AI for the Canadian Public Sector: From Procurement to Production
Canada created a Minister of Artificial Intelligence and Digital Innovation in April 2025. The AI Strategy for the Federal Public Service 2025-2027 was published by Treasury Board Secretariat. The 2025 Budget cited AI as a tool to help the Department of Justice, Public Services and Procurement Canada, Fisheries and Oceans, and Natural Resources Canada meet 15 percent savings targets over three years. Shared Services Canada is developing an AI tool for use across the federal government. The GC AI Register was launched as a minimum viable product to catalogue production AI systems in use across departments. Canada signed AI alliance declarations with Germany, Norway, and Finland in early 2026 and established the G7 AI Network with a mandate to develop a roadmap for scaling AI in public sectors.
The institutional architecture for public sector AI in Canada has never been more developed. The political commitment has never been more explicit. The gap between commitment and production deployment has never been more consequential, because the budget savings targets assume that AI is actually operating in production by the time the savings are counted.
Most federal departments are not there. The pattern is consistent: a pilot that works, a procurement process that takes longer than the pilot, an algorithmic impact assessment that surfaces governance questions that were not addressed before the pilot started, and a production deployment that either does not arrive before the political moment moves on or arrives in a form so hedged by compliance requirements that it does not resemble the pilot that generated the original enthusiasm.
Understanding why that pattern is so persistent, and what the departments that are getting to production are doing differently, is the practical question that federal and provincial public servants working on AI programs need an honest answer to.
The Governance Architecture and What It Actually Requires
The Directive on Automated Decision-Making is the central governance instrument for AI that affects federal program decisions. It has been reviewed and updated four times since its original publication, with the most recent update expanding scope and adding new measures for explanation requirements, bias testing, data governance, Gender-Based Analysis Plus, and peer review for higher-impact systems.
The directive uses an impact level framework to calibrate governance requirements. Level one systems, with low impact on individuals and decisions, have minimal requirements. Level four systems, those that fully automate life-altering decisions with no human review, have comprehensive requirements including peer review, algorithmic impact assessment, bias testing, and plain-language explanation of every automated decision to affected individuals. Most AI systems in federal programs sit at levels two and three, where the requirements are substantial but not prohibitive if addressed systematically from the beginning of program design.
The consistent failure mode is treating the directive as a compliance hurdle at the end of a deployment process rather than as a design framework from the beginning. A department that builds an AI system for benefits eligibility determination and then performs the algorithmic impact assessment before production deployment will discover governance gaps in the system design that require architectural changes, which means rebuilding work that was already done. A department that performs the algorithmic impact assessment at the design stage builds the required explanation capability, the bias testing methodology, and the human review workflow into the system from the start, at a fraction of the cost of retrofitting them later.
The updated GC AI Register, launched as an MVP in 2026, is the first systematic attempt to create a government-wide inventory of AI systems in use across departments. Its existence as an MVP reflects the reality that nobody currently has a complete picture of what AI systems federal departments are running, which is the same model inventory problem that OSFI's E-23 creates for federally regulated financial institutions. Both problems have the same root cause: AI deployment has been faster than governance, and the governance infrastructure is catching up to a production landscape it did not help design.
The Procurement Problem and What It Actually Looks Like
Public sector AI procurement is not slow because public servants are inefficient. It is slow because the procurement system was designed to ensure fairness, prevent corruption, enable audit, and protect the public interest in ways that are fundamentally incompatible with the iterative, experimental character of AI development and deployment.
The standard federal IT procurement process runs through CanadaBuys, with requirements defined in advance through a statement of work, vendors evaluated through a scored assessment, and contracts awarded based on that evaluation. That process works well for procuring technology with well-defined specifications. It works poorly for procuring AI capability, where the specifications for what constitutes a good solution are often not knowable until the solution is being evaluated, the technology landscape changes materially between when requirements are defined and when the contract is awarded, and the iterative improvement that makes AI systems valuable requires ongoing collaboration between the department and the vendor rather than delivery against a fixed specification.
Public Services and Procurement Canada's AI Source List addresses part of this problem by pre-qualifying AI vendors, reducing the time required to move from a recognized AI need to a procurement that can be executed. The Spring 2026 Economic Update included a new SMB procurement program partly aimed at reducing barriers for smaller Canadian AI vendors. The budget's explicit reference to preferred procurement of made-in-Canada sovereign AI tools creates a policy lever for directing public sector AI spending toward Canadian companies.
These are meaningful improvements to the procurement system. They do not resolve the fundamental tension between procurement design and AI development methodology. The departments that are getting AI to production within the constraints of the current procurement system are using procurement instruments differently, not waiting for the instruments to change.
The approaches that work are specific. Using existing software licenses to access AI capabilities within already-contracted platforms rather than initiating new procurements: a department that has Microsoft 365 E5 can access Copilot capabilities within its existing contract, with no new procurement required. Using professional services contracts rather than software contracts for AI development, which allows for the iterative collaboration that software procurement prevents. Using standing offers from the AI Source List for rapid access to pre-qualified vendors when a specific AI use case has been identified and scoped. And using the ProServices and TBIPS standing offers for the data engineering, governance, and change management work that surrounds AI deployment but is not AI procurement itself.
The Use Cases That Are Getting to Production
The public sector AI use cases that are consistently reaching production in Canadian federal and provincial departments share identifiable characteristics that distinguish them from the ones that stall.
