The Conversation Everyone Is Having in the Wrong Room
Federal and provincial governments across Canada are producing AI strategies, establishing AI centres of excellence, and publishing ambitious transformation roadmaps. Ministers announce AI commitments. CIOs publish priorities. Departments form working groups.
And then the procurement process begins, and progress stops.
The honest diagnosis is one that few consultants are willing to give their government clients: the barrier to AI adoption in the public sector is not technical readiness, staff reluctance, or executive buy-in. It is the procurement system itself. SOW structures, standing offer arrangements, and evaluation frameworks built for traditional IT delivery are structurally incompatible with how effective AI work actually gets done.
Until that changes, AI transformation in government will consist largely of pilots that never reach production.
How Traditional IT Procurement Works — And Why AI Breaks It
Government IT procurement is designed around knowable, specifiable outcomes. A department identifies a need, writes a statement of work, solicits bids against defined requirements, evaluates against a scoring matrix, awards a contract, and holds the vendor accountable to the specified deliverables.
This process works tolerably well for infrastructure projects, software licences, and systems with stable requirements. It fails almost completely for AI work.
Effective AI development is inherently iterative. You learn what works by building and testing, not by specifying in advance. A language model that performs well in controlled evaluation may behave differently against real operational data. An automated workflow that looked efficient in design may create unexpected bottlenecks in practice. The requirement that seemed clear in the SOW turns out to be underspecified once the team starts working with actual data.
Iterative discovery is not a vendor failing — it is how AI development works. But government procurement frameworks treat scope changes as problems, budget adjustments as red flags, and evolving requirements as evidence of poor planning. The incentive structure pushes vendors to over-specify upfront, deliver what was written in the SOW regardless of whether it solves the actual problem, and avoid raising issues that might trigger procurement review.
The Standing Offer Problem
Many federal departments procure AI services through standing offers and supply arrangements — frameworks that were established to streamline purchasing of standard services. The logic is sound for commodity procurement. For AI work, it creates serious problems.
Standing offers tier vendors by price category, which means departments are often selecting from vendors who won the supply arrangement years ago, before AI capabilities matured, by pricing commodity services. The vendors who are actually capable of delivering complex AI work may not be on the arrangement at all, or may be positioned in categories that don't reflect the nature of the engagement.
The task-based professional services frameworks are particularly ill-suited. AI transformation work spans strategy, data engineering, model development, change management, and ongoing iteration. Forcing this into standardised task classifications means either selecting the wrong task types or creating a patchwork of separate engagements that nobody is accountable for integrating.
What Actually Needs to Change
The fix is not to work around procurement — it is to reform how AI work is procured. Several specific changes would have material impact.
Phased contracting with defined gates. Rather than specifying the full engagement in a single SOW, structure AI procurement in phases: discovery, proof of concept, production development. Each phase has clear deliverables and a decision point. The department retains the right to proceed, pause, or change direction. Vendors are accountable to phase outcomes, not to a multi-year specification written before the team understood the problem.
Outcome-based evaluation criteria. Scoring matrices that weight technical compliance over demonstrated AI capability consistently select the wrong vendors. Evaluation criteria need to reward evidence of past AI delivery — working production systems, measurable outcomes, relevant domain experience — not the ability to write a compelling response to a requirements checklist.
Agile-compatible contracting language. The Treasury Board Secretariat has published guidance on agile procurement. That guidance is not consistently applied, and departments that want to use it often face pushback from their own procurement functions. Making agile contracting the default for AI engagements, not the exception, requires clear policy direction and procurement officer training.
Shorter initial engagement windows. A 12-week engagement to validate an AI approach and demonstrate proof of concept is far more valuable than a 3-year contract structured around uncertain requirements. Short initial engagements reduce risk, create clear accountability, and allow departments to change vendors or direction without being locked in.
The Role of Leadership
Procurement reform in government requires political will, not just process improvement. CIOs and CDOs who are serious about AI transformation need to be having explicit conversations with their procurement functions about the structural mismatch between AI work and current contracting frameworks.
This is not a comfortable conversation. Procurement officers are not wrong to apply the frameworks they have — those frameworks exist for legitimate reasons, including accountability, fairness, and audit requirements. The problem is that the frameworks were not designed with AI delivery in mind, and nobody has made changing them a priority.
That changes when leadership makes it one.
The departments that will deliver on their AI commitments in the next two years are those where the CDIO and the Chief Procurement Officer are aligned on what iterative AI contracting needs to look like — and have the authority and appetite to make it happen.
The rest will continue announcing strategies and launching pilots that quietly disappear.