Private AI for a brokerage ranges from hosted chatbots with contractual data protections to local models that keep documents on your own computer. Use an approved business workspace for low-sensitivity drafting. When client documents must stay off a provider’s systems, a verified local setup gives you that control.
The choice depends on what you enter and who must be prevented from accessing it. A paid plan, Australian server or training opt-out answers only part of that question.
What Private AI Means
In its strictest sense, private AI prevents a provider from keeping, training on or reading the brokerage’s data. A local model can meet that goal when processing stays on the brokerage’s device and cloud features are disabled.
Vendors also use private AI to describe narrower protections. A hosted chatbot might exclude prompts from training while retaining chat history and allowing authorised support access. Another might promise zero retention after processing, while still receiving readable text to generate the answer.
Separate the promises before choosing a tool. No training describes a permitted use of data. Zero retention describes what remains after processing. Provider-inaccessible processing describes who can read the content. Encryption during transfer and storage does not, by itself, prevent access while a hosted model processes a prompt.
A secure AI chatbot can protect access well and still retain client information under its terms. Your brokerage needs the protection that matches the document, rather than the broadest marketing label.
Private AI Options From Least to Most Control
The options below generally increase the brokerage’s control over deployment. Australian hosting alone does not give more control than a stronger contract elsewhere. Provider access and retention still need separate assessment.
| Option | Data handling | Brokerage effort | How the cost is structured |
|---|---|---|---|
| Mainstream business chatbot | Hosted processing with business terms. Training exclusions can coexist with saved history and permitted provider access. | Set up accounts, permissions and staff rules. | Usually recurring subscriptions per user. Enterprise terms can require a quote. |
| Privacy-focused hosted tool | Can restrict training and retention more tightly. Readable processing and exceptions depend on the service. | Review the contract and configure approved models or modes. | Subscriptions or usage charges. Additional protections can belong to a different tier. |
| Australian-hosted model or service | Can keep specified storage or processing in Australia. Offshore support and subcontractors remain separate questions. | Obtain a written description of every processing location and access route. | A service subscription, usage charge or managed-service quote. |
| Model run locally or on premises | Prompts can remain on brokerage-controlled equipment when external processing and syncing are disabled. | Maintain devices, model software, access controls and document storage. | Hardware and support costs, plus any software or model licence. |
ChatGPT Business is an example of the first category. OpenAI’s business privacy commitments exclude business content from model training by default. The same policy permits specified human access for support, abuse investigation and legal compliance. Those are different protections from preventing the provider from reading the content.
A privacy-focused service needs a narrower written promise about each model and feature you use. A no-retention chat mode does not establish the rules for uploaded documents, web search or connected storage. Treat those as separate routes through the service.
Local models remove the hosted provider from the prompt-processing path when the whole workflow runs locally. They also move security and maintenance responsibilities to the brokerage. An online document converter or cloud-connected add-on can send the same material outside that boundary.
Which Option Fits a Brokerage
Match the AI option to the information the task needs. Email drafting without client details has a different privacy requirement from summarising a named client’s bank statements.
| Broker task | Practical option | Accuracy and convenience trade-off |
|---|---|---|
| Draft a general follow-up email | Approved business chatbot using a generic scenario. Add identifying details afterwards in the brokerage’s own system. | Fast to use. The broker still checks promises, tone and facts before sending. |
| Summarise client statements | A brokerage-approved document service with suitable terms, or a verified local workflow when external disclosure is prohibited. | Document extraction can omit transactions or misread figures. Reconcile the summary to the original statement. |
| Search public lender policy | A hosted search assistant using a client-free question. | Links help locate documents, but the broker must confirm the lender’s actual policy and applicable conditions. |
| Search private broker policy documents | A permission-controlled hosted collection or local document search using authorised copies. | Convenience depends on keeping documents current and preserving who may read each source. |
Consider this hypothetical task: a broker wants wording explaining a variable-rate review. A generic prompt needs no client name, address or account balance. The same broker asking AI to identify every recurring payment from a bank statement exposes a much larger set of personal information.
Choose the simpler approved hosted option for the first task. For the statement task, decide whether the brokerage permits that disclosure before choosing the tool. Local processing is useful when keeping the document on controlled equipment is a requirement and someone can maintain the setup.
