Reviewing AI tools already used at work
Find AI features used by your team, review the data and connectors they handle, and set clear rules for each use with proportionate privacy safeguards.

Your team may use AI through a chatbot, a meeting assistant, a browser extension or a feature inside software you already pay for. An application you already trust can gain a new AI feature without your business reviewing how that feature handles information.
A shadow AI review finds the AI uses the business hasn’t checked yet. The useful result is a record of the work being done, the information involved and the conditions under which the team can continue using it.
Ask about the task and the account
Start with a short conversation: what are people using to draft, summarize, transcribe, search or analyze work? Ask whether they use a business account, a personal subscription or a free service, and whether the tool connects to email, files or another application.
Record one entry per use case. A tool that’s fine for public marketing copy may need a separate review before processing a client’s records. Capture the task, department, account type, input categories, connected systems, output use and a person responsible for the decision.
Sort the inputs into a few plain categories, such as public, internal, confidential and regulated, so staff can apply the rules without having to interpret policy language.
Use descriptions of the data during discovery. Don’t ask staff to submit copies of sensitive prompts or client files just to show that a tool is being used. Provide a private route for reporting a possible exposure.
Review what leaves the business
Check the terms and settings for the exact product and subscription. Establish what information is sent, how long it’s retained, who can access it, whether it’s used for model training and how deletion works. Include uploaded files, transcripts and connected data sources.
The Cyber Centre’s generative AI guidance recommends assessing providers, defining permitted uses and reviewing generated content. Answer those questions for the configuration your team is using, whatever the product’s name suggests.
Inspect connector permissions separately. Identify whether a connection uses one person’s access or broader application permissions, and whether it can send messages or change records. Remove access that isn’t needed and confirm how to disconnect it. Use the Copilot permissions guide for the separate task of reviewing existing Microsoft 365 access.
Keep the review proportionate
Canada’s privacy regulators’ generative AI principles address appropriate purposes, necessity, transparency and accountability. Check which privacy laws apply before putting personal information into an AI workflow. Removing a name alone may leave enough detail to identify someone.
If discovery needs employee monitoring, define its purpose, collection, access and retention first. The OPC’s workplace privacy guidance explains why transparency and less intrusive approaches matter. The law that applies depends on the province, sector and employment context. Don’t treat a signed policy as blanket permission to inspect personal accounts or browser histories.
Decide what’s permitted
For each use case, record an actionable decision:
- Approved: name the business account, permitted data and required review.
- Limited: specify the inputs or actions that aren’t allowed.
- Replace: assign an owner to move the work to a tool that’s been reviewed.
- Stop: explain the concern and provide a workable way to complete the task.
Start unresolved trials with public or synthetic data. Set a reviewer for outputs before they affect a customer, employee or business record. Permission to generate a draft doesn’t extend to letting an agent send it or update a system automatically.
Recheck when the workflow changes
Review these decisions when a vendor changes terms, an integration gains access or the team wants to use a new data category. Record unresolved questions and an owner. A tool can still be in use when it’s missing from a discovery report.
For the broader software inventory, see finding unsanctioned cloud apps. Contact us if you need help reviewing AI workflows and the systems they connect to.


