We Pointed AI at $534 Billion in Federal Grants. Here’s What It Found.

1.275 million records. Seven analytical phases. And some findings every Canadian should see.

Earlier this week I published a piece about Agency 2026, the national AI hackathon my Ministry is hosting in Ottawa on April 29. I wrote about what agentic AI can do when you point it at public data. I talked about the challenges our teams will tackle. I talked about the guardrails.

Now let me show you what happens when you actually do it.

In preparation for the hackathon, my team used agentic AI to conduct a full independent review of Canada’s federal grants and contributions dataset. Every dollar the federal government has recorded giving away through grants and contributions, across every department, for every fiscal year on record from 2006 to 2025.

That is 1,275,521 records. $533.65 billion in original grant spending. All of it published on the Government of Canada’s own Open Data portal, available to anyone.

We downloaded every record, cleaned the data, loaded it into a database, scored it against a risk framework, and cross-referenced the highest-risk findings against public internet sources.

The data download alone took 12 minutes. The full pipeline, from raw data to validated findings, was done in a few hours. Work that would take a team of analysts weeks or months, completed before the end of the day.

Here is what we found.

The AI Built Its Own Analytical Pipeline

Before I get to the findings, it is worth understanding what “agentic AI” actually means in practice. Most people picture a chatbot answering questions. What we used is fundamentally different. An AI agent is given a goal, access to tools, and the ability to work autonomously.

The agent downloaded all 1.275 million records from the federal government’s Open Data API in 128 sequential batches. It designed and built a purpose-built database with five reference tables, twelve performance indexes, and three analytical views. It identified and fixed data quality issues: 19,204 records with inconsistent agreement type codes, 100 records with free-text province codes like “Rome” and “N/A” instead of proper codes, and 318,030 amendment records that had to be separated from original grants to avoid double-counting.

Then it ran eight advanced analytical scripts. Provincial equity analysis. For-profit deep dives. Amendment creep detection. Recipient concentration scoring. Zombie and ghost-capacity entity identification (if those terms are new to you, I explained them in my earlier article on the hackathon). A comprehensive risk register scoring 109,795 non-government entities across seven dimensions.

Every script is open source. Every query is reproducible. The read-only database credentials are available for anyone who wants to verify the work independently. We will publish the full repository, scripts, and database access alongside the Agency 2026 hackathon on April 29.

A team of human analysts could have done this. It would have taken weeks, maybe months. Our AI did it in hours.

The Risk Model Caught Real Scandals

We built a risk scoring framework that evaluates every grant recipient across seven dimensions: cessation of activity, identity gaps, amendment patterns, funding concentration, dependency on a single source, opacity (missing descriptions or expected results), and outsized scale. Each dimension scores 0 to 5, for a maximum risk score of 35.

The model flagged 846 entities as CRITICAL risk and 6,931 as HIGH risk.

Then we did something important. We took the top 25 highest-scored entities and searched public sources to see if the risk model was right.

Three of them were confirmed failures that Canadians already know about.

WE Charity Foundation: $543 million. Wound down after a political scandal in 2020. The federal government awarded WE a $543 million contract to administer the Canada Student Service Grant before the conflict-of-interest revelations forced a reversal. The charity subsequently shut down its Canadian operations.

Sustainable Development Technology Canada (SDTC): $134 million. Abolished by the federal government in 2024 after the Auditor General found 90 conflict-of-interest violations in its grant approvals. The RCMP is investigating.

Canada World Youth: $37 million. Shut down in 2022 after decades of federal funding. Combined: $717 million in confirmed waste to three entities that our AI model flagged before any internet research was performed.

The model identified them purely from patterns in the data: cessation of activity, unusual funding concentration, identity signals.

That is the power of this approach. The data tells you where to look. The AI processes more of it, faster, than any team of auditors could manage.

Not Every Flag Is a Scandal

This is where intellectual honesty matters.

The risk model also flagged companies like FCA Canada, Toyota, and Husky Oil as potential “zombies” because their last federal grants were several years ago. Internet verification quickly confirmed these are not zombies. FCA became Stellantis. Husky merged with Cenovus. MDA Systems rebranded as MDA Space. These are corporate name changes, not disappearing acts.

Similarly, organizations like STARS Air Ambulance and Deep Earth Energy scored as high risk because pre-2018 data quality gaps (missing business numbers, no industry codes) inflated their scores. Both are clearly still operating.

