Some business people have recently passed the peak of their inflated expectations regarding generative AI. Having had their first experience implementing AI-based services and failing, they have come out of their cloud 9 and feel generative AI isn’t able to produce what they expected it to do.
However, it’s important to remember that generative AI is still evolving and has yet to reach its full potential. With new, more powerful models being launched almost weekly, and smaller models becoming more capable, the future of generative AI is bright, expanding its reach to laptops and mobile devices.
Blame it on marketing.
So what’s happening? I blame it on the marketing departments of AI companies, who are just too good at promoting the incredible things their services offer (to consumers) and forget to mention the pitfalls for businesses.
Basic LLMs are not business applications.
Too many people still think ‘just’ using a basic LLM service like ChatGPT will be sufficient for their use case. Business-driven use cases often require traceability or a guaranteed certainty that the generated answer is accurate. And yep, that differs from what basic LLM setups are created for or capable of.
Using AI in your business is hard expert work.
There are no free rides when implementing an enterprise-grade AI solution. You might have heard of RAG or Retrieval Augmented Generation and wondered what all the ‘fuzz’ is about. RAG is part of a more extensive set of technologies capable of making LLM-based solutions fit into your business use case. Read up on them, and you will discover that there is a lot you can do to improve the outcome of generative AI. However, implementing this requires expert knowledge to create the appropriate infrastructure, and there is a lot of trial and error. This underscores the need for professional guidance in implementing generative AI.
Want to better understand what you can do to improve a LLM based AI setup? Look at the articles in the Enterprise AI section of my blog.
Just throw AI at it
Another cause of this Peak of Inflated Expectations is the ‘Just throw (Generative) AI at it, and see if it sticks’ strategy. This approach, if we can really call it a strategy, involves haphazardly applying AI technologies without a clear understanding of the problem or the technology’s capabilities. When this approach inevitably fails, it can lead to disillusionment with AI’s potential. This will continue as companies think it’s a competitive advantage to say that they use AI, no matter if it makes sense.
Throw at it what it needs.
Instead of adopting a ‘throw at it and see if it sticks’ approach, businesses should take a more strategic and critical approach to solving their automation problems. This involves using the right tool for the right problems, empowering you to make informed decisions about AI implementation.
Just use business logic.
The first question that needs answering is:
- Can we solve this with business logic?
Business logic provides stability, control, and traceability with predefined rules and logic. If this suits your use case, then there is no need for AI. Business logic is the way to go if you can determine the rules, handle the manual updates when rules change, and determine when the rules change. (You could let AI detect when the rules change!)
Business logic can be effectively used in rule-based decision-making, process automation, and data validation scenarios.
Machine Learning the ‘easy’ AI.
Next, when business logic doesn’t fit, you ask what kind of AI would fit best: machine learning or an LLM (Generative AI)?
First, check whether the most basic one of the two fits: Machine learning. Why start with machine learning?
Because:
- It offers the most control; you pick the algorithm and can set thresholds,
- It can also work with smaller datasets,
- And in general, it is easier to implement.
Use cases for ML are classification, regression, and recommendation systems.
Enterprise-grade Generative AI
If that doesn’t work, you are up for the extensive work: generative AI. Generative AI, while powerful, comes with its own set of challenges. These include the need for large amounts of data, potential biases in the training data, and the complexity of the models. Overcoming these challenges requires expertise and knowledge to ensure stability, control, and traceability in your AI system. Also, remember the tools mentioned to ensure the quality and trustworthiness of the outcome of the LLM.
It doesn’t have to be one technology.
But you can also be smart and combine different technologies. For example, take an incoming email to which you want a standardised reply that includes information from the original email.
You could:
- Train an ML model on labelled email data.
- Extract features (e.g., subject, sender, keywords) and use them to predict the category (spam, urgent, etc.).
- Automatically route emails to the appropriate folders or teams based on ML predictions.
And reply by:
- Using a standard email template (greeting, signature, etc.).
- Augment the template with personalised content from the original email using an LLM.
- Add relevant information (e.g., customer name, issue details) into the reply.
- Ensure the tone and style match the original email.
It’s just one example of how you can combine different techniques with strengths and weaknesses into a solution that offers the traceability, certainty, and accuracy you need as a business. This approach addresses the limitations of individual technologies and opens up possibilities for innovation and efficiency.
Please don’t use the ‘throw at it and see if it sticks’ method, but take a serious look at all the options and think outside the box. You may need more than one technology to get what you need.
Greetings, and good luck in your automation/ AI journey.