Broker guide
Private AI for Mortgage Brokers: Keep Client Data Private
Private AI keeps client data out of public chatbots. Compare business plans, private cloud deployments and local models for a brokerage.
- Published
- Updated
Private AI is AI whose provider cannot retain, train on or disclose your data beyond the agreed contract and applicable law. A brokerage can use a business assistant, a private cloud deployment or a model on its own computer. The choice depends on who can access client information and who will maintain the system.
For a small brokerage, a managed business workspace is usually the simplest starting point. Choose private cloud or local models when you need controls the managed service cannot provide. In every setup, approve the specific use of client data before staff upload a file.
What Private AI Means
Private AI puts a defined boundary around how the system handles your information. That boundary covers the provider’s permitted use of prompts and uploaded files, as well as the answers it produces. It also determines who can access stored conversations and how long copies remain.
The word “private” alone tells you little about those arrangements. A private AI platform can still run on infrastructure shared with other customers. A model running inside your office can still send information elsewhere through a connected feature.
For example, a brokerage might approve an assistant to summarise a client’s employment history from a fact find. The approval needs to cover that document and its contents. Approval to draft a generic marketing email doesn’t establish permission to upload the client’s file.
The Office of the Australian Information Commissioner (OAIC) recommends keeping personal information out of publicly available generative AI tools. Its guidance on using commercial AI products also calls for due diligence on the intended use. A private arrangement still needs that assessment.
Three Ways to Run Private AI
Business plans, private cloud deployments and local models give your brokerage different levels of operational responsibility. Compare the complete service, including document storage and support access, before deciding where the model will run.
| Setup | Control over client data | Cost structure | Work for the brokerage |
|---|---|---|---|
| Business plan of a public assistant | Contract and workspace settings govern provider processing. The provider operates the service. | Subscription billing, with plan-specific limits | Manage staff accounts and approved settings. Review connected services. |
| Private cloud deployment | Your deployment can restrict network access and choose storage arrangements. Hosting and model-provider terms still apply. | Hosting, model usage and implementation costs | Configure access and retention. Maintain the application and monitor it. |
| Local model | Prompts can stay on your own device when every processing step runs locally. | Hardware and staff time, plus any software licensing | Secure the device and maintain the model. Manage saved files and backups. |
A business workspace fits a team that wants an operated service and can accept its processing terms. A private cloud deployment suits a brokerage with an IT owner and specific requirements for access or integration. Local processing fits work that must stay on an approved device, provided that device can run the required model.
“Self hosted” can mean a server rented from a cloud provider. “On premise” means the system runs on infrastructure at your premises. Neither phrase establishes that support logs or backups stay in the same place.
What to Check in a Private AI Contract
A private AI contract must explain permitted processing and what happens to every copy of client data. Read the data processing agreement alongside the service agreement and your order form. A marketing statement about privacy cannot supply missing contractual terms.
| Contract term | What the brokerage needs to establish |
|---|---|
| Training | Whether prompts, documents and outputs are excluded from model training, including any opt-in exceptions |
| Retention period | How long conversations, uploaded files and operational logs remain, including backups |
| Deletion | What deletion removes, the time allowed and exceptions for legal retention |
| Storage location | Where each category of data is stored or processed, including support access from other countries |
| Sub-processors | Which other organisations process the information and how changes are notified |
| Breach notice | When the provider must notify you and what assistance it must give |
For a named example, OpenAI’s data processing addendum says it processes customer data “only in accordance with Customer Instructions”. Those instructions include the agreement and service configuration, with an exception for applicable legal requirements. The addendum also addresses sub-processors and deletion after the agreement ends.
ChatGPT Business is covered by OpenAI’s Services Agreement. That agreement excludes using customer content to develop or improve the services unless the customer explicitly agrees. It still permits processing to deliver the service, comply with law and prevent abuse.
These terms show why a training exclusion and a retention promise answer different questions. A provider can refrain from training on a document while retaining it to provide the service. Your approval must address both uses.
Secure AI Chatbots for Client Work
A secure AI chatbot needs protections that match the client information and the task. A familiar brand or popular assistant doesn’t establish that fit. For document analysis, assess access to the original file as well as the resulting summary.
In a business workspace, give staff individual accounts and restrict administrative access. Review sharing permissions and connections to other services. A managed assistant such as ChatGPT Business is an example of this setup, with the contractual limits described above.
In a private cloud deployment, restrict the application to approved staff and protect the model endpoint from public access. Include document extraction, search indexes and logging in the assessment. Sending a PDF to an external extraction service creates another processing step, even when the model itself is private.
For a local model, keep processing on the approved device and disable external features for confidential work. Your local large language model guide explains the hardware and setup decisions. A local large language model (LLM) generates text on your computer, but saved chats and synced folders still need protection.
A hypothetical controlled workflow starts with an approved document in restricted storage. The assistant receives only the information needed for the task. The broker checks its summary against the original before adding it to the client record.
When Private AI Falls Short
Private AI cannot prevent a staff member from uploading the wrong file or sharing an answer with the wrong person. Your brokerage needs an AI policy that names approved tools and permitted tasks. It must also explain which information staff can enter and how to report an accidental disclosure.
Limit access to people who need the information for their work. Train staff on the difference between a personal account and the approved business workspace. Include the handling of uploaded documents and the removal of access when someone leaves.
Privacy also doesn’t establish accuracy. An assistant can keep a document within the agreed processing boundary and still misread income or invent a detail. Require a broker to check consequential answers against the source before they influence the file.
The OAIC’s guidance treats personal information in AI outputs as part of the privacy assessment too. Include generated summaries in your access and retention rules. Approve a defined workflow with a named owner before extending private AI to another type of client document.
