Understanding that AI is often not able to provide the instant solution you were looking for is a tough lesson to learn. There are, however, key steps that you can take to improve AI outcomes for your business.
Both working with and witnessing software vendors as they clamour to add AI to their product, it has become very apparent to me that there is a significant gap in the understanding of customers and seeing beyond the AI ‘shininess’ that is being presented to them.
The main reason for my thinking here is that ultimately, AI is only as good or efficient as the data that it has access to, or it is trained on. This is one of the key reasons for the disparity in capabilities in generative AI tools, as the quality of their training data is a key limiting factor. It leads to AI having the hallucinations and inconsistencies that we all read about or have experienced firsthand.
What should you consider when buying a tool with AI capabilities?
In addition to the accuracy and real intelligence that the tool may have, it is vital for buyers to consider not just the outcome they are looking for, but the task complexity that they are asking the AI to take on.
Imagine a huge plate piled high with spaghetti. You might be able to see a few ends (or beginnings) poking out from the pile in front of you. In some rare cases you may be able to see both the beginning and end of a single piece. Now imagine that the spaghetti represents data and start and end points in a query that you are looking to interrogate and get answers to. Finding the right piece in all that spaghetti mess is hard enough. Cutting through the noise, interpreting the start from the end and then how it related to all the other pieces is a seemingly massive task… because it is.
Expecting to apply AI to huge scale tasks and have it take care of things is where problems begin and projects start to fail or become in reality very lacklustre.
There are in fact, multiple reasons for AI struggling to interpret data:
- Data quality: If the data is noisy, incomplete, or biased, the AI model may produce inaccurate or unreliable results. High-quality, well-labelled data is crucial for effective AI performance.
- Complexity of context: Understanding human language involves grasping context, nuances, and cultural references, which can be challenging for AI. Ambiguities and subtleties in language can lead to misinterpretations.
- Training limitations: AI models are trained on specific datasets. If the training data doesn’t cover certain scenarios or topics, the model may struggle with those areas. This is known as the problem of generalization.
- Evolving language: Language constantly evolves with new slang, idioms, and expressions. AI models need regular updates to keep up with these changes, or they risk becoming outdated.
- Complex queries: Some questions or tasks require deep understanding and reasoning, which current AI models might not fully achieve. Complex, multi-step problems can be particularly challenging.
- Technical limitations: Despite advancements, AI models still have limitations in processing power and algorithms, which can affect their ability to interpret data accurately.
All of these reasons go some way to explain why the best technology implementations start by looking at the way in which organisations are working with data rather than just applying a product and expecting it to work.
Back to the spaghetti
So how do companies make their AI teammate work smarter and achieve better results? There are two key factors to consider here.
One is to reduce the amount of data (spaghetti) that the AI is being asked to interpret. This can be done through using tools that remove duplication and only retain the data that is meaningful or required for future interrogation.
The second is to utilise a tool that improves the way in which your data is organised and stored, reducing complexity and lack of completeness. If you imagine that pile of spaghetti laid out in rows rather than in a pile, it is easier to find each strand and to identify the start and end points – much like it arrived in the packet!
A better outcome for all
This means that there are also two approaches to improve the situation for the customer. One is to build in data consolidation and performance enhancements that work in tandem with the AI. The other is for customers to take the initiative themselves and apply such technology to improve their data and processes before buying an AI solution.
The second approach means that the customer has more options in selecting their AI solution as they are well prepared to get the best out of what it can offer. Importantly, using either of these approaches will mean that the project is far more likely to succeed and to deliver the desired results.
