top of page

How to Build an ML Flywheel That Propels Growth (Instead of Running You Over)



Creating a self-perpetuating cycle of data-driven innovation that accelerates growth sounds enticing. However, building a flywheel that truly propels your business forward is challenging. This guide will equip you with the knowledge and tools to navigate the path to ML-driven success.


Creating Value and Momentum

Creating a machine learning (ML) flywheel can significantly enhance your product, revenue, and valuation. The goal is to create a product that becomes smarter and more efficient with each use, leveraging data to fuel algorithms, generate insights, and make better decisions, ultimately improving the customer experience.


A Significant Investment with High Risks

Let's be clear: this is a major investment with considerable risk. It calls for a different approach than what you're likely doing with your product teams today. ML and AI are evolving rapidly, making it challenging to keep up, let alone implement effectively. This guide will help you understand what ML flywheels are, how they add value, and what to focus on.


Flywheels are Trendy

A flywheel is a heavy wheel attached to a rotating shaft, which helps smooth out power delivery from an engine or motor by regulating speed fluctuations. It takes significant effort to start, but once in motion, each rotation creates more momentum. The concept has been applied to various business areas, from marketing to sales.


Jim Collins popularized the idea in "Good to Great" and a supplemental paper. Collins met with Jeff Bezos and the Amazon team when they were struggling in the early 2000s, leading to excitement about a flywheel focused on customer experience. Andrew Ng from


Success isn't driven by a single factor, but by a series of interconnected elements that build upon each other to create a closed loop. If you can generate momentum, it becomes a powerful force. Here are some useful diagrams from my friends at ML4Biz.org



How It Works

The basic ML flywheel works as follows:

  1. Product creates value

  2. Value attracts users

  3. Users generate more data

  4. Data improves the AI

  5. Improved AI enhances the user experience

  6. Enhanced experience attracts more users


With AI-powered products, this cycle can continue indefinitely as additional data further improves the underlying models.


Examples:
  • Blue River Technology: An agricultural AI system that detects and eliminates weeds, improving its computer vision models with each use in the field.

  • Flywheel.io: A platform for medical AI product development, focusing on medical image annotation. It standardizes data curation and automates machine learning pipelines, accelerating time-to-market for medical AI products.


A High-Stakes Investment

Unlike the small, regular bets you may be accustomed to with your development team, an ML flywheel is a significant investment with potentially huge upside and unpredictable results. A pilot could easily be a 6-9 month effort with no guaranteed results beyond valuable learning. A portfolio approach is necessary, ensuring you can make some big bets. Your investment will need to include talent, data, proper data infrastructure, data curation infrastructure, and compute resources.


The Need for a High-Value, Well-Defined Use Case

The flywheel must address a use case that impacts a key customer need. Solid customer research helps create context and clarity, which is then translated into input-output format for the ML team. ML and AI practitioners stress the importance of well-defined, valuable use cases with clear success metrics. ML doesn't create things but can make them better, faster, and cheaper once baseline performance is measurable.


In a hype cycle, we sometimes focus on technology and forget the value that comes from truly understanding the client and providing context. Business leaders need to sign off on SME-defined use case performance metrics (i.e., business leaders determine what's important, and SMEs determine how to measure it). Establish a baseline performance on current product performance as the starting point. Don't implement ML before assessing what you're trying to improve upon.


Fueled by Your Unique Data

The most valuable flywheels come from your own unique data. This is data that large providers don't have, combined with an understanding of what that data means to your customers, integrated into a product that helps them progress. This is your most defensible position and creates the most value for the business. Ideally, data is proprietary to your product, abundant, high-quality, and paired with inputs and outputs. Value really starts accruing from data exhaust at product-market fit, and most of the value comes from continued customer use.


Durable product moats with AI and ML are often composed of elements such as data moats, business use case moats, talent moats, compute moats, and customer helpfulness moats. These factors often go unconsidered by many business leaders.

Examples of companies leveraging unique data:

  • Amazon, Spotify, Netflix, and Stitch Fix all have recommendation engines based on their own user data.

  • Canva started with customizable templates and has learned over time which templates, elements, and visual styles are preferred by customers based on type, allowing them to make better recommendations.


A Different Approach from Traditional Product Development

These projects are fundamentally different from traditional product development and require a different approach. The project needs to be very tightly defined from the start, and the team needs more than two weeks to make progress. While accountability checks are still necessary, they will look different, and you should work with your team to plan accordingly.


Key differences to expect:
  • A focus on exploration and experimentation. A good early experiment, as recommended by AJan Van Looy, is typically a 10-20 engineering day PoC to test the hypothesis that a specific result can be achieved by model X trained on data Y.

  • Significant time spent on data collection, cleaning, preprocessing, and feature engineering.

  • Early and frequent risk management.


Risk should be an integral part of the process, with key risk indicators (KRIs) identified and measured early. These projects involve upfront work to establish objectives, prepare data, and develop models before interaction is possible, followed by extensive tuning, testing, and validation.


Cross-Functional Team Collaboration

AI/ML teams will be more effective if they are connected to engineering, product, and design teammates. This allows for more context, more feedback on model performance, and better integration into the customer experience. ML engineers, like the rest of the team, want to do good work and deliver for customers. Cross-functional efforts have a better chance of making it to production.


Focus on Learning

There is much to learn about culture, roles, risk, investments, governance, operations, technology, and experience. ML practitioners often feel that leadership lacks understanding of how these projects work, creating unreasonable expectations around skills and outcomes. Overnight success is unrealistic, and no one can tell you exactly what will work for your organization. You'll need to apply what you've learned, make decisions, and continue to iterate.


As mentioned in the post about adoption, a broad training program and focus on making the cultural shift to one that supports ML and AI is necessary.


Image: An old chalk drawing on the driveway


Resources


bottom of page