Hi there!

Let’s say you’re a startup founder or project manager who wants to create a new AI-powered solution. The success rate depends on factors like your experience, track record, AI expertise, and more. In this article, I’ll outline different aspects to make success more likely.

I like seeing the project as destined to succeed, then managing risks. Here, I’ll share experiences from projects where I was a researcher or product lead, and from my company which works with clients across industries.

So, let’s explore this risk management checklist further.

1. Concrete Business Goal

It’s obvious, but it’s true. It’s essential to have a clear understanding of the business challenges that the AI initiative aims to address. Without a concrete objective, it’s difficult to define success or measure the impact of the project.

Inadequate goal: Develop an AI system to minimize risks.

Better approach: Create a risk-minimizing solution that reduces financial losses by 30% within 1 year.

The best way is to use the SMART goals setting or a similar framework

2. Financial Effect

AI projects can be costly, so it’s important to consider the financial impact of the project. This includes the cost of developing and implementing the project, as well as the potential return on investment.

Before starting a project, calculate its impact, preferably in financial terms. For example, determine if the project will be profitable or not. Even if everything goes as planned, it may still be unprofitable.

Inefficient idea: Reduce costs by 1% per year on $10 million in revenue with $200,000 annually for system support. This does not account for development and implementation costs, which are significant.

Long-term payback effect: Turnkey development costs $300,000, and annual support is $50,000. Aim to reduce costs by 1% per year on $10 million in revenue over 10 years. Initially, there will be losses, but the long-term impact will be positive.

Effective idea: Sign 10 $1 million software supply contracts per year. Development costs $700,000 and takes 6 months. These are rough estimates, but the general meaning should be clear.

3. Trade-off Between Time-To-Market and Model Accuracy

There is often a trade-off between quickly developing and deploying an AI system and how accurate it will be. It’s important to balance these factors based on the project’s needs.

Sometimes it is better to sell and deploy a rules-based system or use ready-made models and libraries than spend a long time developing a system. When the system generates revenue, improve the models.

There is a famous quote by Reid Hoffman, the founder of LinkedIn:

If you are not embarrassed by the first version of your product, you’ve launched too late.

This approach often works when you can sell the product with the current default model efficiency. It is harder in a competitive market or when the system can cause losses, e.g., for trading or investing.

4. Intellectual Property

AI projects often involve creating and using new algorithms or models. It’s important to consider how to protect intellectual property and who will own the rights to the technology developed.

The requirements depend on how you plan to use your work:

• Keeping research private.

• Using libraries and models for commercial products.

• Using libraries and models only for research.

• Buying licenses for API services or software.

• Other options.

  
  

Consult lawyers to understand the laws in your area.

5. Exclusive and Appropriate Data

AI projects often require access to large amounts of data. It’s important to ensure that the data used is high-quality, relevant, and exclusive to the project.

In general, applied research is divided into two groups when creating a product:

• Search for the optimal open-source or pre-trained model to integrate into the solution. This is common for products where AI is not a core feature.

• Train custom AI models to improve quality or add new features.

  
  

At the pre-project stage, determine whether an off-the-shelf or custom AI model best fits your needs. Cost, timeline, and risks depend on this decision. For custom models, determine:

• Do you have proprietary data? If so, is it labeled? If not, plan to label it. This can be done in various ways, e.g., web scraping, acquiring new data, crowdsourcing, etc.

• Do you lack sufficient data? Then collect and label new data.

6. Domain Expertise

AI projects often require expertise in a specific domain, such as healthcare, finance, or law. It’s important to have the right domain experts involved in the project to ensure that the technology is developed and deployed in a way that meets the needs of the domain.

On the one hand, domain knowledge of the business is required. Representatives of the customer solution (e.g., owner, C-level executives, product managers, etc.) are responsible for this, as they understand the business best.

On the other hand, experience with the technologies of a particular AI domain is a key requirement. It is important to have specialists in this area. Some criteria:

• Proven experience in the form of implementation of several projects in this field;

• Senior level of grade in the teams for the development of these solutions;

• A degree in computer science is desirable to minimize the risk of incorrect design.

  
  

In other words, a time series specialist who has not worked with computer vision should not lead the project in terms of technology, even if he is good in his domain.

7. Integration With Other Systems

AI projects usually don’t work on their own. They require integration with other systems, such as data management or business intelligence tools. It’s important to consider how the AI system will integrate with these other systems and whether any additional development work will be required.

A few key things to think about first:

• Cloud or in-house infrastructure?

• Enough resources for building and running the system?

• How do regulations affect security, data handling, and storage?

• If system-on-chip (SoC) devices are used, understand the architecture and operational principles involved. Choosing incorrectly may require completely redesigning the entire solution.

8. Model Explainability and Transparency

Some business decisions require transparency in the decision-making process. This could be due to regulations, industry standards, or a company’s internal policies. In these cases, the solution cannot function as a “black box.” It must implement algorithms that are human-interpretable.

They tend to be simpler but have lower accuracy. This should be considered when designing a technical solution, evaluating its economic impact, and selecting a team to implement it.

Conclusions

According to many types of research, 80–90% of AI projects fail to make it into production or to deliver any business value.

However, by following the key ingredients outlined in this article and considering potential risks and challenges, businesses can minimize the risk of failure and increase their chances of success.

In my experience at AI Works Lab, you can decrease this rate to 30%. By doing so, businesses can gain a competitive edge, drive growth and success, and ultimately deliver value to their customers.

Let me know in the comments if you would like to dive into other subjects of application AI in your business domains or make more details about specific points from this article.