AI Best Practices – 10 Tips To Successful AI Projects

We’re lucky to live in an era where we can supercharge our businesses with the power of AI.

However, onboarding or building generative AI, chatbots, robots and other AI-based solutions brings a fresh set of challenges.


Best practices are the best way to deal with such challenges. This article is a list of 10 best practices gathered through our own experience. Tips and tricks we collected during AI projects that we onboarded or built for our clients. While there may be other best practices these 10 specific ones were the keystone to ensuring that the core of all our AI project is solid and future-proof.


So here we go…


1. Don’t Start With Coding. Talk To The End Users And Build A Clear Business Case. 


An AI project does not start with getting an AI engineer to develop an AI model. The first step is to build a holistic plan focused on people rather than on coding. The holistic plan should also make sure the AI model is not built in a vacuum. The first step is to meet the end users and understand to full context of where the AI model is going to be deployed. This will ensure that you have full clarity of the business case you are catering for.


The most vital element to keep in mind: focus on people, not on coding.


2. Don’t Let AI Jargon Get In The Way.


AI is a vast subject with a lot of jargon. When words like neural networks, supervised/unsupervised learning become the main theme of a meeting some team members will get lost. If not properly tackled this can create major communication issues within the project team. There are two things you can do to ease this :

  1. Upskill your non-technical staff to understand the basics of AI. Get everyone on the team to understand what it building a “digital brain” means.
  2. Ban your AI engineers from using AI jargon during project meetings with non-technical staff. All AI concepts must be explained using simple words.


3. No Data, No Party.


AI projects’ primary cause of failure is surprisingly data. You can think of data like the food or water to your body. Without it, an AI model cannot be trained. Nowadays businesses store a lot of data, but it is not the amount that matters, but how ready and how relevant the data is in relation to the model being built. Supervised AI models need labeled data. Unfortunately, it is a very common occurrence that a team identifies a good use case for an AI model but then cannot get hold of labeled training data. Synthetic data might be a possible solution but that brings with it a fresh set of challenges.


4. Team, Tools  & Budget : The Other Main Ingredients You’ll Need.


Data is not the only ingredient an AI project needs. Team, tools and budget are the other three main ingredients.


Team : Behind every successful AI project there is a team of people. The main team roles are shared between the domain expert that knows the ins and outs of the challenge the AI model is being applied to, and the head of the AI engineering team. Very often the team also requires one or more data engineers to handle the data flow, to clean and prepare the data and to manage and store the output of the AI model, and one or more ML Ops engineer to handle the operations around the model. There is more to an AI project than just onboarding a ready-made AI model or building it.


Tools : The team implementing the AI project, not just those directly onboarding/building the AI model, needs to right tools and the right hardware. And by tools, we are not just referring to using ChatGPT! You might need a set of software tools : data labelling software (like LabelBox), IDEs (like Visual Studio), frameworks (like PyTorch), cloud resources (like Azure) , databases (like MongoDB), code versioning (like Bitbucket) and a set of hardware tools, mainly GPUs (like Nvidia GPUs) which can be either physical or in the available in the cloud. You might also need tools for security, compliance and scalability.


Budget : Like all other projects, an AI project needs a budget. Depending on the level of risk that you see, it might be a good idea to bump up the contingency percentage of the project. If you usually allocate a 20% to 25% contingency, especially in your first AI project, be a bit more conservative and set the contingency higher, to say 30% to 35%.


5. Be Strategic On Whether To Build It Or Buy It.


There are many ways to procure an AI model. You can build your own internal team of AI engineers. You can outsource the project to a company that is focused on implementing AI projects. Irrespective of who builds the “digital brain” it is important to have within the team one or more persons who can take an informed decision on what kind of AI model fits best into your business case. AI is a vast field. As a first step this person need to choose the type of AI model that fits the business case – whether to apply Supervised Learning (i.e. learning from examples), Unsupervised Learning (i.e. learning to identify common patterns), or Reinforcement Learning (i.e. self-learning based on ‘gamification’).


6. Accuracy Is Not 100% – Keep That In Mind.


Unlike software systems built on conditional logic (if-then-else) and rule based systems the accuracy of an AI model is not 100% and will never be. It is relatively easy to get to say 75% accuracy, it is harder to get to 95% accuracy, and it is much much harder to get accuracies close to 99%. So keep this in mind all the time. You cannot deploy an AI model in a business critical use case assuming it can directly replace a manual system and get everything right, since the accuracy of the AI model is never 100%. The accuracy can vary a lot if the data fed to the AI model is not similar to the dataset it was trained on.


7. Be Aware Of Biases.


Humans have biases. AI models have biases too. Biases present themselves not only in processing data about humans (for example an HR AI system sifting through Cvs for a managerial post, and because the data it was trained on has a bias for males in managerial positions it start giving preference to males when sifting through CVs!!!) but biases will always be there in all AI models processing all types of data. To reduce biases ensure that your training dataset is “clean” and as diverse as possible, fairly representing all the types of data. And regularly monitor, not just the accuracy of the AI model but also its results for all known biases.


8. Build Iteratively.


It is almost impossible to build an AI model and to put it in use right first time. So plan your AI project in a way that you build a good enough AI model, connect it seamlessly to the other parts of your business and then be ready to continue to improve on the accuracy and bias-free version of the AI model. It is very common that new training data is made available well after the first versions of the AI model are built. This provides the opportunity to retrain the model and get better results.


9. Seamlessly Integrate Your AI Model


Having an accurate unbiased AI model is great, however an AI model cannot live in a vacuum. We need to be able to reliably pass requests to it for processing and we need a way to reliably capture its replies. Thus we need to plan to seamlessly integrate the inputs and the outputs of the new AI model, via APIs as an example. You should also consider onboarding it into a continuous improvement framework, via CI/CD as an example . 


Also like any other software system the AI model needs to be hosted in a secure, scalable setup.


10. Monitor & Maintain, Constantly.


The fact that an AI model’s accuracy may change drastically depending the data input fed to it makes monitoring and maintenance an essential element of ant AI project. Imagine a self-driving car that is deployed on our roads. It might cater for say 99% of the scenarios, but there will be that 1% which was not covered in its training, and thus it might take the wrong decision. 


Irrespective of whether your AI model is as mission-critical as a self-driving vehicle or not it will always have room for improvement. Thus 1. continuous monitoring of scenarios not catered for and 2. the mechanism to prompt the retraining of the model must be part and parcel of the AI project, not an afterthought.


We hope these practical tips and tricks help you in adding AI to your business smoother, faster and better.


If you have any questions, comments or know of more AI best practices we can add to the list do get in touch.