How to apply artificial intelligence to solve business problems | by John Ade-Ojo | Jan, 2022

In the business capture phase you work together with your business SME or analyst to map out your business problems. Here are some guidelines for this:

• Business Area & Owner: Take this as a reference point, you want to know who could potentially fund a full implementation if your lab were successful.
• Stakeholders and their roles: You want to note who the SME contacts are, what their position or title is, who the analysts are, the data scientists, sponsors etc.
• Problem: Write down the business problem to be solved, you need to understand it by talking to your business SME. If the problem seems too complicated, break it down into simpler subproblems and deal with them separately. Consider answering the following questions: What is driving the change? Who owns the processes? What does a successful solution look like?
• Business advantage: You estimate the benefit to the company if your solution were implemented. Initially this will be a “back of the envelope” calculation based on the best possible results. You will refine this later when the prototype is built. You want to express this business advantage quantitatively and qualitatively.

Once you have captured your business problem, you can try to frame it at a high level in terms of AI&ML.

• Value proposition: Try to understand who the end user is, what their goals are and how they can benefit from machine learning or AI.
• Decision: Intelligent systems can be effective in making decisions. Determine the decisions that your system makes or supplements, the scale and the frequency. In regulated sectors such as banking, there are often restrictions on ‘black box’ models. So you will also need to understand how interpretable the decision-making process should be.
• ML services: Your intelligent solution may require you to use a combination of ML or AI services provided by a cloud platform or ML PaaS. It’s a useful exercise to work these out at a high level and decide what they might be. Broadly speaking, ML services fall into the following categories: language, NLP and text extraction, computer vision, and speech recognition.
• machine learning: Not all solutions require you to build a custom machine learning model. You have to judge whether you should do this or not. Keep in mind that building your own custom ML solution is often more complex than using ML services.

Data fuels your machine learning engine, so it’s important that you document your data requirements in detail. Initially, consider your training and inference requirements. If you’re training a machine learning model, what features can be predictive? If you’re just making inferences, what data is needed? For example, if you are building an ID verification engine, you may need some photo IDs, you need to know how to capture this data.

Data management & purchasing

Also consider your approach to data management and purchasing.

Do you use structured/semi-structured data, unstructured data, or both?

There are different requirements for dealing with structured and semi-structured data versus unstructured data. The way you store your data, the processing costs and the available volume are just a few things to consider.

In the industry, there are strict rules about data and how it is used under the GDPR. You want to have a strategy for meeting this. Determining whether personal data is used, the data source, primary contact for the data and how long you want to keep it will help you get through the inevitable data audits.

Finally, you want to capture your acquisition costs, data is not free and your value proposition can be greatly reduced if expensive data is required.

Data Labels

In an industry setting, you may want to take a supervised machine learning approach, but you may not have labeled data. You want to specify your data labeling strategy here. Will you be using a service like AWS’s Mechanical Turk, or will you be learning unsupervised to create your labels?

Bias Limitation

You need to understand the bias in your dataset and implement strategies to manage it. Examples of prejudice are:

• Sponsorship bias*: Are your data sources from a sponsoring agent? Sponsor agents tend to suppress data that can be reputation-damaging. For example, data on the influence of tobacco on health collected by a cigarette manufacturer.
• Self-selection bias*: A form of selection bias where your data sources come from those who voluntarily provided them. This is most of the research data.

*Source: Machine Learning Engineering, Andriy Burkov (2020), p. 44–47

Reach the details of your intelligent application. You need to use the Business Capture and AI Problem Framing exercise to educate you here. Things to consider are:

• Approach: Describe in writing the approach you intend to take to resolve your problem. You should have a good idea of ​​what machine learning services you need and whether you need to train a model.
• Performance requirements: What are the minimum performance standards you expect? You need to see this in terms of latency requirements and performance of any models you build or use on test data. Your performance goals should align with your business requirements.
• outputs: What will be your output? Does your intelligent system provide inferences to another app, or is it used to generate a dashboard of insights? This affects the design patterns you choose later.
• Research: You need to get a sense of the landscape of apps already out there that can solve your problem. It is often more efficient to go for an ‘out of box’ solution rather than trying to reinvent the wheel. Another thing to consider is whether the problem you are trying to address is a typical ML use case. There is a lot of uncertainty about the success of ML projects, it helps to know if this has been resolved in the past to give you an idea of ​​your chances of success.
• High level solution design: Map your solution design end-to-end. If your solution is cloud-native, keep it at the service level and don’t go into detail about your ML pipeline yet (if you’re training a model). These diagrams are helpful for engineers, who can help you scale for production at some point.
• Machine learning strategy: If you are training your own model, list the models you are going to try along with a base model. It is worth mentioning the limitations of each model here as well. You will also need to map out a more detailed machine learning pipeline here, including steps for data segregation, data scrambling, feature engineering, model training and evaluation loops etc. You will probably need to loosely define your strategy for monitoring ML models you train as they go live. Consider the attributes you want to monitor, how you set up your base performance, and your monitoring schedule. By default, you probably want to check data quality, model quality/performance, model bias drift, and attribute attribution.
• Model training and evaluation: What are your splits for training, validation and testing datasets? What strategies will you use to fine-tune your hyperparameters and test your model?
• Platforms, tooling and infrastructure: Think about the services you need and how much they will cost you. If you’re using a cloud platform like AWS, there are tons of machine learning services you can use, and you can estimate costs using the AWS calculation tool. There are similar options for other cloud platforms, including Azure and GCP.

Whether you’re training your own custom models or leveraging existing ML services, you need to track their performance against your requirements.

Tailor-made machine learning solutions require you to have a scientifically rigorous strategy for evaluating model performance to avoid overfitting and maximize the chance that your model will work well in production.

Remember, your end goal is to build a model that can work in production and add value to your business. Once you’ve mapped out all the requirements and built the prototype, you’ll have a better understanding of how feasible your solution is to scale. Things to consider here are:

• Data infrastructure: Data access, volume and quality.
• ML/solution: Technical resources available, existing solutions, knowledge of solution.
• Processes & Systems: Are business process changes, system changes, or organizational changes required to implement your solution?
• know how: Is the tech and domain knowledge available to you? And how long does it take to retrain your team to meet the requirements?
• Solution to live: How long will it take for your solution to be viable?

The initiative for applied artificial intelligence has released a white paper that gives you a method to assess the feasibility of your intelligent solutions.

You must be able to assess the value versus feasibility of your solution. You want to go for solutions that have high value and high feasibility, very few solutions you design fit into this category from the start. You should revisit the original design to see if you can take a more feasible approach. If you’re in an organization that’s at the beginning of the intelligent application adoption curve, take a look at several of these solutions to see where you can invest.

Note: You would have estimated the value of your solution at the capture stage of your business.

• top left: Ambitious projects that are currently difficult to execute, but have high business value.
• Bottom left side: Trash, relatively low business value and low feasibility.
• bottom right: Low hanging fruit, low business value but high feasibility.
• top right: Missiles, high business value and high feasibility.

Applying artificial intelligence to solve business problems is difficult. Because of all the moving parts, it’s definitely worth standardizing your approach as much as possible to make the process more efficient. Feel free to use what is useful to you here and modify it as needed.

️ I like to help people by sharing my experiences with data science in the wild. If you are not already a member, please consider subscribing to Medium to receive more helpful content from me.