Thinking about adding ML to your product? Start here first.

Before diving into your first ML project, consider this sobering statistic.

85% of ML projects fail.

A common reason cited is that product teams focus on the wrong areas.

If you sprinkle ML like fairy dust over your entire codebase, you’ll just become another statistic.

So what can you do? Let’s get into it!

Incorporating ML is all about finding the right opportunities and dodging the pitfalls, a bit like a game of snakes and ladders.

So, what’s the playbook?

Below is your 1-hour activity that product teams can use to spot the most promising places to introduce ML into your product. I’ve also included some useful ML tools and services.

Step 1: Find the ladders 🪜 [when to ML]

1) Identify repetitive, predictable tasks 🌀

If a task makes you feel like you’re in Groundhog Day, it’s probably a good candidate for ML. Think about areas in your product where users repeat actions, and you could automate decisions with data.

Examples:

• Email sorting: Auto-labeling or prioritizing emails based on content

• Recommendations: Suggesting related products, content, or actions

• Image tagging: Auto-tagging photos based on image recognition

ML Products:

• Google Cloud AutoML: Automates training models to sort, tag, or categorize data like emails or images.

• Amazon SageMaker: Great for building, training, and deploying custom models for recommendations or sorting.

• IBM Watson Visual Recognition: Fast image tagging and visual pattern detection.

Play: Start small. Identify a specific task your users repeat, and explore using these tools to automate and optimize it.

2) Personalize experiences 🎯

Everyone loves a personal touch, but manually tailoring experiences for each user? Yikes. Enter ML. By analyzing user behavior, you can adapt the experience to each person’s preferences.

Examples:

• Content feeds: Curating articles, videos, or music based on engagement

• Shopping suggestions: “People like you also bought this flamingo-themed coffee mug!”

ML Products:

• Dynamic Yield: AI-driven personalization for eCommerce, mobile apps, and websites.

• Algolia Recommend: Adds personalized product or content recommendations using pre-trained models.

• Optimizely: Experiment with personalized experiences across platforms and channels.

Play: Look at your user journey. Where could ML make the app feel like it’s speaking directly to each person?

3) Predict what happens next 🔮

If you find yourself guessing about the future, demand, churn, capacity, that’s a prediction problem, and prediction is exactly what ML is built for.

Examples:

• Churn prediction: Flag accounts likely to cancel before they do

• Demand forecasting: Anticipate inventory or server load

• Lead scoring: Rank prospects by likelihood to convert

ML Products:

• Amazon Forecast: Time-series forecasting with no ML expertise required.

• Google Vertex AI: Build and deploy custom prediction models end to end.

• DataRobot: Automated modeling for churn, demand, and risk.

Play: Pick one number you currently guess about. Ask whether a model trained on your history could guess better.

4) Catch the weird stuff 🚨

When you need to spot the needle in the haystack, fraud, spam, anomalies, ML shines because the patterns are too subtle and too numerous for hand-written rules.

Examples:

• Fraud detection: Flag suspicious transactions in real time

• Content moderation: Surface spam or abusive content

• Anomaly detection: Catch unusual spikes in your metrics

ML Products:

• Amazon Fraud Detector: Pre-built fraud detection trained on Amazon’s own data.

• Sift: Fraud and abuse prevention as a service.

• Hugging Face: Open models for moderation and classification.

Play: List the “needle in a haystack” problems your team handles manually today. Those are prime ML territory.

Step 2: Dodge the snakes 🐍 [when NOT to ML]

For every ladder there’s a snake. Slide down one of these and you join the 85%. Before you commit, check that you’re not standing on any of them.

🐍 Snake 1: You don’t have the data. ML learns from examples. No data, messy data, or too little of it, no model. Be honest about what you actually have.

🐍 Snake 2: A simple rule would do. If an if / else statement solves it, ship the if / else. ML adds cost, latency, and unpredictability you don’t need for a solved problem.

🐍 Snake 3: You can’t tolerate being wrong. ML is probabilistic, it will be wrong sometimes. If a single bad call is catastrophic and there’s no human in the loop, think twice.

🐍 Snake 4: You can’t measure success. No metric means no way to know if the model helps or hurts. If you can’t define “better”, you can’t build it.

Step 3: Roll the dice 🎲 (15 min)

Now put it together. On your whiteboard, list 3–5 candidate problems. For each one, score it against five questions:

1. Is it repetitive or prediction-shaped? (ladder)

2. Do we have enough good data?

3. Would a simple rule solve it just as well? (snake if yes)

4. Can we tolerate being occasionally wrong?

5. Can we measure whether the model is helping?

Action: Pick the single candidate that’s a clear ladder on every question. That’s your first ML project. Park the rest.

ML isn’t fairy dust, and it isn’t a strategy. It’s a tool that’s brilliant for a narrow set of problems and a money pit for everything else.

Find the ladders, dodge the snakes, pick one clear win, and ship it. Do that and you’ll be in the 15% that actually make ML pay off.

See you next week. 👋🏼