They are internal-facing rather than citizen-facing in their first deployment. The governance requirements for AI systems that affect government employees are lower than those for AI systems that affect citizens' access to benefits or services. A document summarization tool for public servants, an internal search system for policy and regulatory information, or a meeting transcription and summarization service can be deployed without the full algorithmic impact assessment required for citizen-facing decision support. Starting internally builds the organizational capability and confidence for more complex external deployments while generating demonstrable value that justifies the investment in higher-governance applications later.
Transport Canada's AI-powered Regulatory Platform, launched in June 2025, exemplifies this pattern. It is an internal tool that transforms regulatory information into structured, searchable data for public servants making regulatory decisions. The direct citizen impact is indirect rather than direct, which reduces the governance burden while producing genuine productivity improvement that is visible and measurable. It is built on a clear data asset (federal regulations), serves a defined user community (regulatory staff), and has a measurable outcome (reduced time to find and analyze regulatory information). Those three characteristics, clear data, defined users, measurable outcome, are the design features that make AI deployments succeed in public sector environments.
They are designed around existing data rather than requiring new data collection. The GC's most tractable AI opportunities are in applying AI to information the government already holds: service transaction histories, regulatory and legislative texts, program application data, correspondence volumes. These data assets exist, they are held under existing legal authorities, and making them more useful through AI does not require new data collection that would trigger additional privacy assessments. AI programs that require new data collection, new data sharing agreements, or new consent frameworks add governance complexity that extends timelines materially.
They have explicit human oversight built in, not added as an afterthought. The Directive on Automated Decision-Making's requirements for human review at levels two and three are not obstacles to AI deployment. They are risk management design requirements that, when built into the system architecture from the start, produce systems that are more trustworthy, more defensible in audit, and more likely to receive sustained ministerial and executive support through the organizational changes that production AI deployment requires. Systems designed to assist human decision-makers rather than replace them are also more acceptable to the public servants whose work they are changing, which addresses the adoption problem that KPMG found is the primary obstacle to AI value realization in Canadian organizations.
The AI Governance Infrastructure That Is Still Missing
The responsible AI framework described in this series has its direct public sector equivalent in the GC's Algorithmic Impact Assessment and the Directive on Automated Decision-Making. The infrastructure that exists is more developed than most people outside the federal government realize. The infrastructure that is still missing is equally important.
A functioning model inventory is the most significant gap. The GC AI Register MVP is a starting point, not a solution. Until the federal government has a current, accurate inventory of every AI system in operation across departments, including the systems that were deployed before the governance frameworks existed, it cannot govern its AI portfolio at enterprise level. Individual departments are managing their own systems within their own governance frameworks, but there is no enterprise view of AI exposure, AI capability, or AI risk across the GC. The AI Strategy for the Federal Public Service 2025-2027's commitment to building out the AI Register is the right priority, and it is not a quick one.
Shared AI infrastructure for common capabilities is the second gap. The most widely used AI capabilities in federal departments, document summarization, internal search, meeting transcription, regulatory analysis, are being built or procured by multiple departments independently. Each department is solving the same problems with different tools, different governance approaches, and different data configurations. Shared Services Canada's enterprise AI tool is designed to address this, but building enterprise tools in a federated government environment requires sustained political sponsorship that persists through the organizational inertia, procurement complexity, and departmental resistance that enterprise platforms always encounter.
AI talent concentrated in mission-critical programs rather than distributed across compliance functions is the third gap. The AI talent shortage documented in the Emerging IT Talent Trends post is more acute in the public sector than in the private sector, and the departments that have AI talent are deploying it disproportionately on governance and compliance documentation rather than on production deployment. Rebalancing toward production with governance built in from the start, rather than governance documented after the fact, is the management challenge that federal AI leaders are navigating in 2026.
What the Renewed AI Strategy Needs to Prioritize
The renewed AI strategy expected in 2026, informed by over 11,000 submissions to the October 2025 national sprint, needs to address production deployment and not only research and infrastructure investment. Canada's AI challenge is, as one senior strategist described it, much less about invention and much more about diffusion, scalability, and execution. The country that was first to launch a national AI strategy in 2017 has produced world-class research and genuine structural disadvantage in translating that research to production value in the public sector and commercial economy.
The specific mechanisms that would accelerate public sector AI production deployment are known and not novel. Procurement reform that matches procurement instruments to AI's iterative development character. Shared governance infrastructure that departments can use rather than each building their own. AI-ready data standards that departments build toward rather than discovering they have not met when AI programs stall. And a performance framework for government AI that measures production deployment and measurable outcomes rather than procurement activity and pilot completion.
The departments and agencies that are not waiting for the renewed strategy to address these mechanisms are the ones that will have production AI programs with documented outcomes when the next political cycle evaluates whether the 15 percent savings targets that the 2025 Budget committed to were realistic. That documentation will shape the next round of AI investment decisions, and the departments that have it will be in a materially better position than those that are still explaining why their pilots have not scaled.
Talk to Us
ClarityArc works with federal and provincial public sector clients on AI program design, governance frameworks, algorithmic impact assessment preparation, and the transition from pilot to production that most public sector AI programs struggle to complete. If you are working on a public sector AI program and want support that understands the GC governance environment, we are ready to help.
Get in Touch