Shadow AI: The Risk of Staff Using Unapproved AI

What Shadow AI Looks Like

Shadow AI includes an unapproved chatbot, browser extension or AI feature inside software your brokerage already uses. The National AI Centre’s shadow AI guidance defines it through use outside an official policy or approved approach.

Several brokerage tasks can involve shadow AI:

• A loan processor pastes a client’s income details into a personal chatbot to draft an email.
• A broker uploads a fact find to an unapproved summarising tool.
• A meeting assistant records a client conversation without the business approving that service.
• A staff member enables an AI extension that can read work pages.

These are hypothetical examples. The common problem is that the brokerage hasn’t assessed the tool and the way the employee uses it.

ChatGPT can be shadow AI when staff use it outside your approved approach. A business-approved account used for an allowed task falls within that approach. Approval must cover the account and use case, because permission to draft generic marketing copy doesn’t authorise uploading client documents.

A Personal-Account Example

Imagine a processor wants a clearer email about missing documents. They paste the client’s name, address and income explanation into their personal account, then copy the answer into your email system.

The finished email looks ordinary. Your client file may contain the message without showing where the source information went. Removing the client’s name afterwards doesn’t undo the earlier upload.

An approved workflow starts with a generic request for missing documents and adds client details in the brokerage’s email system. Where a task genuinely needs personal information, approve that handling separately before staff use it.

The Risks of Shadow AI

Shadow AI exposes your brokerage to client-data handling you haven’t assessed and answers nobody has checked. It also leaves you without a reliable record of what staff shared.

The Office of the Australian Information Commissioner (OAIC) recommends avoiding personal information in publicly available generative AI tools as a matter of best practice. Its commercial AI privacy guidance explains that privacy obligations apply to AI inputs and outputs where the Privacy Act covers the business.

In practice, focus on these risks:

• Client documents leave approved systems. A payslip, bank statement or fact find can reveal personal information beyond the text the employee wanted summarised.
• An incorrect answer enters the file. A generated summary might omit a liability or change an employment detail, affecting the broker’s assessment.
• The business can’t reconstruct the work. A personal account can leave the original prompt and response outside your file records.
• Access remains with an individual. When an employee leaves, the brokerage may lack control over the account holding work conversations.

A polished paragraph doesn’t show whether the underlying facts are correct. Compare any AI-assisted client summary with the original documents before relying on it.

If you discover a client-data upload, stop further sharing and record the tool, account, information involved and people who could access it. Preserve relevant evidence through your incident process. Escalate to your privacy or compliance contact to assess containment and any notification obligations.

How to Detect Shadow AI

Start by asking staff about their actual tasks, then compare their answers with the software and accounts your brokerage controls. A small brokerage can do this without buying a monitoring product.

1. Ask each person which AI tools they use for work. Include personal accounts, free trials, meeting assistants and browser extensions. Ask what job each tool helps them finish.
2. Record the information each task uses. Distinguish generic writing from client emails, uploaded documents and recorded conversations. Keep client information out of the inventory itself.
3. Review business records you can legitimately access. Look at software subscriptions, expense claims and approved integrations. With staff, check extensions and AI settings on work devices under your workplace policies.
4. Sample a few completed tasks. Ask the employee to explain how they produced a recent summary or email and where they checked its facts. A change in writing style alone doesn’t prove AI use.
5. Compare the findings with your approved list. Mark each use as approved, requiring assessment or stopped. Give the follow-up to a named person.

Explain the purpose before the discussion: you need to understand data handling and give staff usable tools. The National AI Centre recommends making it safe for people to disclose their AI use.

An empty expense report doesn’t establish that staff aren’t using AI. Free tools and personal accounts won’t necessarily appear there. Repeat the conversation when staff introduce a new workflow or software feature.

How to Prevent Shadow AI

Prevent shadow AI by approving useful tools for specific jobs and teaching staff the limits of that approval. A blanket instruction to avoid AI leaves the original task and time pressure unresolved.

Before approving a tool, assess its data handling for the intended task. Review who can access inputs and outputs, retention and deletion terms, model-training use and account controls. Record the approved account type and settings with the decision.

Write those decisions into your AI policy template. State which tools staff can use, what information they can enter and who checks the result. Include a route for requesting a new tool and reporting accidental sharing.

The Australian Cyber Security Centre’s guidance on engaging with AI recommends checking generated results and understanding how tools store and use data. It also advises following organisational policies and obtaining approval before entering sensitive information.

Make AI training for staff part of the workflow. Demonstrate an allowed task with fictional client details, then show an upload that breaches your policy. Ask each employee to identify the problem and explain the approved route.

For example, an exercise can pair a generic document-request email with a full client fact find. Staff must recognise which information the approved drafting task permits. The person reviewing the result then checks its factual content before use.

Keep approval decisions under your wider AI governance process, with someone responsible for reviewing changes in tools and tasks. Before closing a shadow AI finding, confirm that the unapproved use has stopped and the employee can complete the task through the approved workflow.