This Time It’s Different
Note: Finally have a picture opportunity with Adam, Founder of Siri
While everyone flocked to SF Tech Week, I drove to Mountain View to hear three generations of AI entrepreneurs at the Computer History Museum: Jerry Kaplan (’80s expert systems), Adam Cheyer (Siri), and Daniela Rus (robotics, Liquid AI). If history repeats, what—if anything—is different this time?
Jerry reminded us that booms rhyme. The ‘80s wave ran on brittle assumptions—Lisp machines, rule stacks—and collapsed when reality hit production. What’s new today isn’t just the algorithms. It’s the institutionalization of startup culture and how fast ideas get distributed.
Adam framed Siri as a “do engine,” not an AI brand. In the late ‘90s and early 2000s, “AI” wasn’t a selling point; execution was. His point still lands: AI is great at knowing. We need knowing and doing together to create value.
Rus told a great origin story: explain a technical idea clearly enough to a journalist, and suddenly an investor asks, “Why aren’t you building this?” Clarity creates momentum.
I’d heard Adam speak a few months earlier and went back specifically to listen again. Off stage, he repeated a prediction that’s reshaping how I evaluate agent vendors: orchestration agents will slide out of hype by 2027.
That line stuck with me because it maps directly to what I’ve seen working with enterprise merchants over the past few months. Today’s best agent applications mix deterministic rules (policies, limits, approvals) with probabilistic reasoning (retrieval, planning, content generation). The probabilistic layer means the model produces different outputs each time. That variability is fine for drafting an email. It’s unacceptable for initiating a payment, updating customer credit terms, or triggering a compliance workflow.
The benefit of rapid prototyping tools is they let teams build muscle memory through iteration. Tools like Replit and Lovable let non-technical users ship functional demos in hours. I’ve watched merchants build working prototypes without writing code. The learning curve compresses when you can test an idea in the same afternoon you have it.
The Gap Between Demo and Deployment
But making those prototypes enterprise-grade is where the effort multiplies. Managing an AI product requires exponentially more work than planning and building it—because you’re not managing deterministic logic anymore. You’re managing:
Version drift. The model changes; your outputs change. How do you audit that?
Guardrails at scale. What happens when the agent hallucinates a discount code or misroutes a refund?
Stakeholder trust. Finance, compliance, and operations teams need to understand why the system made a decision, not just what it decided.
Which is exactly why this time is different—and why the fundamentals still decide outcomes.
What’s Genuinely Different
Distribution velocity. Workflows ship in days, not quarters. I’ve seen merchants go from “we need to automate dispute responses” to a working Slack bot in 72 hours. Drafts are cheap; iteration is almost free. The constraint is no longer build capacity—it’s decision-making speed and organizational alignment.
What Never Changes
Obsession > idea. The best founders are pulled by a problem they can’t ignore. In a world where anyone can spin up an agent prototype, the differentiator is depth of understanding—knowing the edge cases, the failure modes, the political blockers inside the enterprise buyer’s org chart.
Craft > output. In a world of infinite drafts, taste, error states, and trust are the moat. Your agent can generate a merchant onboarding email in three seconds. Can it detect when the tone is wrong for an enterprise prospect versus a solo founder? Can it fail gracefully when it doesn’t have enough context? That’s craft.
What Gets Overlooked
Data planning > Agent. Personalization requires understanding customers. Understanding customers requires data. Most organizations failed at customer segmentation because of lack of data. Before you deploy an agent to personalize offers or automate outreach, audit whether you have clean, accessible customer data with the right attributes. If your customer records are fragmented across systems or your segmentation logic lives in someone’s head, the agent will amplify the gaps, not fill them.
What This Means for Fintech
If you’re evaluating agent platforms or building AI workflows internally, here’s what to prioritize:
Define ownership at the edge cases. Map every workflow to the question: who is accountable if this goes wrong? That clarity will tell you where agents can run autonomously and where they need to stay in co-pilot mode.
Ask vendors how they handle guardrails. If the answer is vague, the product isn’t ready for production.
The hype will cool. The architecture will mature. The question isn’t whether your team can build an agent—it’s whether your organization can audit one.