Procurement
How to write a procurement-ready AI pilot statement of work
An artificial intelligence pilot statement of work in the Canadian public sector can fail on three things, and none of them is technical. The scope does not say what the pilot will decide. The security requirements name a sensitivity level without naming the evidence. And the intellectual property terms leave the buyer unable to test what it bought.
Each is fixable in a paragraph, if the paragraph is written before the requirement goes out.
Know which vehicle you are in before you write
Public Services and Procurement Canada (PSPC) maintains an artificial intelligence source list of pre-qualified suppliers. It was published 15 January 2019. The departmental page describing it, last modified 1 October 2025, states that 74 vendors initially qualified and that the number has since reached 145. The companion supplier list page, last modified 18 August 2026, carries substantially more entries than that. Count it on the day you cite it rather than repeating a figure from a page that has not moved in a year.
Two structures sit on top of each other, and they are easy to confuse.
Three capability categories: insights and predictive modelling; machine interactions; cognitive automation. Note that these are the actual category names. “Machine learning” and “computer vision” are not categories on this list, and a requirement written against those words will not map.
Three dollar bands: up to $1,000,000, up to $4,000,000, and up to $37,500,000, all before taxes.
The standing invitation to qualify that fed the list closed 15 January 2025 and is marked expired. Public Services and Procurement Canada states that a replacement invitation to qualify is open. Confirm the current solicitation number directly with the departmental artificial intelligence procurement team rather than from a secondary listing.
Know whether you are competing it
The Treasury Board of Canada Secretariat contracting policy notice on trade agreement thresholds, issued 23 December 2025, sets the values in force for 1 January 2026 to 31 December 2027. One set covers both years.
The values below apply to federal departments and agencies buying services.
Thresholds for services
The Canadian Free Trade Agreement threshold is $139,000. The threshold under the Comprehensive Economic and Trade Agreement, the World Trade Organization Agreement on Government Procurement and the Comprehensive and Progressive Agreement for Trans Pacific Partnership is $239,200.
One caution that is easy to get wrong. These thresholds determine whether the trade agreements cover a procurement, and therefore whether open tendering and non discrimination obligations attach. Crossing a threshold does not by itself prohibit sole sourcing. The agreements contain limited tendering exceptions and the Government Contracts Regulations set separate bid solicitation exceptions. Say “covered” in your documents, not “must be competed”, unless you have checked the exception against the file.
A separate threshold is worth knowing early. The Secretariat’s guidance for drafters of Treasury Board submissions, last modified 1 April 2026, requires an appendix to the submission for proposals involving more than 10 new full-time equivalents.
Write the scope around a decision, not a capability
A pilot statement of work should name the decision the pilot exists to inform, the evidence that would support each possible answer, and the date the decision gets made. Three sentences. Without them, the deliverable becomes a demonstration and the evaluation becomes a matter of taste.
Then write the boundary. If the workflow is adjacent to determinations about people, say explicitly what the system will and will not do. Verifying that an application is complete is assisting. Determining that an applicant does not qualify is deciding. Those two sentences belong in the scope section, not in an annex.
Attach the compliance clock to the schedule
The Directive on Automated Decision Making requires an Algorithmic Impact Assessment at the beginning of the design phase and again before the system goes into production. The second one confirms the results describe the system as built. The institution publishes the results in both official languages on the Open Government Portal. Since the amendments effective 25 April 2023, it must also make peer review findings public, in full or as a summary, before a system launches.
Put those as dated milestones in the statement of work schedule with a named owner. An assessment that has to be produced but is not scheduled can become the reason a pilot misses its go live.
The administrative decision test decides the scope of the Directive, not whether a system is internal. Since 2023 the Directive expressly covers internal services including hiring, performance management and employee security screening. Partial automation counts too: a system that provides an officer with an assessment influencing a decision is in scope.
Name the security evidence, not the sensitivity label
“Must meet Protected B requirements” names a category, and a category is not a testable requirement. The evidence is what you procure.
