How to Develop your AI-MVP?
This article was born out of a few conversations I had regarding Minimum Viable Products (“MVPs”) and AI products in particular. These discussions led me to think about the essential components of an AI MVP. Are there minimum requirements for Artificial Intelligent Products? I did some research and wrote this article as a guide for myself and others.
An MVP reduces risk and costs when creating complex products that are subject to uncertainty and unpredictability. It is important to create and deploy a simple product in a short time frame that can be used by users and still provide value. The MVP can then be used to learn new insights that will help to improve the product in future iterations.
What type of product do you need?
It is important to think about who we are creating the AI-MVP for and what knowledge or expertise the user has regarding the AI product. An internal data product may not require as much development as a private user. An internal dashboard that displays customer churn might be simpler than an integration to display route alternatives within a navigation app. You should make minimum viable products as easy as possible when creating them.
What function will our product serve?
Is it going to sort the items into predefined categories, predict future values, or suggest options to users?
AI can sometimes seem like an amazing tool that can do anything. It is a powerful tool, but it still relies on mathematical algorithms to produce outputs.
Here are some examples:
- Historical patterns
- Get involved
- Take a decision
- Create: deep fakes, speech, text
What type of AI model are we looking for?
Are we able to do this stage without a deep neural network?
This was something we mentioned earlier. Some many models and algorithms can be used to create your AI-MVP. It is important to choose the simplest one that can still be created while still delivering value.
What integration could be made of the AI-MVP into existing systems?
Integrating your AI-MVP into existing systems is a great way to make it useful and add value. You can integrate APIs, plugins and add-ons. Or simply e-mail us with your updates.
This reduces the need to create new routines and mental models for users to use the product. Keep in mind that our focus is still on the minimum viable product. It will be easier to plan integrations in future iterations if you have a clear idea of where the product can be integrated.
- Which systems and tools are already in use by the users of the system or tool we are creating?
- What integrations could you make with existing workflows to incorporate our AI-MVP?
Domain knowledge is required for the AI-MVP.
We must consider not only how the product can be integrated into existing systems but also existing workflows. How easy is the product to integrate into users’ daily routines? This is also important because new habits and behaviors are difficult to develop if the product is not usable.
What routines or habits does the user have in the event of a problem?
What would the user do with the AI?
Value at Day Zero
It is important to continuously improve the product through iterations and to increase the user’s value. However, this does not necessarily mean that the first iterations of the product should not produce any value. It is important to remember that your first user should feel a benefit from your product from the very first time they use it. To ensure that your product is reliable and creates value, you might want to include non-AI support like rule-based technology or a human-in-the-loop.
What other non-AI methods could be used to support the AI-MVP in creating value starting at day zero?
We have covered the outline of the problem and function of our AI MVP. We now need to focus on the data, which is the most important part of any AI model. We can’t do anything without it.
Which data assets are your primary models using for training?
Based on the model functions and framing you created in your first iteration of your model, map what data assets are needed to build your model.
Are you able to use the data already available to create a somewhat efficient model? Where can you find the data if not?
You have a lot more data than you need to build your first model. However, be aware of the licensing. You may need to obtain it in some cases.
What is the source of your data?
Data that is unique to an individual or company is called proprietary data. Data that is proprietary can include data that provides competitive advantages or data that will not be available to the public outside of the organization.
What amount of data integration, cleaning and other activities must be done before your data is usable for training?
When it comes to building ML models, data is everything. It is important to know how much data you have to curate in order to quickly create a simple model.
Are you imagining additional data in the future that can be used to improve your models’ performance?
Think ahead. Is the product itself, or will other factors provide new data that can be used to improve the model in future?
Also, you should consider the Vs of data: Velocity (Velocity), Variety, and Value.
Volume refers to how large the data is. What data do you need to build your model? How does it scale as the number of users and usage increases?
- Velocity: How fast are you collecting data?
- Variation: How are you collecting your data? Are you collecting structured, semi-structured or unstructured data?
- Veracity: Are your data accurate and reliable? This will directly impact how to trust you have in your model.
- Value: Do not just collect data. You should collect data that could be used to solve the problem or improve the model.
Feedback, learning, and measurement
The learning stage is a crucial part of the Lean Startup model. Your product’s main function is to collect insights that can be used in its next iteration. Before you deploy or build your AI-MVP, it is important to think about what you are looking for and how you could incorporate these measurement devices into the process.
- What data are you looking for to improve the model in the next iteration of the model?
- What data are you able to collect about the product’s usability?
- What kind of quantitative feedback can you get from users?
- What is the metric that you use to determine whether the model is successful? How can you decide which features to eliminate or enhance in the next iteration of the model?
- What could be an A/B test to measure performance?
The cost of error
AI models are built on probabilities and are susceptible to making mistakes in the initial iterations. Before you create your model, consider the cost of error. False positives and false negatives can lead to the most serious of outcomes. Self-driving cars can have high error rates, for example. It could also run over pedestrians, which is quite an error. However, it is cheaper to stream a poor movie suggestion than it is for the streaming service.
You can use the cost of error as a guide in your development phase. The more accurate your model needs to be, the more time it will take to develop and tweak it. Sometimes, a model with 80% accuracy is sufficient. Any additional development time will not be necessary. Sometimes you might need more than 99 percent when the stakes can be higher.
If error costs are high or accuracy is difficult to attain in the first iteration of the project, you can always add a human to the loop to help when things go wrong.
In summarizing, think about the possibility that AI algorithms could be incorrect and what will happen.
Let’s Start AI-MVP Development with Brisk Logic:
You should now have an idea of the purpose of your AI-MVP. It should address a problem, deliver value and have the features that make it possible. Keep your focus on the MVP. To gain trust from stakeholders, you must demonstrate value as quickly as possible. AI development can be complex. It is common for AI projects to be abandoned during the development phase before they are deployed. Keep it simple, contact Brisk Logic, to begin with, AI-MVP, and then iterate as needed.