The payoff gap.

I missed a few issues of this newsletter. Some of that was a heavy summer of client work. Some of it was taking a bit of a break to get married earlier this month. But, frankly, most of it was feeling boxed in. I’d started to carve out a lane writing about how professionals actually work with AI, and every week that lane felt smaller against what was actually moving. What moved this summer was policy.

In June, after a federal task force digested eleven thousand submissions, the Prime Minister launched AI for All: two billion dollars, a target of a quarter-million jobs, and four pillars of ambition. What it didn’t come with was much policy. The numbers are aspirational, the legislation is unwritten, and the follow-through gets decided starting in three weeks, when the House comes back with an online harms bill in the queue and consultations closing. I kept writing nothing, because the thing I wanted to say didn’t fit the newsletter I’d been writing.

So the newsletter changes a bit today. I’ll still write about practice, but the frame is different now: if Canada gets AI policy wrong, how any of us work with agents is a rounding error. The coming gap runs between countries, not between people who prompt well and people who don’t, and wealth follows that line. So do skills.

To start I went looking for the adoption gap, and it wasn’t where I expected. Statistics Canada’s latest business survey has AI use at about one in five Canadian firms, triple what it was two years ago and statistically tied with our American counterparts. We did the easy part. The payoff is what’s missing: control for firm differences and the measured productivity advantage of adopters shrinks to nearly nothing, and no G7 country can find the gains in its macro numbers yet. Adoption was the on-ramp. The road is agents doing real work inside firms, and whether that lands here is a policy choice being made in the next twelve months.

The models are evenly distributed now. The advantage won’t be.

My six calls:

1. Judge the follow-through on diffusion, nothing else. Canada wrote the world’s first national AI strategy in 2017. It bought us research leadership, and one of the weakest commercialization records in the G7. AI for All is the third pass at this file, and its headline numbers are frankly aspirational, two hundred billion in growth and two hundred and fifty thousand jobs by 2031, with no published account of how anyone will know we’re on track. Aspiration is what strategies are for, so I won’t begrudge it. The test is the implementation: does it move AI from pilots into production inside ordinary firms, and will the government publish adoption and deployment numbers quarterly, beside the other economic prints, so anyone can check whether the ambition is landing.
2. Leave AIDA buried, and modernize privacy law instead. The last Parliament’s horizontal AI act died with prorogation. That was a mercy. Regulating a general-purpose technology through one omnibus statute freezes small firms in legal uncertainty while doing little about actual harms, which live in specific uses: fraud, discrimination, unsafe products, deceptive marketing. We have laws and regulators for each of those already. Resource them and give them AI-specific guidance. Then spend the legislative energy where uncertainty actually chills deployment. Our federal privacy law is a quarter-century old, and every serious AI project I’ve seen slows down at the same question: can we even use this data? Answering that in statute would do more for responsible adoption than any AI act.
3. Ottawa should buy before it regulates. The strongest signal government could send is a purchase order. Pick high-volume, low-drama workflows, benefits backlogs, procurement paperwork, translation, records management, and deploy agents into them with published results, including the failures. Design for recourse: actions stay drafts until a person promotes them, and the logs are good enough to reverse anything. The public service’s caution is legitimate, and an undo button answers it better than abstinence does. A government that runs agents credibly can regulate them credibly. Not before.
4. Treat the information layer as national infrastructure. This is where my old lane and the new one meet. Two years from now every country has the same models. They will not have the same quality of information for those models to work with: machine-readable government data, standardized records, registries that publish structured data instead of PDFs, firms that have written down what they actually know. That substrate is what agents run on, and it compounds. Compute sovereignty matters for a handful of frontier firms. Data readiness decides the rest of the economy. It’s unglamorous and it photographs badly, and I’d trade a billion dollars of it for the next moonshot announcement without blinking.
5. Energy policy is AI policy. Sovereign compute is a grid question. If we want training and inference to happen in Canada, the binding constraint is megawatts and the years it takes to approve them. British Columbia and Quebec sit on some of the cleanest power advantages on the continent. A country serious about AI approves generation and transmission on timelines measured in months, and says so out loud.
6. Solve the apprenticeship problem before it solves itself badly. Agents eat junior work first, the research and the first drafts, and junior work is where every senior professional came from. No firm has an answer. Mine included. The market alone won’t produce one, because each firm optimizing its own headcount collectively starves the pipeline. Rebuild co-op and articling models around supervising and verifying AI work, and put it on the skills agenda now, because in ten years this decides whether firms here can still grow their own senior people.

I don’t hold these six with equal confidence, and I’m more certain of the direction than the details. But the window is real. The strategy launched in June, the legislation behind it is still unwritten, and several consultations close within weeks. The people drafting the follow-through are reading the room right now. It’s worth being in it.

The number: 19.2 percent. Canadian businesses using AI, per Statistics Canada this summer, up from 6.1 two years ago and tied with the United States. Adoption caught up. The payoff hasn’t, and policy decides whether it does.

Steal this: one question that sorts any AI policy proposal in about ten seconds. Does this make it easier or harder for a twenty-person firm to put this technology into production? Most proposals fail it, and most of the people drafting them have never had to answer it.

Cool thing I saw: an investor named Matthew Smith spent eighteen months modelling nearly every gas well, pipeline, and processing asset in the United States, and concluded that AI data centres and LNG exports will need more gas than the country can produce and deliver starting around 2028. His numbers, not mine, and plenty of people push back on them. But sit that forecast next to two provinces holding some of the cleanest surplus hydro on the continent and the fifth call above stops sounding abstract. The episode is “Natural Gas: The Next Bottleneck” on Invest Like the Best, worth the drive home.

Next issue: when an AI agent lobbies, who registers? Nobody’s lobbying law contemplated software making the calls, and it’s a better question than it sounds.

I advise organizations with interests on some of these files. The views here are mine.