In this edition of the newsletter, we dive into the AI-driven technical evolution of KOHO, a Canadian fintech company dedicated to helping over 2 million users build wealth through smarter spending and saving tools. KOHO’s VP of Technology, David Kormushoff, and Technical Product Manager, Katrina Stevenson, share how the company is methodically scaling AI developer productivity without compromising engineering culture. Their reflections capture a transition from chasing short-term speed to building a “balanced scorecard” where AI adoption, code quality, and developer experience are measured in tandem. Towards the end of our interview, we had just enough time for a rapid fire round!

Complementing KOHO’s story, we are introducing the AI Corner, where we highlight recent breakthroughs shaping the next wave of agentic AI along with what we, at Sagard, are reading; from the viral rise of the “OpenClaw” agent and its companion social network, Moltbook, to Anthropic’s “Claude Cowork” and Axiom’s landmark AI reasoning performance on the hardest math competition in the world, the Putnam Competition.

Measuring AI Developer ROI Without Breaking the Culture

KOHO has been methodically experimenting with AI across its engineering organization, with an emphasis on long-term impact rather than short-term hype. In this conversation, Katrina Stevenson, Technical Product Manager, and David “DK” Kormushoff, VP of Technology, share how KOHO thinks about developer productivity in an AI-enabled world. They discuss why traditional metrics still matter, what surprised them about adoption across experience levels, and how they addressed concerns around trust, quality, and security. The discussion offers a grounded perspective on treating AI not as a shortcut, but as a new way of working that requires patience, guardrails, and cultural alignment.

Here are our key takeaways from the interview:

1. Productivity Follows Enablement, Not the Other Way Around: AI productivity gains typically appear after teams adjust their workflows and build comfort with new tools. Early dips are often part of the learning curve, making experimentation and enablement essential before optimizing for output.
2. Measurement Requires Balance, Not New Metrics: Effective AI measurement combines existing productivity metrics with quality signals, developer experience, and usage data. This balanced approach helps teams understand whether AI adoption is driving meaningful, sustainable impact.
3. Adoption Is About Mindset, Not Seniority: AI adoption varies more by how individuals define their value than by experience level. Engineers focused on customer outcomes tend to adopt faster than those who equate value primarily with writing code.
4. Guardrails Enable Trust and Scale: Strong engineering fundamentals, tool standardization, and clear guardrails help address concerns around quality and security. These controls make it possible to scale AI adoption without compromising reliability or trust.

Let’s dive in.

To start us off, what are your roles at KOHO?

Katrina: I’m a technical product manager at KOHO, and my role focuses on building AI tooling and driving AI adoption across the company. That includes developer productivity, but also exploring how AI can support how teams work across different departments. A big part of my work is making sure we’re learning from the data while also listening closely to how people feel about these tools.

David: I’m the VP of Technology, reporting to our COO. I oversee what we think of as KOHO’s AI-native organizational transformation. Developer experience and productivity are important pieces, but the broader mission is about change management, helping teams across functions and seniority levels adapt to AI as a new operating model, not just a new tool.

Should throughput and speed be the primary metrics for AI-driven developer productivity?

David: Throughput and speed are useful, but they can be misleading if they’re treated as the main success criteria early on. AI changes how people work, and there’s almost always an adjustment period. If you optimize for productivity too quickly, you risk misreading what’s actually happening, because people may initially slow down as they learn new workflows. The more meaningful gains tend to show up after that adaptation phase.

Katrina: From a measurement standpoint, we landed on using familiar metrics rather than inventing new AI-specific ones. The reason is simple, you need a clear baseline to understand impact. We built a scorecard that combines impact metrics, quality signals, developer experience feedback, and AI usage data. That lets us see not just whether output changes, but how usage correlates with quality and satisfaction over time.

Did KOHO build its productivity dashboards in-house or rely on vendors?

Katrina: It’s a combination. We use Sigma as our dashboarding layer and feed it data from several sources, including our developer productivity tools and internal surveys. There were also a few AI-specific metrics we wanted that didn’t exist out of the box, so we set up custom data pipelines to capture those and bring them into the same view.

KOHO reported a 150% year-over-year increase in code output with fewer developers. When code abundance becomes the norm, how does it affect the roles of junior and senior engineers?

David: One important clarification is that quality didn’t decline, it stayed steady or improved. As AI reduces code production as a bottleneck, teams can spend more time on testing, documentation, and trying ideas earlier. Over time, that can actually improve the overall health of the codebase.

For senior engineers, AI often removes repetitive work they’ve been doing for years and creates more room for architecture, exploration, and proactive thinking. For junior engineers, AI can act as a fast feedback mechanism, a way to explore options, ask questions, and build context before bringing things to a human mentor. That tends to make mentoring interactions deeper and more efficient.

To get David’s full take on this topic, watch the following video snippet:

Breaking the fourth wall; This is a critical observation; one that is aligned with Anthropic’s CEO, Dario Amodei’s comments at the World Economic Forum at Davos earlier this year. He predicts that the industry might be just six to twelve months away from AI handling most of software development [source]. Additionally, the creator of Anthropic’s Claude Code, Boris Cherny, tweeted how Claude Code wrote 100% of the lines he contributed to Claude Code in the month of December’25 [tweet].

Back to the interview.

Do you see differences in AI adoption based on seniority?

