Leadership Principles in the age of AI - move quickly with Bias for Action
In 2020, when I lead the innovation program at AWS, I was working with a waste management company in the Nordics who came to me with an ambitious plan. They wanted to install IoT sensors in trash bins across customer sites, deploy cameras in waste rooms, and use computer vision to optimize collection routing and planning and predict fill levels in real-time. The expected investment was substantial - hardware deployment across thousands of locations, connectivity infrastructure, computer vision systems.
As part of the process, we asked ourselves: "Can we look at what data you already have?"
They went back and examined their existing data: years of collection records, customer types and sizes, seasonal patterns, pickup schedules, tonnage by location, and other logs. We also explored readily available external data - weather patterns, local business activity, industry-specific waste generation patterns, public holiday schedules.
The historical data - properly analyzed with AI - could predict waste generation patterns with remarkable accuracy. We could forecast which customers needed more frequent pickups, identify seasonal variations, optimize routes based on predicted fill levels, and anticipate extra service needs. The solution we built cost a fraction of the IoT deployment they'd originally thought of and worked with their existing operations instead of requiring massive infrastructure changes.
We used Amazon's Working Backwards approach to ensure we started with customer needs - reducing overflows, optimizing costs, improving service reliability - rather than starting with impressive technology.
This connects directly to Amazon's "Dive Deep" leadership principle: "Leaders operate at all levels, stay connected to the details, audit frequently, and are skeptical when metrics and anecdotes differ. No task is beneath them."
AI creates a particular trap for leaders: it's easy to feel like you understand what's happening without actually understanding it. You see demos that look impressive. You attend conferences where vendors show polished use cases. But do you understand the limitations? Can you explain what would cause it to fail? And critically - do you understand what data it depends on and whether you actually have access to quality data?
Dive Deep doesn't mean leaders need to become AI engineers or data scientists. It means asking the right questions and not accepting surface-level answers.
Before evaluating any AI tool, map out what data you have, what you could access, and what would be ideal. The waste management company almost spent millions on sensors before they systematically examined their existing data.
Ask your team: "Show me the actual data we'd feed this AI system. Not sample data - our real data. What does it look like? What's missing or inconsistent?"
Think of the waste management AI case. Even if you have existing customer historical data, how would the system handle new customer predictions? Rather than treating this as a of the concept, you could design a hybrid approach: new customers start with industry and similar business averages, then the system would learn their specific patterns over 8-12 weeks. Understanding the failure case leads to a better solution.
For the waste management company, we initially thought to measured prediction accuracy. But diving deeper, what actually matters is service reliability (no overflows) and cost efficiency (no unnecessary trips). You should shift the metrics to focus on business outcomes, not just AI performance.
In the waste management case, it's not enough to rely on your customer service data. You need to seek other data sources, closer to the field. This could be people planning and operating the routes, those speaking with customers, and of course - the customers themselves. Through this you can learn about access time windows, vehicle capacity limits, and driver break requirements that aren't in the data model. Adding these constraints doesn't just improve accuracy - it earns the trust that can make adoption possible.
In AI projects, your data matters more than your technology choice. The most sophisticated AI algorithm is worthless without relevant, quality data. Meanwhile, straightforward AI applied to rich, relevant data can create tremendous value.
An instructive example is that customer complaint data is often biased toward large, vocal customers. Smaller customers who have issues often don't complain - they just quietly switch providers. You need to deliberately incorporate retention data and proactive surveys to get a more complete picture.
Pick one AI initiative in your organization. Block two hours this week to Dive Deep, with a focus on data:
Examine 50-100 actual records that would feed this system
Identify what's missing, inconsistent, or potentially biased
Talk to three people who generate or would rely on this data daily
Ask: "What does this data not capture that's important?"
The leaders who thrive in the AI era won't be those who chase the most impressive AI demos. They'll be those who understand that AI success is built on data foundations - and are willing to examine whether those foundations are solid before building on top of them.