I was a few years into my time at Amazon Web Services when this principle was introduced in 2021. If I'm being honest, it felt a bit artificial at the time - unlike the original principles that had been forged through years of Amazon's operating culture, capturing it's core beliefs and mindset. This one seemed more like a public statement than an operational guide. The principle reads:
"We started in a garage, but we're not there anymore. We are big, we impact the world, and we are far from perfect. We must be humble and thoughtful about even the secondary effects of our actions. Our local communities, planet, and future generations need us to be better every day."
But watching how AI has evolved in organizations over the past two years, I've come to see this principle differently. It may be the most important leadership principle for the age of AI - precisely because it addresses something most organizations are getting wrong.
The responsibility gap
There's a gap forming in many organizations. AI capabilities are scaling faster than AI governance. Teams deploy more use cases while oversight stays static. Data flows expand while policies remain unchanged. The practices that worked for a pilot - informal review, ad-hoc oversight, "let's see what happens" - become dangerous at scale.
I see this pattern again and again: A team runs a successful AI pilot. Leadership is excited. The mandate comes down to scale it across the organization. But nobody pauses to ask: What worked at small scale that won't work at large scale? What risks were acceptable with 50 users that become unacceptable with 5,000? What secondary effects were invisible in the pilot but will compound with adoption? Or - in some cases - people do care, but then find it hard to define the scaled way forward and get stuck at the pilot stage.
The pilot had one person reviewing outputs. Now there are too many to review. The pilot used anonymized test data. Now it needs production data with real customer information. The pilot was optional. Now it's embedded in a critical workflow.
Why AI amplifies this challenge
Traditional technology scales predictably. If your software handles 100 transactions, you can model what happens at 10,000. The failure modes are largely the same, just faster or more frequent.
AI doesn't work this way. AI systems can behave differently at scale. They encounter edge cases the training data didn't cover. They interact with user behaviors that shift as adoption grows. Small biases in the model become systemic biases in the organization. A hallucination rate that seemed acceptable or was even undetectabe in demos becomes a trust-destroying pattern in production.
And unlike traditional software, AI's "secondary effects" are harder to anticipate:
An AI that speeds up decisions might inadvertently disadvantage certain customer segments
An AI that personalizes recommendations might create filter bubbles you didn't intend
An AI that automates judgment might erode the human expertise you'll need when the AI fails
Humility as an operating principle
The principle calls for being "humble and thoughtful." In practice, this means acknowledging what we don't fully understand about these systems - and building safeguards accordingly.
For AI leaders, humility looks like:
Admitting uncertainty: "We don't fully know how this model reaches its conclusions" isn't a weakness - it's accuracy. Leaders who pretend otherwise are building on false confidence.
Designing for failure: Not "Will this AI make mistakes?" but "When this AI makes mistakes, how will we catch them and recover?" Human oversight isn't a lack of trust in AI - it's responsible scaling.
Monitoring secondary effects: Measuring not just "Did the AI complete the task?" but "What else changed as a result?" - looking into customer satisfaction, employee experience, unintended patterns in outcomes.
Slowing down to speed up: The pressure to deploy more AI use cases is intense. But one serious incident - biased outputs that reach customers, a data breach, a public embarrassment - can stop your AI momentum entirely. Responsible practices and foundations protect your ability to move fast sustainably.
Responsibility scales with capability
When you're in the garage, mistakes are learning opportunities. When you're at scale, mistakes have consequences that extend far beyond your team.
This doesn't mean moving slowly. It means ensuring your governance, oversight, and responsibility practices scale at the same pace as your AI capabilities. The question to ask isn't "Are we moving fast enough?" but "Is our responsibility keeping pace with our capability?"
Organizations that get this right don't treat responsible AI as a compliance checkbox or a brake on innovation. They treat it as the foundation that enables ambitious scaling - because stakeholders trust them to handle the impact.
We started in a garage. Most of our AI initiatives did too - small pilots, limited scope, contained risk. But we're not there anymore. The question is whether our practices have grown up accordingly.
When organizations achieve success with AI and operate at scale, the responsibility extends far beyond quarterly results. Technologist Tristan Harris delivered a powerful wake-up call at TED2025, drawing direct lessons from social media's catastrophic rollout to illuminate what's at stake with AI deployment. In this 15-minute talk, he challenges leaders to reject the false narrative of inevitability and choose "a narrow path where power is matched with responsibility at every level".