Overcoming the challenges in building AI solutions

Overcoming the challenges in building AI has evolved from a tech-first topic to customer experience and productization. AI is the most significant transition the internet has seen since the web browsers. This is the opportunity of a lifetime. Even better, it’s an opportunity well within our grasp. If you’re an aspiring disruptor, passionate about Customer Experience, or a product person, read on to learn how to make AI a reality. If you missed our previous post,  Creating Solutions with AI, check it out.

We continue on the AI solutions theme. First, we’ll list the essential requirements of building AI, identify which have been well addressed, and what you should focus on. Second, we’ll study a common concern among our clients: how to keep your data safe with AI. Third, we encourage you to become a transformative force in AI rather than a sitting duck waiting for someone else to disrupt your business.

“It helps to think that your objective isn’t building AI.”

The complexity of building an AI solution has decreased over time in three areas

Building an AI solution used to require deep pockets and a Ph.D. battalion. The following three fundamental requirements used to be showstoppers (but are not anymore).

1. Infrastructure: solved. Public cloud services (e.g., AWS, Azure, GCS) became readily and cost-efficiently available in the early 2000s.
2. AI algorithms and open-source ML libraries (e.g., TensorFlow, PyTorch, and Keras) became available in 2015 onwards and have significantly lowered the barrier of entry.
3. Data has become less of a blocker for several reasons. Deep learning techniques can learn from mountains of data (available on the internet or collectible otherwise) without extensive feature engineering. Pretrained models are fit for specific tasks. Automation (data labeling, annotation, identification, error correction, and data format normalization) has eased the use of specialized datasets where required. 

You could think of AI as an ecosystem in which you don’t need to possess the abovementioned expertise. Focus on what matters. So, what does?

Focus areas when building a successful AI-enhanced solution

The challenges in building AI have evolved. It helps to think that your objective isn’t building AI. Traditional AI projects are tech-first explorations where required resources and outcomes are unknown. Instead, your goal is improving the solution, and you’ll use AI to the extent it’s feasible, viable, and desirable.

This brings us to the two areas where you need access to expertise.

1. Customer experience can be summarized in the words of Steve Jobs “You’ve got to start with the customer experience and work back toward the technology – not the other way around.” AI is no exception. You should use AI where necessary to address pains and bring gains but not change the world with a tech-first attitude.
2. Productization is much more than building features and often overlooked by the overly techy AI community. A solution must be designed, packaged, and delivered as a product or service. The solution must be desirable for the users, technically feasible stand-alone and as part of the systems and processes it integrates to, and financially viable. Finally, considerations like ethics, data privacy, transparency, and bias mitigation are hallmarks of successful solutions.

Using AI to enhance products has become accessible to the masses. Accessible doesn’t equal trivial. Reading Steve Jobs’s biography doesn’t make you a design expert. Having a product manager in your organization doesn’t mean she has the skills or time to build a killer product. You guessed right. A-CX can help.

FOLD (Fear of losing data) is a mighty blocker

We take data privacy, cyber security, and compliance seriously and have implemented secure backend solutions for highly demanding environments. Our clients regularly raise data-related concerns when exploring AI. This chapter introduces some common topics, explains how Chat based AI works, and presents a solution to the data concerns. 

Uploading data to an online AI model raises concerns:

1. Data privacy may be compromised, resulting in unauthorized access, breaches, or leaks. You don’t need to hold PII, PHI, or confidential data to consider this a valid concern.
2. A model may claim ownership or a broad license over the data it processes. Proprietary data may be leaked, and data protection regulations violated unintentionally.

“Use AI where it helps to address pains and bring gains but don’t change the world with a tech-first attitude.”

ChatGPT and related basic terminology

Users have cast their votes. Chat is the interaction method of choice with AI. Let’s quickly recap the terminology behind it. Alert! This gets nerdy for a while. OpenAI is the for-profit company behind ChatGPT. GPT stands for Generative Pre-trained Transformer, the” brain” behind the app. GPT is an example of LLM (large language model.) It’s trained with vast amounts of text to teach it to simulate human conversation. Notably, its dataset reaches early 2021 only and lacks any events and developments after that. Languages don’t only refer to natural languages. LLMs can process images, audio, speech, code, and many other patterns the human brain wouldn’t consider a language. GPT and similar models usually run online as they require significant computing capacity. As the models are online, their developers can collect user input to test and further improve their models.

Addressing the FOLD blocker

Imagine if you could run this magic machine safely in your isolated environment. Well, you can! The new Alpaca model released by Stanford is fine-tuned from Meta’s LLaMA and is at the time of writing cutting edge. We’re not talking about a local client with an API connection to the online model. We’re talking about running the entire model in isolation on your computer or cloud so the content never leaves the container. Once again, technology isn’t the challenging part here. It’s selecting the suitable model that fits your unique needs. For some use cases, the models come pre-trained. When a pre-trained model is complemented with your dataset, it’s called transfer learning. Contact us to discuss your project or share experiences with like-minded people!

Getting ahead of the curve

Have you filled in a form and gotten an error message ’All required fields must be filled’? A good UX used to mean the missing input field is highlighted with a red underline. What if the system wrote you in clear text what you missed and asked you to give input in natural language? How many websites have search functions that don’t find relevant matches? How often are the FAQ sections outdated? Could you automate customer care or have a chat function that is helpful? These are some examples of AI-enhanced use cases.

Do you still feel hesitant? For us, the decision-making chain looks pretty obvious. 

1. Someone will disrupt your business with the help of AI soon or sooner.
2. Overcoming the challenges in building AI isn’t primarily a technical task anymore. You don’t need deep pockets or access to superhuman brainpower.
3. You know your business. We know how to build winning products. Together we’ll nail the customer experience, address the pains and bring gains.

AI is the most significant transition internet has seen since the web browsers. If you don’t disrupt your business, don’t worry. Someone else will. On the other hand, many markets can be won by moving early and doing things right! The barrier of entry gets lower by the day, so time is of the essence. Let us know how we can help you get started.