David: We didn’t see a strong senior-versus-junior divide. The more meaningful difference is how individuals define their value. Engineers who see their impact in terms of outcomes, shipping value to customers, tend to adopt AI more easily. Engineers who see their value primarily in the act of writing code can struggle more at first, because AI challenges that mental model.

Katrina: When we looked at the data, it supported that view. There wasn’t a meaningful adoption gap tied to seniority level alone.

What assumptions did you make about people adopting AI that you later had to revisit?

Katrina: We initially assumed that because we hire for agency, most people would naturally figure AI out once the tools were available. That turned out to be true for a small group, but many people needed more support. We also learned that early messaging around “10x productivity” wasn’t realistic, there’s usually a learning curve before gains appear.

Creating space for conversations, addressing concerns directly, and running peer-led AI jam sessions helped people build confidence. Over time, teams started coming in with their own ideas rather than waiting to be told how to use the tools.

For a deeper dive on this topic, watch this video snippet:

David: I also underestimated how important joy and curiosity are in adoption. We talked a lot about productivity early on, but what really moved the needle was helping people have an “aha” moment, building something they’d been curious about or experimenting without pressure. I also learned how individual this transformation is. Even with shared tools and guardrails, people engage with AI in very personal ways.

Up until this point, what’s the strongest argument you’ve heard against adopting AI, and why wasn’t it enough to stop you?

David: The most common concern is that AI will degrade quality or erode trust, whether through hallucinations, maintainability issues, or security risks. I take those concerns seriously, but the fundamentals of good engineering don’t change. Strong testing, CI/CD, observability, code review, and security practices mattered before AI and matter even more now because output accelerates.

From a security perspective, knowing where data comes from and where it goes is essential. With the right infrastructure choices and safeguards, these risks can be managed. Avoiding AI entirely would suggest those risks are unsolvable, and I don’t believe that’s true.

Katrina: I often hear, “What if it’s wrong?” My perspective is that adoption doesn’t mean handing full control to AI. There’s a spectrum between human oversight and AI agency, and different problems sit at different points on that spectrum. When you’re deliberate about where AI fits and put guardrails in place, the objections become design constraints rather than reasons not to engage.

We had the time for a few rapid fire questions.

OpenAI Codex or Claude Opus 4.5?

David: Claude Opus 4.5.

Katrina: Same.

Cursor or Replit?

David: Replit.

Katrina: If I’m being honest, neither! I vastly prefer Lovable..

What advice would you give to companies or individuals still not adopting AI?

David: Start by playing. Don’t begin with the hardest problem you already know well. Use AI to explore something you’ve always been curious about but never had time to pursue. Let experimentation and enjoyment guide your learning.

Katrina: Create dedicated time, like a hackathon, where teams can experiment without the pressure of day-to-day work. Giving people permission to focus and play makes it much easier for adoption to take hold.

To learn more about the experiments that KOHO conducted to track developer productivity, take a look at this blog post.

---

Introducing the AI Corner

Where we highlight key updates from the fast-paced world of AI, and what we are reading at Sagard.

• Anthropic just dropped a blog on how software engineering interviews need to change in the age of AI. It includes details on how AI crushes traditional interview problems, why we still need strong SWEs (for reasons beyond “can you code”), and how to design interviews AI can’t game (Hint: make them “weirder”).
• Extending Anthropic’s wins, earlier this year, Anthropic launched Cowork: “Claude Code for the rest of your work.” Anthropic introduced Cowork as a research preview that gives Claude controlled access to a user-selected folder (read/edit/create files), pushing assistants from “answering” into artifact creation + workflow execution. If you are a non-technical individual and have been putting off using GenAI for personal workflow automations, there has never been a better time to start.
• For our fintech readers, we have some good news – “AI for Tables” takes a big step forward! Researchers introduced TabDPT (like chatGPT but nerdier), a tabular foundation model trained to generalize across new structured datasets using in-context learning so it can make predictions on a new dataset without per-dataset retraining or hyperparameter tuning, and shows strong benchmark performance across classification and regression. Finance runs on tables: risk features, transaction histories, fraud signals, credit attributes, portfolio records, regulatory templates. A foundation model that can adapt quickly across tabular tasks could reduce the time-to-value on new datasets where teams usually spend cycles on feature engineering and model selection. [Github]
• In a historic leap for reasoning-based LLMs, Axiom has conquered the hardest math competition in the world, the Putnam Competition. Axiom’s AI autonomously solved all 12 problems with the precision of formal Lean proofs. This breakthrough marks a turning point in machine intelligence, as the AI not only solved 8 of the world’s toughest mathematical challenges within the official exam clock but also pioneered its own unique logical paths that often diverged from human intuition. Last year’s highest score was 90/120. AI scored 120/120, suggesting AI has evolved from simple pattern recognition to mastering the rigorous, verifiable depths of complex mathematical reasoning. 
• Created by Peter Steinberger, Clawdbot Moltbot OpenClaw (third time’s the charm) is a rapidly ascending open-source AI agent that distinguishes itself from standard LLMs by autonomously performing real-world tasks, such as managing emails, scheduling, and shopping, all directly within a user’s operating system. Despite its complex setup and significant security warnings regarding data privacy and malicious manipulation, the tool has gained massive global traction, particularly in Silicon Valley and China, due to its “persistent memory” and ability to integrate with various LLMs. The buzz surrounding OpenClaw is further amplified by Moltbook, a Reddit-like social network where these agents interact independently, and post comments via APIs. Enter the world of sci-fi here.