In February 2024, technology giant Microsoft published a blog post about the three major artificial intelligence trends that the company intends to follow over the next few years. The trends mentioned include small language models focused on narrow tasks, multimodal AI solutions, and scientific modelling. Microsoft executives have been making sizable investments in AI, particularly with Bing Chat for search engines and the Copilot multimodal AI tool, so it makes sense that they are mentioned within the three development trends they are following.
For many business owners, company executives, and top-level managers, AI implementation is not something they are contemplating beyond using current off-the-shelf solutions such as Microsoft Copilot and Google Gemini. Many others, however, are already implementing AI solutions for customer service, information security, inventory management, accounting, supply chain operations, and other business processes. Unlike solutions such as Google Workspace, a productivity suite that supports business in general, many AI solutions are being implemented in a bespoke fashion which means that they are being developed or configured to meet specific business needs.
Boxed and off-the-shelf solutions such as ChatGPT by OpenAI offer a starting point for many businesses, but they are filled with limitations and inaccuracies caused by their LLM scope. We have already seen this with a chatbot implemented by Air Canada for customer service; in 2022, it provided incorrect information to a passenger inquiring about partial refund policies, and the company eventually lost a civil claim in a British Columbia court. The true power of AI for business lies in its potential to be customized for specific business goals; with this in mind, let’s delve into the AI development lifecycle to understand how you can build a solution tailored to your unique needs.
Charting the AI Development Process Through Scope and Objectives
Getting back to the Air Canada incident, which was caused by a pre-trained AI model, we can see that the planning stage of the AI project lifecycle was not effective in choosing the right solution. With the current rate of inaccuracies found in major LLMs such as Google Gemini being as high as 30%, you don’t want such a solution to handle a legal scope such as refund policies. A better solution could be a small language model specifically trained on existing documentation and extensively tested through several customer service scenarios. The objectives are generally easier to formulate than the scope, but they should be realistic; if you aim to exceed the accurate productivity of employees, you will increase the risk of mishaps exponentially.
Building the Foundation Through Data Collection and Preparation
The AI concept development stage is generally followed by data collection and preparation. In this step, you want to remember the “garbage in, garbage out” (GIGO) axiom of computer science, which happens when flawed information underpins any data system. The opposite of GIGO involves delivering high-quality results from high-quality data. Remember that AI models learn from the data they are fed; if the data is incomplete, inaccurate, or irrelevant, your AI solution will assimilate these flaws and deliver them at the worst times. Raw data often requires cleaning and organizing before it can be used for training, and this might involve removing duplicates, correcting errors, and coming up with standard formats.
Shaping Your Solution Through AI Model Development and Training
This stage tends to be most exciting in the journey from AI concept to deployment. This is when the AI consulting services firm you retain will recommend how to build the solution. In some cases, an off-the-shelf or pre-built AI platform may be suitable; other cases may call for granular development. As this stage unfolds, you get to see how your scope and objectives start coming to life. In all cases, training is the bulk of AI development, and it is partially handled by machine learning (ML) to speed up the process.
Fine-Tuning Your AI Solution Through Testing and Validation
Of the various AI development stages, this one may not be as exciting as the previous one, but it must be undertaken before AI application deployment. You can’t skip or “speed-run” this stage; otherwise, you may end up with an Air Canada situation or worse. Before deploying your AI solution, it is crucial to assess its effectiveness through rigorous testing, which in some cases involves feeding the model unseen data and evaluating its performance against predefined metrics. If the scope and objectives of your project call for LLM implementation, you will have to implement several guardrails through testing and validation, which can be a long and tedious process involving key employees. In the past, major tech companies such as IBM have abandoned their projects at this stage because the AI output could not be validated; this is avoided now during the scope and objectives stage, especially when you work with expert consultants.
Putting Your AI Solution to Work Through Deployment and Monitoring
Once your AI solution has been meticulously trained, the next step involves deployment, which should have gone through dry runs during the previous stage of testing and validation. The specific process will vary depending on your chosen platform and data infrastructure. Some deployments may involve integrating the model into existing systems, setting up a dedicated server environment, or going with the cloud computing route, which is what many companies choose. This should not be considered the finish line of your project; you want to monitor everything from output to access and from performance to resource usage. You may need to retrain your AI solution with new data to maintain accuracy and ensure that it stays aligned with evolving business needs, and this would call for further monitoring.
As an executive or business owner, there’s a good chance you may need to catch up to competitors who have already implemented AI. While it is always important to sharpen your competitive edge, you don’t want to move too fast by skipping the stages listed herein. If you understand each stage, from defining objectives to deployment and monitoring, you can properly navigate the complexities of AI development. Beyond keeping you competitive, AI can also unlock new levels of productivity in your business, but only if it has gone through proper implementation.
