Designing the Future: Inside Canva's AI Strategy with John Milinovich, GenAI Product Lead at Canva

Episode Summary

Executive Summary: Canva’s John Milinovich frames AI product strategy around automation vs augmentation: remove annoying tasks when outcomes are clear, and provide concept-level tools when users are still exploring. He argues AI is shifting design from pixels to objects to concepts, and that success depends on strong evals, human control, and building for the problem rather than any one solution.

Main Topics: Automation vs. augmentation in AI design tools (Priority: 5/5): Milinovich distinguishes tasks users want fully automated from creative work where users want to stay in the driver’s seat. Canva maps its AI portfolio to this split, using automation for tedious tasks and augmentation for fuzzy, exploratory creation. From pixels to objects to concepts (Priority: 5/5): He describes Canva’s original innovation as moving design from pixels to objects, and AI’s next shift as enabling concept-level manipulation across text, images, and multimodal workflows. Product strategy: embedded AI vs AI-native experiences (Priority: 4/5): Canva favors embedding AI at the point of need for existing workflows, but is also building more standalone AI-native products like Dream Lab to expand creative use cases beyond the editor. Evaluation, taste, and human-in-the-loop quality control (Priority: 5/5): Because design has no clean ground truth, Canva uses rigorous custom rubrics, trained evaluators, comparative testing, and online experiments to judge quality and avoid homogenized outputs. Model strategy: build, partner, and ecosystem (Priority: 4/5): Canva uses a three-pronged strategy: its own in-house models, frontier model partnerships, and ecosystem integrations. Milinovich sees fine-tuning as useful in specific cases, but often prefers long-context prompting for style tasks. AI product principles for founders and engineers (Priority: 4/5): He emphasizes deep customer centricity, problem orientation over solution attachment, shipping quickly while maintaining escape hatches, and being willing to pivot when a solution tops out. AI’s impact on architecture and services industries (Priority: 3/5): Drawing from his architecture background, Milinovich argues AI can heavily augment architecture and related services, especially in bid visualization, permit generation, and vertical workflows, but human sign-off and taste remain essential.

Key Arguments: Users split roughly evenly into three groups: those who want more design time, less design time, or about the same; Canva must support all three with different levels of automation and augmentation. Automation should remove friction from clear, annoying tasks; augmentation should help users turn vague ideas into concrete outputs while keeping them in control. AI is moving design from object-level manipulation toward concept-level manipulation, especially in multimodal creative tools. The best interfaces will not be text-only; design is multimodal, so users should interact via clicks, gestures, voice, and other context-appropriate controls. Evaluation is foundational in AI product development; in design, teams need custom rubrics, expert human evaluators, and comparative testing rather than generic benchmarks. Fine-tuning is not always the best answer; for many writing/style use cases, large-context prompting with rich examples can outperform fine-tuning and reduce overfitting. Cost should not be over-optimized too early; focus on making the product work, then good, then fast, then cheap. AI will likely augment rather than replace creative professionals in fields like architecture, especially where legal accountability and human taste are required. The application layer is where major AI value will accrue, especially in services-heavy, under-digitized industries. Founders should build for a specific persona and problem, not for a generic model capability or abstract solution.

Data Points: Canva user scale: 200 million+ users - Described as Canva’s current global user base AI feature usage: 10 billion+ times - Canva AI features have been used across the platform Design preference split: 1/3 more time, 1/3 less time, 1/3 same time - User research on how much time people want to spend designing Design categories supported: 500+ - Scale of graphic design categories on Canva Background-removal crop rubric: 3 criteria - Example rubric: right item picked, cropping good, overcropped vs undercropped Human evaluator agreement: 70-90% prediction quality - Approximate range for narrow domain tasks using trained evaluators/models Fine-tuning example count for diffusion LoRA: Tens of examples - For style-consistent image generation Fine-tuning example count for some evaluation tasks: 1,000-5,000 examples - Examples cited for background generation or social media post evaluation AI cost reduction: ~95% reduction in the last year - Milinovich’s explanation for deprioritizing cost early in development Global GDP: ~$110T today; ~$130T by 2030 - Used to frame the economic opportunity for AI applications Tech sector share of GDP: ~$5T - Current portion of global GDP captured by tech Podcast intro context size: ~50,000 tokens - Host’s workflow example for style-matching writing Architecture win-work share: Majority pre-bid effort - He notes much architecture effort is spent winning projects before client selection

Pivotal Quotes: "At the highest level, there is automation and augmentation." — John Milinovich: He opens the framework for Canva’s AI product strategy by separating clear-task automation from creative augmentation "We believe it's moving to a world from like object-level manipulation into a world of like concept-level manipulation." — John Milinovich: He explains how AI changes the abstraction layer of design, following Canva’s shift from pixels to objects "Make it work, make it good, make it fast, and then make it cheap." — John Milinovich: His development philosophy for AI product iteration and cost management

Implications: For AI builders, the winning pattern is likely hybrid: automate obvious drudgery, augment creative judgment, evaluate with humans, and keep users in control. For design and services industries, AI will expand output and compress workflows, but not eliminate taste, accountability, or human sign-off.