5+5 Truths You Should Know About AI

A Foreword 

Before I dive into the details, I’d like to set expectations straight. I’m not a writer, nor am I an engineer, so don’t expect this to be the typical article filled with technical jargon or politically correct statements. My motivation in sharing these truths with you comes from a place of honesty and a commitment to integrity, a value I hold dear as a business development director at Scopic but also as a parent and a member of the tech community. 

Whether you’re an entrepreneur or an intrapreneur, a leader, or a visionary within your company, there are a few things you should be aware of when considering AI for your business. My goal is simple: to share some of the realities of implementing custom AI solutions I have experienced, hoping to aid your decision-making process. No beating around the bush—just the plain truth. Here’s what I think you need to know: 

Building a custom AI solution isn’t as straightforward as some might lead you to believe. Sure, if all you need is a simple AI chatbot for your website, the process might be relatively quick and easy. But if you’re aiming for an AI that truly adds value and drives change in your business, prepare yourself for a complex journey. 

Effective AI software development involves deeply understanding your business needs, being clear about your vision, and actively engaging in the process. When we built Orthoselect’s AI for segmentation for their dental software application, it took us over 6 months and over 1.5k hours to achieve the desired outcome. This wasn’t a one-size-fits-all approach; it required constant communication, iterative testing, and refining. We keep refining our solution and maintaining the model to this day.  

This ML solution had an immense impact on the business; it managed to reduce the technician’s time from 2 hours to correctly configure the braces on the 3D model of the patient’s mouth—to just 5 minutes or less (how about that 2400% boost in productivity), making Orthoselect the fastest provider of their innovative indirect bonding/bracket transfer trays. 

Similarly, we are currently working on Abby Connect’s AI Voice Assistant (an AI contact center solution). We have invested 2.5k hours so far and had to make adjustments along the way to address issues with latency and prompts, which lead to combining multiple services to achieve the desired user experience.  

Key takeaway: AI-powered business solutions that deliver real impact take time, effort, and commitment from both sides. Don’t expect miracles in a month. Be prepared to invest your time in testing, feedback, and data sharing to get the AI working as expected. 

AI’s performance hinges on the quality, quantity, and processing of data. Think of data as the fuel that powers the AI engine—if the fuel is poor quality, the engine won’t run efficiently, if at all. 

• Data Quality: The phrase “garbage in, garbage out” couldn’t be more accurate when it comes to AI. If your input data is flawed, you’ll end up with flawed outputs. For instance, if you’re feeding an AI model with inaccurate or biased data, don’t expect it to make smart decisions. While modern AI systems can make inferences or predictions that seem intelligent, they still rely heavily on the quality and breadth of the data they are given.  

• Data Quantity: AI thrives on large datasets. The more data it learns from, the more accurate and reliable it becomes. Data quantity is why companies like Google and Amazon have such powerful AI systems—they have access to vast amounts of data that continuously feed and improve their models. Not that you should compare yourself to these titans, but you should make sure you have sufficient data to achieve good results for your use case. 

• Data Pre-processing and/or Augmentation: Raw data is rarely in a state that’s ready for AI to train on. Almost always, data pre-processing is needed, whether it’s cleaning it of formatting to fit the AI model’s requirements. For example, if you’re using an open-source model, you might need to either manipulate your data to meet the model’s parameters or change the way the data is loaded into the model.  

In another case of Orthodontic’s AI, the client had a small dataset of 1000 samples. To augment it, we built an algorithm that would synthesize new samples from combining existing ones, and then we selected the quality combinations only to increase the dataset from 1k to 3k in size. 

Key takeaway: Data handling is often the most time-consuming part of developing an AI application, but it’s crucial for success. If you don’t have the data, start collecting it now. The better your data, the better your AI will perform. If you want to know if your data is good enough, schedule a free consultation to discuss your project! 

AI isn’t a plug-and-play solution. Depending on your use case’s complexity and any compliance implications mandated by your domain or industry practices, research, testing, and model refinement may be required. Sometimes feasibility assessment is the way to go where the use case is too big to achieve in a realistic timeframe or within your realistic budget. Cut your spending and do a feasibility assessment first.  