Name the assessment path. The Canadian Centre for Cyber Security Cloud Security Assessment Program, page dated 19 June 2026, runs three baselines. Cloud Medium and Cloud High both cover information up to Protected B and differ on the level of injury from a compromise of availability and integrity. Decide which one applies and put it in the requirement.
Name the attestations. The Cyber Centre compares existing industry attestations rather than assessing every provider itself, and its assessment process document names four. The first is system security plans produced for the Federal Risk and Authorization Management Program. The second is American Institute of Certified Public Accountants (AICPA) System and Organization Controls (SOC) 2 Type II reports. The third and fourth are International Organization for Standardization and International Electrotechnical Commission (ISO/IEC) 27001 and 27017 reports. List those four.
Name the guardrails. The Government of Canada Cloud Guardrails have been mandatory since May 2022 under a standard attached to the Directive on Service and Digital. Implementation, validation and compliance reporting are required within the first 30 business days of a department getting access to a cloud account.
And check the hosting model early. Under the Directive on Service and Digital, some proposals require submission to the Government of Canada Enterprise Architecture Review Board. They are proposals for services categorized at Protected B or below that use a deployment model other than public cloud for application hosting, deployment or development. A pilot that quietly assumes on premises hosting has acquired a governance step that nobody scheduled.
Write the artificial intelligence specific clauses, because the standard set does not have them
The Standard Acquisition Clauses and Conditions manual carries its own notice that it “is no longer being maintained as of June 2023.” A targeted search of the archived clause set in September 2026 surfaced no clause addressing artificial intelligence, machine learning or algorithmic systems. Existing contracts using those clauses remain valid; new artificial intelligence requirements need language drafted for the purpose.
The clearest current Canadian government direction on what that language should contain comes from the Cyber Centre primer on the top ten artificial intelligence security actions. It was published May 2026 and last modified 29 May 2026. Its own words: “Require transparency, audit rights, liability and use restrictions”.
It adds: “Contracts with AI vendors should include explicit data usage, privacy, audit, and liability clauses”. And again: “prohibiting unauthorized use of organizational data for model training”.
Elsewhere in the same document, it calls for enforcing “no train” defaults and maintaining an artificial intelligence bill of materials covering models, datasets and dependencies.
Six clauses, then, at minimum:
No training on buyer data, by default, enforced in configuration and not only asserted.
Audit and test rights over the delivered system, including the right to run the buyer’s own evaluation set against it.
Data location and flow, stated for the prompt path as well as the corpus, including any third party model the supplier calls.
Logging access, with the fields named: tools used, actions taken, approvals given, key inputs and outputs, and language of request where the service is public facing.
A bill of materials for models, datasets and dependencies, delivered and updated.
Model change notification, because a hosted model can change underneath a working system without a contract event.
Clause two matters more than it looks. The Directive’s own procurement related guidance points departments to legal services on “licensing issues, including with respect to trade secrets.” It also names “any constraints they may impose on a department’s ability to access and test proprietary systems.” That constraint is negotiable at requirement time and effectively immovable afterwards.
The test
Source
Dollar figures on this page are the sources’ own or recorded amounts, not our prices. Our prices are in Canadian dollars (CAD), plus GST/HST.
Before it goes out, hand the statement of work to someone who was not in the project and ask them three questions. What decision does this pilot inform? What evidence proves the security level? And on the last day of the contract, what can we test ourselves without the supplier’s help?
A statement of work that survives those three questions will survive the review board.
Source
Sources: Contracting Policy Notice 2025-8, trade agreement thresholds, for 1 January 2026 to 31 December 2027. PSPC Artificial Intelligence Source List, modified 1 October 2025. ITSM.50.100. Guidance for Drafters of Treasury Board Submissions, modified 1 April 2026. Read on 12 September 2026 and due for re-check by 12 March 2027. Regulatory writing on this site is analysis, not legal advice.
Previously: from demonstration to operating service. Next in this series: open-weight versus proprietary models, a Canadian cost model.