Broker guide
AI Risk Assessment for a Brokerage Loan Task
Approve an AI task before client data leaves your brokerage. Record the tool, data, controls, residual risk and a clear approval decision.
- Published
- Updated
An AI risk assessment helps you decide whether a specific tool can handle a particular loan-file task with the data you intend to give it. Record an approval, a conditional approval or a decline before anyone sends client data to the service.
Define the Tool, Task and Data Boundary
Start with one tool and plan, one workflow and one class of data. Name the workspace and input method. An approval for a brokerage-managed workspace with manual uploads covers that setup alone.
A personal account or an automated connector needs its own assessment.
Record who owns the task, where the input comes from and what the tool is meant to produce. List everyone whose information can appear in the input, including joint applicants and people named in transaction descriptions. Identify the systems that receive the input or output and the staff who need access.
Use synthetic data when you’re testing the process. For live work, determine whether a smaller extract preserves what the processor needs to see. If the extract still identifies someone or can reasonably be linked back to them, handle it as personal information.
Map the Data Journey and Failure Consequences
Follow the data from the loan file to the tool and back. At each step, identify who controls the copy, who can access it and what will happen when you need to retrieve or delete it.
| Stage | What to record |
|---|---|
| Preparation and upload | Who prepares the extract and where temporary copies sit. Identify which account receives it and how staff check the file before sending it. |
| Processing and storage | Identify the service that receives the file and any intermediary that forwards it to a model provider. Record where the prompt, attachment, output and logs are stored, who can access them and for how long. |
| Sharing and return | Check whether staff, workspace administrators, provider personnel or other processors can see the content. Record every onward transfer, including transfers through apps or shared links. Name the brokerage system that receives the reviewed output. |
| Retrieval and deletion | Record who can retrieve the output for a file review and who can remove working copies. Check what deletion covers across the brokerage, any intermediary and the model provider, including backups and retention exceptions. |
Assess what would happen if each control failed. An original statement sent to the wrong workspace could disclose account details and unrelated transactions. An omitted debit could understate expenses.
If a broker treats the draft as evidence, that error could carry into the loan submission. If the brokerage cannot recover the checked output or remove redundant copies, it may struggle to explain a decision or control further access.
For brokerages covered by the Privacy Act, check whether your proposed use or disclosure is permitted under Australian Privacy Principle (APP) 6. Assess whether APP 8 applies to any disclosure to an overseas recipient. Under APP 11, take reasonable steps to protect personal information you hold.
Document the purpose for which the information was collected and any applicable exception if the new use or disclosure is for a secondary purpose.
The Office of the Australian Information Commissioner (OAIC) recommends keeping personal information out of publicly available AI chatbots in its guidance on commercial AI tools.
Assess a Bank Statement Summary for a Loan File
In this hypothetical file, a processor wants a draft list of recurring expenses for the broker’s review. The proposed tool is ChatGPT Business in a brokerage-managed workspace using a manual upload. The original statement stays in the loan file for reconciliation.
Uploading the original PDF exposes the applicant’s name, address, account details, balances and unrelated transactions. The processor can create a smaller transaction sample in the brokerage’s own system. Include row references, relative weeks, debit amounts and generic payee labels that preserve the pattern.
Remove names, addresses, account numbers, balances and unrelated rows. Strip free-text references and hidden document data by making a fresh, plain-text extract and reviewing it before upload.
Suppose the sample shows a $45 utility debit in weeks 1 and 5, with a $45 reversal in week 5. The expected output identifies the two debits and flags the reversal for review.
Before using the summary, the processor checks each amount, direction and row reference against the original. They then scan the full statement for recurring expenses the sample omitted. The broker uses the reconciled result to assess expenses and keeps the source in the file.
Check the workspace’s actual agreement, applicable processing terms and settings for the proposed upload. OpenAI’s business data controls exclude ChatGPT Business content from model training by default and describe separate rules for access and retention.
Check who can view or export content, where it goes, which subprocessors receive it and how each copy is deleted. Compare the data route with the data processing addendum if it applies to your workspace. Confirm that apps and sharing routes outside the assessment are unavailable for this task.
Decline the original PDF for this workflow. A conditional approval for the reviewed extract can require documented terms, confirmed settings and a passed synthetic trial. Keep live inputs inside the brokerage if the service, plan or upload route lacks approval for that data.
Complete the AI Task Approval Record
Copy this record into your brokerage register for a single task. Fill in the evidence and the decision before staff start using live inputs.
Task owner and authorised users:
Tool, plan, workspace and input method:
Purpose and intended output:
People affected and input sources:
Receiving systems and output recipients:
Permitted data:
Prohibited inputs:
Sources of assurance and where they are filed:
Risks and consequences:
Controls and who checks them:
Residual risk after controls:
Decision and reasons:
Conditions to meet before live use, if any:
Approver and approval date:
Review trigger and owner:Approve when the proposed input and workflow are within the agreed boundary and the controls work. For example, the manager can approve the configured workspace for staff training with invented transactions.
Approve with conditions when a measurable restriction makes the task acceptable. For the loan-file example, require a checked extract, authenticated workspace access and signed reconciliation before the result enters the file. Record residual risk, such as re-identification from a distinctive payment pattern, and restrict inputs further if needed.
Decline when the remaining exposure is unacceptable. For example, decline a full statement upload to a personal account when you cannot establish who can access the file or how long the service keeps it.
Base the decision on the actual evidence and applicable duties. A numerical risk rating cannot settle those duties.
Verify Controls and Reassess Changed Uses
Test the intended account and upload method with invented transactions before permitting live inputs. Record what happened against these expected results.
- An authorised staff member can complete the task, and a staff account outside the approved group cannot open the input or output.
- The draft identifies both $45 debits and flags the reversal. The processor records any discrepancy and tests the reconciliation step against the synthetic source.
- Staff can retrieve the checked record from the loan system and remove the test copies from working locations. Check the deletion controls for the chat and attachment separately. Use the agreement to document how provider-side deletion and retention work beyond what the interface shows.
If any required control fails, hold the live workflow until the task owner fixes it and repeats the failed check. File the approval and test results in the brokerage register alongside the terms and settings. Reassess the workflow before using it after a change to the task, tool terms, connectors, data class or output recipient.