I remember on one occasion, we spent 3 months, 5+ developers, over 1,000 hours, and tested over 10 different open-source models—plus built one ourselves from scratch—just to explore the challenges in building an AI-based health monitoring system. Our main issue here was not only the lack of datasets for the client’s use case but also the lack of scientific research conducted in that specific sub-domain.  

In certain cases, the challenges may be just too big to overcome in a realistic timeframe, so if your case is very niche or something no one has accomplished before, be sure that proper research is mandatory to assess the feasibility of that solution. 

• Licensing: Depending on your use case, some AI/ML models or datasets come with licenses that may restrict their use. For example, some models are available under open-source licenses like MIT, which allows for broad usage and adaptation. However, other models may have more restrictive licenses, requiring you to rebuild or significantly modify the model to fit your needs. This is especially true if your use case involves sensitive or proprietary data. 

• Customization or New Model Development: Even when using pre-existing models, customization is necessary. This could mean retraining the model with your data or developing entirely new algorithms tailored to your specific application. For instance, if you’re developing an AI for medical diagnostics, you’ll need to train the model on a large dataset of medical images, which might require additional research, data processing, and development efforts. However, it is always a good idea to search for open-source models to evaluate before a decision is made to build a new model from scratch. Especially if licensing is not a problem, open source saves a ton of time. 

Key takeaway: Be prepared for a learning curve. Research and customization are often required to build a solution that fits your business’ unique needs. Some AI solutions can be deployed with minimal customization, depending on the complexity of the task; others require building models from scratch. It is what it is! 

Depending on its size (AI engineers call it model weight), AI may require significant computing power, and if you want your AI to perform at the speed your customers expect (if you have a customer-facing AI component planned), you’ll need to invest in the right infrastructure. This isn’t just about having powerful servers; it’s about having the right architecture to support real-time data processing and analysis. 

• Speed Matters: Slow AI is not an option in today’s fast-paced world. If your AI takes too long to generate results, your customers will lose patience, and your business could suffer. Ensuring that your infrastructure can handle the demands of AI means establishing high-performance hardware, cloud computing services, and robust data pipelines. 

• Scalability: As your business grows, so will your AI’s demands. Your infrastructure needs to be scalable to handle increased data volumes, more complex models, and faster processing times. This might involve setting up distributed computing environments or leveraging cloud-based AI services that can scale with your needs. 

• Automation: Some automation of the deployment process has to be put in place for future re-training practices. This reduces the need for intervention from a Devops all the time, giving you more independence in this regard. 

Key takeaway: Investing in the right infrastructure is crucial for AI success. While it may be costly upfront, the long-term benefits in terms of speed, scalability, and customer satisfaction are worth it. 

Before diving into AI, it’s essential to evaluate whether it will deliver a return on investment (ROI) for your business. AI isn’t a magic solution that works for everyone—it needs to be aligned with your business goals and provide tangible benefits. Consider the following: 

• Differentiation: Will AI make your product or service stand out in the market? If AI can provide a unique feature or significantly improve customer experience, it could be worth the investment. Be careful how you approach this. Many businesses will dive into AI and some costly solution without consulting their customer base and then raising prices to cover the costs and start ROI fast. This is bad practice. I suggest you conduct a customer survey first before you start anything. If most of your customers see this feature as valuable and adding to their experience (aka they are willing to pay more for it), then keep on planning. On the other hand, if you think this could bring a certain differentiation aspect that opens your business to a different customer group and fulfills their needs, then continue. 

• Cost and Time Savings: Can AI automate routine tasks that are currently draining your resources? If AI can handle repetitive work more efficiently than humans, it could save your business both time and money in the long run. It would allow your best people to find new purpose in the business, and it may also save thousands of dollars in new personnel costs. 

• Data Availability Costs: Do you have enough data to support an AI solution? Does it exist and is open for use, or can you collect it? If data is scarce or difficult to obtain, the cost of acquiring and processing data could outweigh the benefits of implementing AI. Make sure you have samples of your data with you when you discuss with an AI development partner. 

Key takeaway: Conduct an ROI analysis before committing to AI to better ensure it aligns with your business objectives and provides a clear path to profitability. If you’re interested in AI consulting services, get a free consultation today. 

Interested in learning more? Check out 5 more truths you may want to be aware of!

5 more truths about AI