How to choose the first machine learning project for your organization

Start with the low-hanging fruit to reduce friction and build trust.

One of the questions I often get from product managers who want to get started with AI and machine learning in their company is “Where do I start?” Diving into AI/ML can feel overwhelming, especially if your organization lacks experience in the field. Building ML products requires different types of skills and processes that must be incorporated into your company gradually.

In my experience, the key to success lies in starting small and being strategic. The goal of your first project should be to get your feet wet with ML without causing much disruption in your organization. This can help you build the infrastructure and stepping stones to becoming an AI/ML-driven company.

Start with low-hanging fruit

The best approach is to start with "low-hanging fruit"—projects that are relatively simple to implement but offer clear and measurable business value.

Low-hanging fruit in AI/ML projects typically share these characteristics:

- Minimal disruption to existing workflows. These projects can be integrated seamlessly with your current operations. For example, you can use ML to automate one stage of a multi-step task that is usually done manually.

- Clear success metrics, making it easy to measure results from the outset. For example, it can be time saved per task or the amount of work completed in a specific time.

- A quick implementation timeline, allowing you to see value without long delays. Thanks to generative AI tools, the implementation timeline for many prototypes and proof-of-concept projects is shrinking.

Selection criteria for your first AI project

When selecting your first AI/ML project, focus on three key factors: data availability, process characteristics, and impact.

1. Data Availability

Start with tasks where you already have a good amount of data, preferably clean and structured in a data warehouse. If not, look for areas where you have a lot of unstructured documents that hold a lot of value. In many cases, you can get a lot of value from raw documents or get them ready with minimal annotation.

If you have no data available, look for tasks where data collection or creation is easy and low-cost and can preferably be carried out by a small team. Ensure that the data you plan to use complies with privacy and regulatory standards.

2. Process Characteristics

Look for processes that involve a lot of manual effort and are too complex for simple rule-based automation. If it's a task that humans currently find repetitive but hard to codify, that's a great sign that it could benefit from AI. Tasks that have clear input-output relationships with predictable patterns are ideal for machine learning. 

Also try to choose a project that has minimal dependencies on other systems or teams, so you can control the process and deliver results quickly. To reduce the barriers to entry, you can choose tasks that can be improved without full automation. For example, even if the ML model can improve the speed of a task by 10-15%, it can have great value for your organization.

3. Impact  

Ideally, your project should deliver a quantifiable ROI, through cost savings or time efficiency. But the main point of your first project should be to get comfortable with the process of building ML products. Adopting machine learning requires many support functions such as setting up data pipelines and model serving platforms, performance monitoring and model retraining. A small project will help you gradually experiment and build up those pieces before undertaking large-scale projects.

Real-world examples of first AI/ML projects

Here are three practical examples of AI/ML projects that many organizations have successfully implemented as their first foray into machine learning.

1. Customer Support Ticket Classification

This is often an excellent first project for a company that serves many customers online because you already likely have thousands of tickets to train your model. Moreover, tickets often have a lot of unstructured text that can’t be categorized through rule-based approaches, which makes them suitable for machine learning.

Processing tickets requires reading the content, which could take a lot of time. Machine learning can help save a lot of time. Ticket classification is an easy first step that can be implemented quickly and integrated into existing workflows without disrupting them. Customer service staff can gradually incorporate the ML system into their workflow, build trust, and provide feedback that can help improve the underlying model.

Thanks to large language models (LLMs), ticket classification can easily be implemented without the need to train your own models. You can get good results with a handful of annotated examples and gradually improve the model’s performance as you gather more data. Eventually, you can train specialized models for the task once you have enough examples. Furthermore, LLMs can help generate additional insights such as action items and summaries.

The same process can be applied to any other kind of task that requires document classification, such as emails, contracts, resumes, etc.

2. Document Processing/Data Extraction

Documents such as contracts, invoices, and purchase orders come in different formats that can’t be easily processed with rule-based approaches. If your organization handles lots of such documents, extracting information from them can be a good place to start to highlight the value of machine learning in your organization.

Creating an ML system for extracting document data is very easy, especially with LLMs. Like the previous example, you can easily create a working system with a few documents and the expected result. You can also craft your prompts to get formatted outputs that can be easily fed to downstream systems. 

Moreover, you can start small, only instructing the model to extract very basic information, freeing up the user’s time to address more complex tasks. The model can be easily monitored and validated against existing processes. Gradually, as you gather more data, you can shift more workload to the model, including reasoning and summarizing over the content of documents.

Implementing a document extraction system and integrating it into existing tools will be easy. The ROI is clear and users will immediately see the benefits of reducing manual entry, which will help you get stakeholder buy-in for future ML projects. 

Remember, the goal of your first AI project isn't to transform your business overnight - it's to prove the value of ML in your organization and build momentum for future initiatives. Start small, focus on quick wins, and build from there.