Hi {{first_name}},

I sat with an AI job from the moment it was assigned to the moment it came back finished, and what got my attention was the bill. It arrived itemized to the fraction of a cent, about 26 cents for the whole job, with a log showing where every piece of it went.

If you've ever tried to answer your CFO on what AI is costing the company, you know how rare that sentence is. Most of the AI running inside your organization can't tell you what a single task cost. It can only tell you the meter is running.

Clark Lai's venture studio just launched Opal, a platform for running a company's everyday work on AI agents, priced per seat instead of by the token. A technical team connects the systems and sets the rules once, then anyone else describes the work they need in plain language and the agents carry it out. He walked me through it at Motiv's Calgary office and didn’t mince words about why he thought the industry bills the other way.

"It's not in Anthropic's interest for you to use fewer tokens," he told me. "That's where they make money. It's like a slot machine."

Every file your team uploads and every answer that comes back runs the meter, so the tool you bought to save time makes more money the longer it takes. 

Opal, on the other hand, runs from a free tier through $39 and $99 monthly plans, and converts every frontier model's pricing into a single currency where one credit equals a cent. Research from BetterUp Labs and Stanford's Social Media Lab found 40% of workers received this kind of work in a single month, each incident costing close to two hours to sort out. Lai says his company takes no margin on what customers spend, which is why it has every reason to help them spend less, and he puts it at roughly seven times cheaper than a team doing the same work through separate chat subscriptions. "We're experimenting" is a fine answer for the board right up until someone asks what the experiment costs.

What matters more than the price is the paper trail. Every agent, prompt, guardrail and knowledge file carries version history, and every job leaves a log of what each agent did, which tools it called, and what it pulled. For anyone who has wanted to know what an AI was doing inside their company, that log is a key part of the product. Lai says it's hosted in a Canadian data centre, with zero-data-retention agreements so customer data can't be used for training. And it’s built so non-technical people can still get a lot out of it.

"Humans are good for building relationships, making decisions, having accountability, creativity, the things that agents fall short on," says Lai. "Our approach is not let's replace everybody with agents. It's take the best of both worlds and integrate them so they can truly collaborate and get things done."

→ Read the full story on Opal here (it's a 16-minute read for those who dabble in long reads)

Explaining how an AI reached a decision is about to become a regulatory requirement for much of Canada's financial sector, and there are about nine months left to get ready for it.

My colleague Jennifer Friesen reported on OSFI's updated Guideline E-23, or in plainspeak: the rulebook for how federally regulated banks and insurers manage the risk of the models they run. The update now covers AI, and it takes effect May 1, 2027 (Quebec's AMF finalized a parallel guideline for the same date).

An institution will have to know every model it runs and where it came from, including anything bought from a vendor. Every model gets a risk rating based on what it does, how independently it runs, and how much damage a bad output could do. The ones that could affect customer decisions, financial results or regulatory compliance get the full inventory-and-audit process, and the higher the rating, the more the institution has to prove it understands how the model works and can monitor it. Fail the audit and OSFI can require more capital or a detailed fix-it plan on its timeline.

The hard part of this whole thing is technical. 

A modern AI model learns from data instead of following rules a person wrote, so even the team that built it can't always say why it produced one answer over another. A loan officer has always had to tell you why you got turned down. The AI that does that job gets until 2027 to learn how.

Joseph Geraci, founder of the publicly listed Canadian clinical data company NetraMark, has spent nearly a decade building AI in pharma, where regulators have demanded proof of how a model reached its conclusion for years. Speaking remotely to the CIO Association of Canada's Peer Forum earlier this year, he was making a broader point about what autonomous systems need, and he named the thing E-23 now asks for. "We need to move away from logging what an agent did, to auditing exactly why it made certain decisions," he said.

OSFI only binds the institutions it supervises. Everyone else running agents gets to watch the audit happen to someone else first. The difference between a log of what your AI did and an audit of why it decided is the difference between a bad email and a decision you have to answer for.

→ Read: Your AI made a decision, and Canadian regulators want to know how

Canada does not track burnout cleanly in the corporate technology sector. No public study isolates it, so the closest read comes from broader data, and what that data shows is not good. The TELUS Health Mental Health Index scores 3,000 working Canadians out of 100, and in early 2026 it put the technology industry at 59, the third-lowest of any industry it measured, down 4.5 points in six months and one of the steepest declines in the survey.

Marissa McNeelands lived the version of that the numbers describe. She spent a decade in tech, burned out, and found what helped in a hot sauna and a cold plunge. Now she's opening Edmonton's first social sauna in January, built for the people she used to work alongside.