This matters because it demonstrates exactly how these tools should be used.

The AI is not an auditor. It is a triage system.

It processes volumes of data no human team can handle and tells you which 846 entities out of 109,795 deserve a closer look. Then humans verify. Some flags are real. Some are data artifacts. The value is in the sorting, not in treating every flag as a verdict.

Our model’s hit rate on the top 25 was strong. Three confirmed scandals, five genuine investigation targets, seven explained by corporate restructuring, and roughly ten cleared as operational entities with data quality gaps. That is a useful filter.

$71.5 Billion With No Paper Trail

Beyond the entity-level findings, the analysis revealed a systemic transparency problem in Canada’s federal grant data.

$71.5 billion in grants over $1 million each have no description or expected results documented in the public record. The money went somewhere. The dataset tells you who received it and how much. It does not tell you what it was for or what the government expected in return.

That is not a finding about fraud. Most of that money probably went to legitimate purposes. The problem is that there is no way to verify that from the data. The government records what it spent, not what it bought.

Four percent of for-profit grant recipients have no business number on file. These are companies receiving public money without the most basic identifier that would let anyone cross-reference them against a corporate registry.

When we looked at amendment patterns, we found 19,303 grants across the federal government where amendments more than doubled the original value. Some of these have legitimate structural explanations. Others warrant a closer look.

None of these are accusations. They are questions that the data raises and that the data alone cannot answer. That is the point. The data should at minimum be structured to allow Canadians to verify how their money is being spent. Right now, for billions of dollars, it is not.

Three Entities Worth a Closer Look

Beyond the three confirmed failures, our analysis flagged several entities that deserve further investigation. These are not confirmed scandals. They are patterns in the data that raise legitimate questions.

Halagonia Tidal Energy (Nova Scotia, $30 million). A subsidiary of Irish-based DP Energy that received a $29.8 million federal grant in 2018 for a Bay of Fundy tidal energy project. DP Energy subsequently abandoned the project. It joins a long list of failed tidal ventures at the same site, including companies that went bankrupt leaving abandoned turbines on the ocean floor.

TMT International Observatory ($214 million). Canada committed $243.5 million in 2015 toward the Thirty Meter Telescope in Hawaii. Construction was halted by protests from Native Hawaiians who consider the summit of Mauna Kea sacred. Over a decade later, the telescope has not been built, costs have risen from $1.5 billion to $3 billion, and the project faces a $1 billion funding shortfall.

Carisbrooke Shipping ($13 million). A UK-based company headquartered on the Isle of Wight that received a federal Strategic Innovation Fund grant despite having no visible Canadian operations. The company operates cargo vessels between the UK and Canada but is registered and managed entirely out of England.

Each of these deserves a straightforward answer from the federal government. What was the money for? What was delivered? If the entity ceased operations, did taxpayers get what they paid for?

What Happens Next

This analysis was a proof of concept. A single working session using open data and open-source tools. Everything we built is reproducible. The scripts, the database, the risk framework, the methodology. It is all available for independent review.

On April 29 at Agency 2026, teams from across Canada will tackle challenges exactly like this. Zombie recipients. Ghost capacity. Funding loops. Sole-source amendment creep. They will use the same agentic AI methods on the same public data, and they will do it in a single day.

But there is more to this story than the methodology.

When we ran the provincial equity analysis, something jumped out of the data that every Albertan, and frankly every Canadian, should see. The way federal grant money is distributed across provinces is not what most people would expect. The numbers raise serious questions about whether Canada’s federal funding system is working the way it should.

But that is a conversation that deserves its own article. The numbers are significant, and I intend to lay them all out.

Stay tuned for Part 2.

---

This is Part 1 of a three-part series on what AI revealed inside Canada’s federal grants data. Part 2 examines how federal funding is distributed across provinces and what that means for Alberta. Part 3 looks at how Canada allocates research funding between social sciences and engineering, and what the OECD data says about the consequences.

The full analytical methodology, open-source scripts, and reproducible database are available for independent verification. Details will be published alongside the Agency 2026 hackathon.

---

If you found this valuable, share it with someone who cares about how government spends public money.

Subscribe to get Parts 2 and 3 when they publish.

Nate Glubish is the MLA for Strathcona-Sherwood Park and Alberta’s Minister of Technology and Innovation.