ShapedSI Journal · Applied AI
Private AI workflows in Canada: start with the data boundary
A private AI workflow needs more than a reassuring label. Before discussing a model or a server, define what information the system may read, where that information goes and who can approve an action. That boundary is the foundation for an internal assistant, a document workflow or a tool connected to business records.
In this guide
01
Define the permitted information
List the specific sources the workflow needs. A policy assistant might only need an approved set of documents. An enquiry summariser might only need the current request and a service list. Avoid giving an early pilot access to an entire shared drive simply because connecting the drive is convenient.
For each source, record its owner, permitted users, update process and retention expectations. Identify records that should remain outside the pilot. Use fictional or appropriately anonymised examples when exploring a design. A narrow source list makes the first test easier to understand and reduces the number of permissions you have to manage.
02
Ask what private means in the proposal
Clarify where prompts, files and outputs are processed and stored. Ask which organisations can access them, what logging is enabled and whether the provider may use them for another purpose. A local interface does not, by itself, establish where processing happens. Treat hosting location, account controls and data handling as separate questions.
Ask for the settings and contractual terms that support the proposed arrangement. A provider should explain which controls are available in the selected service and which responsibilities remain with your team. Requirements can differ by organisation and information type; involve the appropriate privacy or security owner before connecting sensitive records.
03
Keep access and actions distinct
Reading a document and changing a customer record are different capabilities. Decide which users can retrieve information, which can prepare a draft and which can approve a write. An assistant that answers internal questions does not need the same permissions as an agent that sends messages or updates a system.
Plan how access is removed when a staff member changes roles. Test that one user cannot retrieve information outside their permitted sources. Include a route for reporting an incorrect answer or an unexpected disclosure. Keep enough operating records to understand what happened, while avoiding unnecessary copies of sensitive content.
04
Operate the boundary after launch
Check what happens when a source is replaced, a document becomes outdated or an integration fails. Assign somebody to keep the approved knowledge current. The assistant should be able to say that it lacks enough information, and staff should know where to check the original source.
Before expanding the pilot, review the permissions and the actual examples collected during use. More data is not automatically a better design. Keep additions deliberate, update the instructions and repeat the access checks. A private workflow remains a managed business system throughout its life, including when you eventually retire it.
Bring this to the conversation
A working brief.
- An approved source inventory
- Processing and storage locations
- Documented provider data terms
- Separate read, draft and write permissions
- An owner for access, retention and updates
A few useful questions.
Does private mean everything must run on our own computer?
That is one possible arrangement. The right design depends on the information, access requirements, operating capacity and provider terms.
Can we start without real customer records?
Yes. Fictional examples can help test the structure of the workflow before anyone connects sensitive information.
Original sources
Sources and further reading.
Canadian privacy commissioners’ generative AI principles address necessity, transparency, accountability and safeguards. Use that guidance as a starting point for your organisation’s privacy review.
Shape the next step