AI service. They document the training setting and processing locations. If a required term remains unresolved, staff use the brokerage’s existing approved process until the owner can make a supported approval decision.
Keep the converter’s access limited to the selected statement. It does not need permission to browse the client’s whole folder simply because that folder contains the document.
A CRM Assistant
A customer relationship management system (CRM) assistant can have a different scope from a file upload. Review the integration permissions before enabling it. Determine whether it can read every client record, connected emails or document attachments.
In this hypothetical approval, the assistant drafts an internal meeting summary from selected notes. A staff member checks the draft before saving it. Permission to summarise those notes does not authorise the assistant to send client emails or alter assessed income.
Record the permission settings alongside the permitted document categories. Reassess the use if the assistant gains access to additional records or starts taking actions.
Set Owners, Approvals and an AI Register
Assign the brokerage owner or another authorised senior person to approve AI uses and maintain the register. Record a contact who can suspend access when the usual owner is unavailable.
Before approving a new tool, have its proposed user describe the task and data access. The owner reviews the vendor checks and a trial using invented information. Approval then states the permitted use, required settings and human checks. Obtain any approval your licensee requires before using client information.
Record one row per approved use when the same tool has different permissions. AI data governance needs to show where information goes and what happens to it, beyond the product name.
| Register field | What to record |
|---|---|
| Tool and plan | The product, account owner and exact subscription or contract. |
| Approved task | The use staff can perform, with an accountable person. |
| Data access | Permitted document categories, connected systems and excluded records. |
| Settings | Training options, retention controls and integration permissions. |
| Storage and processing | Countries and relevant subprocessors, supported by the terms or written vendor answers. |
| Output checks | Who checks the result and what they compare it against. |
| Approval evidence | The approver, decision date and location of the supporting review. |
| Review and incidents | The next review date, changes requiring reassessment and any incident record. |
Store the register where staff can find the permitted uses. Restrict access to contracts or review evidence that contains confidential information. A screenshot of a setting supports the decision, but staff also need the rule written in plain language.
Implement AI Governance in Ten Working Days
Use ten working days to establish the brokerage’s initial controls and identify uses that need more assessment. This suggested schedule does not assume every tool will be approved within that period.
- During working days 1 and 2, list current AI use. Include personal accounts and AI features inside existing subscriptions. Ask staff what information each tool receives and whether it can act on their behalf.
- During working days 3 to 5, review each use and record an approval, restriction or retirement decision. Stop unapproved access to client information. For tools being retired, revoke integrations and arrange deletion under applicable retention obligations, preserving records needed for an incident or complaint.
- On working day 6, adopt a written AI use policy that matches the decisions. State who approves new uses and how staff report a mistake. Link the policy to the register so staff can identify the actual permitted tools.
- During working days 7 and 8, brief staff using an ordinary brokerage task. Have them find the approved use, identify the permitted document category and explain who checks the result. Resolve any answer that differs from the register.
- On working day 9, test a hypothetical failure. For example, an assistant changes an income figure while summarising notes. Confirm that staff can stop the workflow and preserve the source record for investigation.
- On working day 10, record unresolved assessments and book the first review. Confirm that every active use has an owner and a decision. Keep restricted uses unavailable until their approval conditions are met.
These records help a licensee trace what happened during an audit or client complaint. The approval explains why the tool was allowed. The register identifies its access, while the source documents and review record explain how a person checked the result.
Reviews, Incidents and AI Agents
Review the register quarterly and review a new use after its first month. These are suggested starting intervals. Increase monitoring when the tool influences client decisions or handles more confidential information.
Send a tool back for approval when its terms change, a new plan changes retention or training, or an integration gains broader access. A model change that affects results also needs attention. Investigate complaints and repeated corrections before continuing the affected use.
An AI agent can perform actions, such as sending emails or updating CRM records. Governing an agent therefore requires permission limits and a way to stop actions, beyond checking a chat response.
For an email agent, restrict the sending account and permitted recipients. Require human approval before client messages are sent. For a CRM agent, restrict the fields it can change and keep an action log. Establish how a person can stop the integration and correct an unwanted change without losing the original record.
The National AI Centre’s implementation guidance calls for human intervention points and more detailed controls for complex or higher-risk use. An autonomous feature requires a new decision even if the brokerage already approves the underlying assistant.
If an AI incident occurs, suspend the affected access first. Preserve the relevant input and output securely, together with action logs and the settings in force. Notify the governance owner and licensee contact, then assess client harm and any notification obligations.
Correct affected records through an authorised person and document what caused the failure. Restore access only after the owner approves the fix and the relevant check passes. The next staff briefing must explain the changed rule before the tool returns to ordinary use.
