Analysis
What the federal AI register shows
Treasury Board published the Government of Canada Artificial Intelligence (AI) Register on 28 November 2025. It is a public dataset of real federal AI systems, their owners and their status. It reads as data, not only as a list. We read it that way. Four things in it are worth an executive knowing.
The dataset
A public inventory that fulfils a strategy commitment
The register was published on 28 November 2025 and last updated on 28 April 2026. Its own notes say publication fulfils a commitment made in the AI Strategy for the Federal Public Service 2025-2027.
The file is a comma-separated dataset on the Open Government Portal with 24 data columns. For each entry it records the system name in both official languages, the institution and a description. It also records who the primary users are and who developed the system. Vendor information and a status with its date follow. Further fields cover the system capabilities and the data sources, whether personal information features in the system, and the applicable personal information banks. The last two say whether the system tells people they are interacting with AI, and what results the system produces.
Read the scope statement carefully, because it governs everything below. The register covers systems that are or have been in use. Some institutions also chose to submit pilots and proof-of-concept work, and some did not. The register is therefore not a census, and the numbers in this article are a description of what institutions reported rather than a measurement of federal AI.
Source
Source: Government of Canada AI Register (minimum viable product), Treasury Board of Canada Secretariat, published 28 November 2025, last modified 28 April 2026. Counts in this article were read from the register's data service on 5 September 2026 and will move as institutions submit entries.
What the data says
Four things the register tells you
Each figure below was read from the register itself rather than from the launch announcement, which reported rounder numbers five months earlier.
How many AI systems has the federal government registered?
412 entries, from 43 distinct federal institutions. The launch announcement in November 2025 described over 400 systems from 42 institutions. The register therefore grew by a small number of entries and one institution across its first five months.
Are these systems used by the public or by public servants?
356 of the 412 entries record their primary users as Government of Canada employees. Another 25 serve both employees and the public, and 23 serve members of the public. Six record neither. That field is populated for 410 of 412 entries, which makes it the most complete substantive field in the register.
So 86 percent of registered federal AI is an internal tool, and 48 entries touch a member of the public at all. An executive assuming the federal AI conversation is mostly about citizen-facing decisions is reading the wrong dataset.
How many are in production?
| Status | Entries | Share |
|---|---|---|
| In development | 182 | 44% |
| In production | 160 | 39% |
| No status recorded | 42 | 10% |
| Retired | 28 | 7% |
More registered federal AI systems are in development than in production. Twenty-eight have already been retired, which is the one figure in the register that describes an ending rather than an intention.
How many involve personal information, and how many tell people?
88 entries record personal information in the system and 224 record none, leaving 100 with the field blank. On notification, 196 entries say the system tells people they are interacting with AI and 79 say it does not, with 137 blank. Read both as coverage before reading them as findings. The first field has a value for 76 percent of entries and the second for 67 percent.
The structural finding
The register carries no impact level
This is the thing a Chief Data Officer should take away from the file, and it is a fact about the fields rather than about the numbers.
The register has no column for an Algorithmic Impact Assessment (AIA) impact level, no risk tier, no assessment score, and no link to a published assessment. Its own notes say it was assembled from existing sources of information including assessments, so the assessments fed the register without their classification travelling with them.
That makes the register a transparency artifact rather than a risk-governance artifact. It answers what exists and who runs it. It does not answer which systems make administrative decisions about people, which sit at Impact Level III or IV, or which required peer review before launch. A briefing that cites register counts as a measure of federal AI risk is citing the wrong instrument. The same is true of a supplier that quotes the total as a market size.
Where the classification question is the one that matters, the instrument is the Directive on Automated Decision-Making and the assessment published against a given system. We set out how that screening works in when an AI assistant triggers an assessment, and how an assessment is completed and scored in completing an assessment.
Practical use
What the register is good for
- Finding comparable work before you scope your own. The description, capability and data-source fields are written by the institutions running the systems. If another department has already built something close to your candidate workflow, this is where you find it.
- Finding the people. Each entry names the institution and the developer. A conversation with a team that has already shipped is worth more than a market scan.
- Testing an assumption about internal versus public use. If a business case rests on citizen-facing automation, the register shows how unusual that still is.
- Seeing what has been retired. Twenty-eight entries record a system that is no longer in use, which is the closest thing the federal system publishes to a record of what did not work.
Warning
What it will not support
A compliance denominator, a risk profile, a spend figure or a vendor market share. The register records none of those. It also excludes AI embedded in low-risk commercial products, and its coverage of pilots depends on whether an institution chose to submit them.
Honesty
What this analysis does not establish
- These are reported figures, not measured ones. Every count describes what institutions submitted. The register's own notes say entries are incomplete and formatting varies, and that the team used machine translation in preparing the file.
- The counts move. They were read on 5 September 2026 against a file last modified on 28 April 2026. The dataset's stated update frequency is not planned, so a reader should re-derive any figure before putting it in a business case.
- We did not read the change log. The comparison between the November 2025 announcement and the April 2026 file is a comparison of two published snapshots. It is not an account of what the register's own change log says moved.
- Institution counts below the top eleven are not published here. We could read the totals for the largest institutions directly and could not read the full list reliably, so the rest is left out rather than estimated.
Next step
Find the comparable system before you scope yours
Bring one candidate workflow to a 30-minute fit check. Part of that session is checking whether the register already holds something close to it, and whether the Directive applies to what you are proposing.