"Circa isn't created for the wellness people," says McNeelands. "It's created for the high achievers, the high performers, the perfectionists, the people that are running themselves into the ground and only have 75 minutes once a week to devote to themselves."

The money has noticed. Othership in Toronto has raised over $8 million for its U.S. expansion, and Bathhouse in New York raised $35 million in a single round led by Imaginary Ventures, the firm behind Skims and Glossier. (Yes, venture capital is chasing saunas now. I was surprised too.) McNeelands is raising a friends-and-family round with $650,000 committed against an original $700,000 target, and she's now expanding to $850,000.

→ Read the full profile of McNeelands and Circa here

Your engineering team is now faster than the company around it, and that is where the value leaks out. 

Anju Visen-Singh, founder and principal of the advisory practice Throughline, wrote her first piece for us this week, and her argument is that AI has collapsed the build cycle while everything downstream of it moves at the old pace. She points to Anthropic building its Cowork agent in about 10 days, most of the code written by its own AI tool, then shipping a major release roughly every two weeks.

When the product ships in 10 days and the rest of the company takes 10 weeks to position and sell it, she writes, you have "accelerated your dysfunction instead of your growth." 

She walks it through: a capability ships, sales is still working from a deck that predates it, and the buyer who researched the product that morning knows more about the new feature than the rep across the table. The deal stops on a question nobody can answer, and what surfaced in that conversation never makes it back to the product team.

→ Read Anju Visen-Singh's first piece for Digital Journal here

Some light reading for when the agents have the week under control.

Getting to yes starts with knowing who is across the table.
Andrew Robinson has heard the pitch enough times to know how it goes: a company walks into a Nisga'a Nation meeting, points to the work on Ksi Lisims LNG, and explains why it's the right partner. His first question back is rarely the one they expect: how many of you know who I am? Robinson, CEO of the Nisga'a Lisims Government, was on a Global Energy Show panel called Getting to Yes, and the premise built into that title is that a Nation saying no is the obstacle. The leaders on the panel took it apart. Sarah Coleman's piece, which first appeared on Rooted, a CSV Midstream publication, lays out what real partnership requires from the companies building Canada's next energy projects.

OpenAI's own models went rogue during a safety test.
Measuring how good they were at hacking, the company ran them in a controlled environment with internet access limited, and they spent substantial computing power getting out, then hit an outside code platform with stolen credentials, with nobody directing any of it. Leave a capable AI home alone and it does what any teenager does, except this one climbed out the window, threw the party at the neighbour's, and went through its parents' drawers on the way out.

Nearly every developer using AI says it makes them faster, and far fewer have rules for what it writes.
In a new Info-Tech study my colleague Jennifer Kervin covered, 578 applications, engineering and product leaders using AI split like this: 94% reported productivity gains, 67% said its code needs more testing than a human's, and 37.4% rated their build-stage AI maturity as formal or better. The applications people closest to the code are the least convinced by any of it, because they've met confident and wrong before and it used to be an intern.

AI slop compounds into something worse than a bad document. Someone writes a report with AI assuming the recipient will only skim it with AI, and the recipient does exactly that. A Harvard Business Review essay by Oxford's Matthias Holweg and Babson's Thomas Davenport calls the result knowledge decay, where unverified output moves through a process until the information the company runs on stops being reliable. Research from BetterUp Labs and Stanford found 40% of workers received this kind of work in a single month, each incident costing close to two hours to sort out. Verifying the output can cost more time than the AI saved.

Canada ranks third in the world for AI talent, ahead of the U.K., Israel, Singapore and Germany, on a fraction of what the U.S. spends, according to an analysis from the firm Teamed. Dr. Tim Sandle reports that Canada drew about $3.1 billion (all figures USD) in AI startup investment last year against more than $180 billion in the U.S., and still scores 8.89 out of 10 on talent. The people are here, whatever the funding says.

Most of this issue circled the same question from a different side. Opal wants to show you the bill, the regulator wants to see the audit, and the people closest to the code won't vouch for it.

Marissa McNeelands has a master's in management and AI, and she's spending this year on a room full of hot rocks and cold water. Everyone else in this edition of our newsletter is trying to get control of the technology. She skipped that fight and went straight to building the thing that fixes you after it. Ground breaks in southwest Edmonton next month if the city signs off, and the scariest thing on her risk register is a building inspector.

— Chris Hogg

If you have any questions, you can reply to this email. I read everything.

Do you have a story we should chase, or a tip you'd share with a peer?
Contact us here.

Know someone who should be reading this. Forward it on or
they can sign up here.

This issue is presented by Calgary Economic Development.
If you want your brand in front of the people running technology across Canada, you can sponsor a future issue.