The 2026 Canadian Sovereign Agent Deployment Playbook
Twelve chapters that take one knowledge workflow from candidate to a production decision.
Insights
Our method for putting AI to work, published in full
A playbook, three whitepapers and guides for four kinds of organization. Start with the featured work, then go to the section that fits yours.
Start here
The playbook and the whitepapers
The long reads the rest of this page builds on.
Designing for Canadian AI sovereignty
Sovereignty as eight control decisions, with the architecture trade-offs and the split of responsibility.
Twelve control domains for a governed AI agent
The control model in full, and the six architecture stages it maps onto.
Choosing a first AI workflow
Four tests for a good first job, four examples and the jobs to leave for later.
Small business
Plain guides for a small business
Four short guides on real jobs, what to check and what to leave alone.
Pick a first job that a person already checks
Four tests for a good first AI job, four examples and the jobs to leave for later.
Answer customer questions from your own documents
The AI drafts the reply from your price list and policies. A staff member checks it and sends it.
Sort what comes in without losing control of it
Email, forms and invoices get labelled and routed. Anything unclear goes to a person.
Test your AI on real cases before customers see its work
Build a test set from your past work, score four things and run it again after every change.
Enterprise
For a larger organization
Evidence, cost and supplier questions for a team taking AI past its first project.
The business case for a six-week knowledge agent pilot
The published evidence is mixed, so your own organization has to test it.
Open-weight versus proprietary models
Ask it as arithmetic. The answer differs by workload inside the same organization.
The eight control decisions, put as questions to a supplier
Asking where data sits answers one question in eight. These are the other seven.
More for a larger organization
- Canada is adopting AI faster than it is governing it. Business adoption has tripled in two years while governance maturity has not kept pace.
- Ottawa is asking five questions about AI transparency. Can your organization answer them?
Public sector
For a public institution
Screening, assessment, architecture and procurement for Canadian public institutions.
When does an AI assistant trigger an Algorithmic Impact Assessment?
The answer changes the design, the timeline and the review path.
How to complete an Algorithmic Impact Assessment
The score sets who approves your system, whether a human decides and how many experts review it.
A practical Protected B AI architecture checklist
A checklist for Protected B deployments, starting with the uncomfortable fact underneath each one.
More for a public institution
- How to write a procurement-ready AI pilot statement of work. Say what the pilot decides, name the security evidence and let the buyer test what it bought.
- What the federal AI register shows. The Government of Canada AI Register, read as data.
- Human approval patterns for government AI agents. Who approves what the agent does, and at which moment.
- Five knowledge workflows Canadian public institutions can pilot safely. Settle what regulates a pilot before it starts.
- A governed research assistant for policy and program teams. Retrieval on an approved corpus, with every citation bound to its source passage.
- Bilingual evaluation for Canadian government AI workflows. An evaluation written and tested in English can pass while French-language users get wrong answers.
Regulated industries
For a regulated organization
Model risk, operating controls and evidence for institutions that answer to a supervisor.
What has to be true before OSFI Guideline E-23 takes effect
The guideline takes effect on 1 May 2027 and covers AI and machine learning models.
OSFI Guideline E-23 and what it means for an AI agent
An agent needs no parallel governance regime, only the evidence that validators and supervisors already expect.
From demonstration to operating service: the AgentOps control model
A demonstration and an operating service answer different questions.
Next step
From reading to a decision
Bring one candidate workflow to a 30-minute call. You keep the notes whether or not we work together.