This week’s issue is being finalized from Pearson Airport while I wait to (re)board my delayed flight to Calgary for Stampede. Stampede, for anyone who's never been, is when Calgary’s business community crams a year's worth of networking into ten days, with a little cowboy cosplay on the side.

Hmm, never a good sign when the airline proactively gives you a food voucher.

Anyway, last week I wrote about AI governance at Thomson Reuters, and how it's helping the company move faster. 

Then I read about Yoshua Bengio's new watchdog AI, and DigiCert's push to cryptographically verify cloud infrastructure. 

Three different stories, and to me at least, they all feel focused on the practical.

Is AI getting boring? Maybe. And that might be the best sign yet that it's ready to deliver.

Caitlin Halferty, head of data and analytics at Thomson Reuters, told me she's seen the same pattern too often elsewhere.

"They'll build a really great data capability, perhaps a pilot, and then they'll go to bolt on the governance later," says Halferty.

At Thomson Reuters, governance work started with finance and expanded from there, pulling customer data out of 23 different sources into one place before any AI model touched it. The company now runs roughly 37,000 governed tables and 350 data sources into one semantic layer, used daily by more than 1,500 people, so two reports never disagree about something as basic as how many customers the company has.

"Because we focused on it early, we're able to move faster," says Halferty. Read more about Thomson Reuters in the article. 

Halferty's focus is on the data layer, before a model ever sees the numbers. The other companies I read about this week are trying to build that same kind of trust further downstream, closer to where the model runs.

Bengio's LawZero paper proposes training a separate AI to watch other AI systems, with no stake in the outcome and nothing to gain from misleading anyone. 

DigiCert and Google Cloud are proposing cryptographic proof that the servers running an AI haven't been tampered with, the same idea behind the padlock icon on a secure website, applied to a data centre instead of a browser.

Guardrails and audit trails matter because nobody, including the people who built the model, can point to the specific reason it made one choice over another. 

Anthropic's CEO, Dario Amodei, noted in an essay last year that when a generative AI summarizes a financial document, "we have no idea why it makes the choices it does," and it can pick one word over another, or slip into an occasional mistake, without anyone able to say why. 

That's the company building some of the most advanced frontier models in the world. It's also exactly why the accountants, the lawyers, and the operations people are the ones in the room now, asking the questions the sales pitch never had to answer.

Some light reading for when you've had enough of AI for one email.

Alberta facility produces first certified carbon removal credits in North America My colleague Jennifer Friesen has the story on Deep Sky's Innisfail facility, with Microsoft and RBC locked in as buyers through 2034. RBC walked away from its 2030 emissions targets earlier this year and needs a climate claim that can survive a regulator or a lawsuit.

AI projects are clearing launch and missing ROI In another article, Friesen digs into new research from Winnipeg's Laivly. Success, in this survey, is apparently a vibe.

A quarter of Canadian leaders don't see AI coming for them Gartner expects a third of all enterprise software to carry agentic AI features by 2028. Those two numbers are set to collide, whether or not anyone updates the strategy deck first.

Fear and anger brew inside Meta amid AI frenzy More than 1,600 staff signed a petition calling it a "data extraction factory," while the company spends up to $145 billion on AI this year. The unrest is the part the spending never shows up in.

Speaking of riding bikes for a good cause. In a few weeks, I’ll be riding the Princess Margaret Cancer Foundation's Ride North. Our family has raised more than $53,000 over the last four years and is looking to add to that number this year.  My 13-year-old daughter is kicking my ass (again) with another $12,000 so far this year. I'm proud and a little embarrassed she's so much better at this. I’ll just leave this here until all the donations roll in.

Maybe it's time for a little boring plumbing talk when it comes to AI. 

Governed tables, formal safety papers, cryptographic certificates, most of it won’t trend on LinkedIn. I'll take boring over broken, or dangerous.

If your AI vendor is still leading with the hype instead of the audit, ask why.

— David