AI Wins Through Review, Not Generation

The SaaS apocalypse narrative is backward: AI isn’t killing software, it’s creating a gold rush for companies that understand the hidden bottleneck--review, not generation. Figma’s Matt Colyer reveals that the real competitive moat isn’t in building AI tools, but in designing systems that manage the flood of AI-generated output while preserving human values and context. This conversation exposes a non-obvious truth: the companies that win won’t be those with the best AI, but those that solve the downstream chaos AI creates. For product leaders, designers, and builders, this is the early signal of a new era where control, curation, and continuity matter more than speed. If you're betting on AI to automate creation, you're missing the real fight--governance at scale.

Why the Obvious Interface Fails: Design Needs Divergence, Not Chat

Most AI tools today funnel you into a text box. You type a prompt. You get a response. It’s linear. It’s convergent. And according to Figma’s Matt Colyer, it’s fundamentally broken for design.

"So much of us are defaulting to like chat as the experience for generative UI and I feel like we're starting to enter the second chapter of that."

-- Matt Colyer

Design isn’t about iterating on one idea. It’s about exploding into many. The core of creative work is the diamond-shaped process: first you diverge--generate wildly different concepts--then you converge--edit, refine, select. Chat interfaces don’t support divergence. They lock you into a single thread, nudging you toward refinement before exploration even begins.

Colyer points out that even the most advanced AI agents today are stuck in command-line mode--great for tweaking a landing page, terrible for brainstorming a new product. The real unlock? Moving AI onto the infinite canvas.

Figma’s on-canvas agent is a first step toward this. Instead of typing prompts, you spawn multiple frames--grayscale, sepia, experimental layouts--and let the agent generate variations side by side. This mimics real design collaboration: “Here’s my take. Your turn.” The creativity isn’t in the output, but in the space between options.

But here’s the hidden cost: most teams are still using AI to accelerate convergence. They’re automating the nth landing page. They’re not using it to escape their own cognitive biases. The divergence phase is where AI could be revolutionary--and where almost no one is investing.

The system responds by rewarding speed, not depth. Teams that generate 25 wild ideas before selecting one are slower in the moment. But over time, they build more original products. The delay is the advantage. Most won’t wait.

The Real Bottleneck Isn’t Creation--It’s Review

We’ve been sold a lie: that AI’s value is in generating content. The truth, Colyer argues, is that we’ve already solved generation. The bottleneck now is review.

"The big thing this year will be about how do we review better. I think that's where the bottleneck is now."

-- Matt Colyer

When AI can produce 50 design variants, 20 code PRs, or 100 marketing emails overnight, the human can’t keep up. The problem isn’t output--it’s evaluation. And this creates a feedback loop: more AI → more output → more review → human saturation → distrust → slower adoption.

Colyer sees this playing out internally at Figma. Product teams are drowning in AI-generated proposals. Engineering is flooded with auto-generated PRs. The bottleneck isn’t lack of ideas. It’s lack of judgment at scale.

What’s worse, the tools to solve this don’t exist. We have AI that writes code. We don’t have AI that reviews code with your org’s values in mind. We have agents that draft emails. We don’t have agents that audit those drafts for brand, tone, or legal risk.

The immediate reaction? Add more humans. But that doesn’t scale. The real solution? Build review into the system. Create agents that don’t just generate--they filter, contextualize, and flag. The next wave of product innovation won’t be about doing more. It’ll be about allowing less.

This is where most AI tools fail. They optimize for the first-order win: “Look how fast we built it.” They ignore the second-order cost: “Now we have to review ten times more.”

The companies that win will be those that design review workflows as carefully as creation workflows. That means proactive agents that surface only what matters. That means convergence tools that cluster ideas, summarize trade-offs, and simulate user reactions. That means building not just agents, but agent oversight.

Context Is the New Competitive Moat

Everyone says “context is king.” Few build for it. Colyer’s insight is simple: the best AI tools aren’t the smartest--they’re the most situated.

"Every problem is a context problem."

-- Matt Colyer

His personal email agent only worked when it had memory. His design agent only creates usable outputs when it understands Figma’s design system. The magic isn’t in the model--it’s in the embedding of organizational knowledge.

This is why generic AI tools fail in real workflows. You can’t just plug ChatGPT into your design process and expect usable results. The agent doesn’t know your spacing tokens, your component library, your brand guidelines. It generates noise.

Figma’s MCP server solves this by closing the loop between code and design. Need to update a GDPR checkbox? Pull the live page into Figma. Tweak it. Push it back to code. The agent doesn’t just move data--it understands the medium.

But the deeper shift is in personalization. Most AI products treat personalization as a phase-two feature. Colyer argues it’s the differentiator from day one. An agent that knows your team’s Asana structure, Slack channels, and org chart isn’t just helpful--it’s indistinguishable from a seasoned employee.

Apple and Google get this. They’re not winning on model size. They’re winning on access to personal context. Apple’s privacy-first approach may actually be its biggest AI advantage: users will trust on-device agents with sensitive data long before they trust cloud models.

The system responds by rewarding integration. The more siloed your tools, the weaker your AI. The more connected, the smarter it gets. And the gap widens over time--because context compounds.

The 18-Month Payoff Nobody Wants to Wait For

Colyer’s journey--from rickety Python script to inbox zero--reveals a pattern: the biggest wins come from enduring the initial pain of running your own agents.

He built an email agent. It failed. He added memory. It improved. He added proactive summaries. It became indispensable. But most people quit before step two.

The immediate discomfort--debugging, trust calibration, workflow redesign--creates a lasting moat. Because most teams won’t do it. They’ll wait for a polished product. By then, the early builders have already embedded AI into their DNA.

Figma’s internal shift proves this. In just months, product ops went from manual onboarding to AI-driven setup. Engineers moved from context-switching to agent-assisted flow. Designers shifted from pixel-pushing to curation.

The payoff? Not speed. Scale without bloat. They’re doing more with the same team.

But this only works if you invest in the unsexy work: data pipelines, memory systems, review layers. The teams that win aren’t the ones with the flashiest AI demo. They’re the ones who built the plumbing first.

Key Action Items

• Build divergence into your AI workflows now. Over the next quarter, experiment with tools that generate multiple options in parallel--not just one-at-a-time refinement. This pays off in 12-18 months as your team develops a broader creative range.

• Shift focus from generation to review. Start designing review layers: agents that summarize, cluster, and score AI output. Flag this as a long-term investment--most teams won’t start this until they’re already overwhelmed.

• Embed organizational context into every AI tool. Don’t rely on generic models. Connect agents to your design system, codebase, and internal knowledge. This creates immediate advantage and compounds over time.

• Run your own agents, even if they’re imperfect. The discomfort of maintaining a personal AI system now builds deep understanding. This pays off in 6-12 months when off-the-shelf tools still can’t match your customized workflows.

• Prioritize proactive over reactive agents. Move from “ask and respond” to “anticipate and deliver.” This requires upfront setup but creates inbox zero--level efficiency later.

• Design for convergence as much as creation. Build tools that help teams reduce, refine, and decide--not just generate. This is where most AI tools fail and where real product differentiation lies.

• Treat personalization as core, not optional. From day one, design agents that learn user context--memory, preferences, team structure. This is the moat that generic AI can’t replicate.