Privacy does not establish answer quality. A local model can make factual mistakes, and a hosted model can cite the wrong policy. The Office of the Australian Information Commissioner (OAIC) recommends human oversight and product due diligence in its commercial AI privacy guidance. It also recommends keeping personal information out of publicly available generative AI tools.
Judging a Secure AI Chatbot
Judge a private AI claim against the exact plan, model and document feature your staff will use. Ask for terms that answer the following questions before approving client information.
- What content is retained, including prompts, uploaded files, extracted text and generated answers? Request the retention and deletion schedule, including backups and legal exceptions.
- Can the provider or its model suppliers train on the content? Request the applicable business terms and any data processing agreement.
- Who can read the content during support, safety checks or incident investigation? Request the access policy and subcontractor list.
- Where does storage and model processing occur? Ask separately about document extraction, support access and external search.
- Which settings can staff change? Request the administrator guide covering sharing, connected services and account removal.
- What evidence supports the security claims? Request the relevant assurance report or certification scope, plus incident-notification commitments.
Use a fictional document for a small approval check. Confirm that an ordinary staff account cannot retrieve another staff member’s restricted file. Delete the test document and establish what remains in history or storage. Keep the contractual deletion period separate from immediate disappearance in the interface.
For a proposed local workflow, have the person maintaining it demonstrate processing without network access. Include document extraction and saved outputs in that check. A local chat window that silently calls a cloud model fails the requirement to keep prompts on the device.
Record the permitted tasks and prohibited information once the option meets your requirements. An answer such as “encrypted and private” leaves provider access and retention unresolved until the supplier explains both.
Sovereign and Australian-Hosted AI
Australian-hosted AI can reduce overseas handling when the relevant data and processing actually stay in Australia. Sovereign AI usually promises broader control over operations and access. Neither label proves that all subcontractors, support staff or backups are local.
Australian Privacy Principle 8 (APP 8) concerns cross-border disclosure of personal information. For a brokerage covered by the Privacy Act, the OAIC’s APP 8 guidance generally requires reasonable steps before disclosure to an overseas recipient. The brokerage can remain accountable for the recipient’s handling. The guidance also distinguishes disclosure from use within the brokerage’s effective control.
An Australian storage location does not settle that assessment. Offshore processing or access can still require consideration. Local hosting also leaves questions about authorised use, accuracy and protection against unauthorised access.
OpenAI lists Australia as a storage-at-rest region for eligible new ChatGPT Enterprise and Edu workspaces. Its residency documentation lists model-processing residency separately, with Australia absent from that list. Some integrations and transient processing sit outside the storage guarantee.
ChatGPT Business has a separate residency rollout. OpenAI says Business safety-monitoring copies remain in the United States, including a temporary copy of prompts and responses when another storage region is selected. A regional setting therefore needs to be read alongside the plan’s actual data terms.
Is Perplexity Safe for Client Information?
Perplexity is a useful hosted option for researching public lender information without entering client details. It is not a reason to upload an identifiable client file without brokerage approval. Consumer and Enterprise plans have different training and retention rules.
Consumer Plans
Perplexity’s data collection policy says Free, Pro and Max enable AI Data Retention by default. That allows data collection for model training.
To opt out, open Account settings, select Preferences and turn off AI data retention. The change applies to later data. It does not remove material already collected for training, and operational processing continues.
Perplexity’s privacy notice covers prompts, queries, uploads and generated outputs. Its account-retention guidance says personal information remains while an account is active. Saved conversations require their own deletion.
The file privacy guidance states a 30-day period for ordinary session uploads. Earlier questions and answers can retain context after the file becomes unavailable. Project and personal-repository files remain until deleted under their respective policies.
Enterprise Plans
Perplexity says Enterprise data is excluded from training its own models. Its model-provider agreements also prohibit external providers from retaining or training on that data.
Enterprise session uploads have a seven-day deletion policy. That does not delete the conversation holding the extracted information. Projects have separate retention, as the Enterprise file limits explain.
Custom session retention requires an organisation with at least 50 seats or one Enterprise Max user under the Enterprise retention controls. Enterprise subscriptions are priced per seat, with monthly or yearly billing.
For public policy research, ask a question such as “Where does this lender publish its policy for self-employed applicants?” Open the lender’s source and confirm its conditions. Keep the client’s name, application documents and identifying financial details out of the search.
