Data Analytics and LLMs

The rise of general-purpose AI

In the past few years, we have lived through what many call an AI revolution, and most people would associate this moment with the rise of ChatGPT. Was it so outstanding compared to all the other technologies? Maybe. But maybe it was mainly because of how seamless it was — it brought the already existing technology of LLMs to a much broader, non-technical audience. Just a simple chat interface with some additional features, and suddenly they could just ask a question and get an answer.

For any technology to be adopted, several barriers need to be overcome, the most important being:

• Cost — An obvious one. Not many people want to invest money in something they are not sure about, especially with non-deterministic technologies such as LLMs, where in many cases you can’t be completely sure it is good enough for your use case. Costs are dependent on all the other variables and barriers, which makes them usually a focal point of every discussion. Budgets ultimately decide what can happen.

• Complexity — When it takes too much effort to set something up before you can try it, you had better be sure it is worth it. If complexity is too high even for a proof of concept, a real use case will not even get started.

• Reliability/Consistency — Let’s forget the “if it works 60% of the time, is it even a solution to the problem?” discussion. For each solution, we should have some idea of how reliable the outcomes are. Are they ambiguous? What if the input itself is ambiguous? There are certainly many questions to take into consideration, so let’s dive a little deeper. Without some understanding of reliability, it becomes difficult to operationalize AI beyond experimentation.

Have you ever wanted to ask questions about your data?

AI and non-determinism in analytics

With modern AI, the door has opened to non-deterministic outputs produced without direct human involvement. Data professionals traditionally operate on the opposite end of the spectrum — deterministic methods, clearly defined and tested outcomes, and ideally at least a rough idea of ROI for everything we do.

When working with data, a significant part of our work is to minimize ambiguity — mainly in data collection or cleaning. However, you do not always have exact data and predefined values. Sometimes all you have is natural language, which is highly contextual and ambiguous.

As the complexity of natural language is very high in itself, the complexity of the methods required to process it also grows significantly. Lastly, with more complex methods and input data, we leave the realm of traditional analytics methods that rely on crisp, deterministic logic and dive into a (controlled) fuzziness, where uncertainty is accepted but bounded and measured to handle the ambiguity of natural language.

Now, let’s take a step back and focus on business. We should never forget ROI — “The more complex a solution is, the more value it needs to deliver.” Natural language processing has always been one of the data science domains requiring specific skills and tools compared to traditional analytics or business intelligence. More skills and more tools inevitably increase costs. So in short, you needed a good reason to even consider processing your textual data.

Do you collect data that you are currently unable to use in a meaningful way?

Pains of advanced analytics

Suppose you are a company that has a data team of a few people consisting mainly of data analysts. You have a low-code ETL tool, a data warehouse, and a reporting tool. You have the basic reporting needs covered, but now you want to apply NLP because you need to start processing customer reviews at scale.

Historically, this would usually require a data scientist. A data scientist can do most things using Python, but you do not have one, so you need a new environment for running Python. Besides having to manage this environment, data leaves the data warehouse and is processed elsewhere. Complexity also increases around data governance and security. You obviously need to run this regularly, which adds another increase in complexity because you need a scheduler.

This is one end of the spectrum. Let’s have a look at a better situation. Let’s say you have a person willing to tackle this problem, because hiring a new person could hardly be justified. Python can be run locally, and you need the analysis once a month, so you just have your colleague run it on their local laptop once a month. This way, the complexity is minimized, and they just need to set up their local environment, choose and learn the appropriate NLP technique, and learn some Python. It feels like an MVP, but it solves the case and the ROI is likely better. However, this approach is fragile and heavily dependent on one person and one machine.

It gets even better with ChatGPT and general-use LLMs — you can likely just send the Excel file there and have it processed. This lowers the skill and setup barrier but introduces serious data governance and compliance concerns, unless you have access to a secured LLM that does not use your data for training.

Now let’s try to imagine an even better scenario — you have Snowflake as the DWH.

Snowflake allows you to use most currently available LLMs with your own data inside Snowflake.

• Snowflake security and governance → No data leaves Snowflake
• Both structured and unstructured data can be stored and used
• No new tools are required, only an increase in credit consumption
• Easy to use (AISQL) — I cannot stress this enough here, but you will learn more in upcoming articles about Snowflake’s AISQL
• Costs are kept to usage only, depending on token consumption
• Complexity does not increase too much
• Reliability/Consistency is the biggest question in this case, and experiments should generally be conducted and evaluated for most LLM use cases.

In the end, the real shift is not simply that LLMs like ChatGPT exist. It’s that platforms such as Snowflake are embedding them directly into the analytics stack — where data already lives. That changes the equation. Data teams can begin treating natural language as just another analyzable data type, expanding their capabilities without introducing an entirely new layer of infrastructure or complexity.
The next step is practical application. In the following article, we’ll move from concepts to implementation and examine Snowflake’s AISQL capabilities in action.

Jan Šulc

Business Analyst at Revolt BI with hands-on experience in Snowflake and modern data platforms. 

---

He designs scalable data models, automates data pipelines, and works on applying AI capabilities within enterprise analytics environments. Jan bridges business needs with technical architecture to help organizations operationalize data and AI.