I was knocked out with the flu last week, which means this newsletter is us catching up on the week's news together in something close to real time. (I got my shot. The flu just didn't seem to know.)

One piece that caught my attention was KPMG Canada opening four AI Labs to help Canadian companies finally turn their AI pilots into something real. The framing in the announcement was that most Canadian AI investments die in the gap between running a pilot and putting something into production.

I thought of the Snowflake article I wrote on Monday. Snowflake had announced new tools that, in plain terms, give companies a way to set rules around what their AI is allowed to do and which data it's allowed to touch.

Then on Wednesday, Info-Tech warned that low-code platforms like Microsoft Power Apps, which let employees build their own simple apps without writing code, are being adopted faster than companies can put rules around them.

There’s a pattern here, even if the products are different. Companies are adopting new tools quickly. They're now figuring out that they need rules for them, and a growing list of vendors are showing up with help.

New tools, like AI and no-code/low-code platforms, are letting non-technical people build and act faster than companies can put rules in place to manage them.

I had lunch with Andrew Forde, KPMG's new national head of AI research, a few weeks ago and we got onto the question of friction. His view was that boards are starting to treat it as a strategic question rather than a technical one. Where do we need to move slow in order to move fast.

KPMG's Labs focus on agentic AI specifically. Where generative AI writes a paragraph for you, agentic AI takes a series of actions on its own. An AI handling a customer's billing dispute from start to finish without anyone in the loop. It’s not hard to see how risk increases with autonomy.

The Labs are intended to let clients build and test agents in a controlled space for a few weeks before anything goes near production.

Snowflake is looking at the controls a layer beneath. Before an agent can be allowed to act, somebody has to define what data it's allowed to touch and which agent is allowed to touch what. The controls only mean something when the company knows what's in its data in the first place.

"If you don't have it documented to train a human, then how do you train an AI agent," says Anahita Tafvizi, Snowflake's chief data and analytics officer, whom I spoke with in January.

Info-Tech's warning signals the same problem, only without AI. Yes, that still happens. When non-technical employees build their own apps with low-code tools, without oversight, the company ends up with weak security and tools nobody owns.

Friction usually gets a bad rap, but some of it is what stops a company from shipping something it shouldn't, or letting an AI agent take an action it shouldn't. The vendors making news this week are all, in different ways, selling that friction back.

Some light reading for when your AI strategy is technically a pilot for the third quarter running.

The workforce data your company is making AI decisions without — My colleague Jennifer Kervin profiles Toast Institute, a new not-for-profit arm of Toast trying to measure what mentorship and sponsorship programs really do for women's careers in tech. Half of the next Champions cohort will get structured sponsor training. The other half will get a manual. CEO April Hicke wants this reframed as talent strategy rather than DEI. Findings will run in Harvard Business Review over three years.

Telus and L-Spark hand five Canadian AI startups the keys to a supercomputer — A six-month accelerator backed by the Telus AI Factory, with hands-on advisory support and full IP ownership for the startups. The cohort covers retail, healthcare, robotics, enterprise software, and industrial automation. One more piece of the Canadian sovereignty story.

IKEA Finland tests pre-booked accompanied shopping and audio store navigation for blind and low-vision customers. The NaviLens app reads coded markers around the store and gives directions in over 30 languages, which means the accessibility build doubles as a translation tool. Useful read if you're thinking about whether accessibility is a feature you bolt on or a design choice you make up front.

Saskatchewan puts $30,000 into Uniting the Prairies — 650 attendees, 80 investors, 160 founders, and Innovation Saskatchewan showing up as the conference concierge instead of a passive funder. Founders who came through UP have raised more than $40 million since 2022. Modest by national standards, real by regional ones.

A stat from Wealthsimple's TLDR newsletter a couple of weeks ago has been kicking around in my head. AI suppliers have spent the equivalent of 2.4% of global GDP on data centres in the last eight years. 

Whether you read that as a bubble or just a very large bet, the scale is staggering. 

That kind of money only pays off when the organizations using it have the structure, the systems, the documentation, and the rest of the boring work in place to make it work. 

Which brings us back to friction. 

Used strategically, it provides the time to ask questions and consider decisions before the temptation to move faster makes them expensive to undo.

Talk to you next Tuesday.

— David