Skip to content 
As data practitioners, we often get the question of how to use AI to extract meaningful insights from the vast volumes of unstructured customer feedback that companies collect.
Recently, we built a demo of such a solution, leveraging the official Customer Reviews Analytics Quickstart and Snowflake Cortex for customer reviews analytics.
This article gives you a step-by-step guide to the implementation from that demo. It also elaborates on the practical value that Cortex provides, the cases where this is the most powerful approach, and when you might want to explore other routes.
The problem with customer reviews
Customer reviews can be a goldmine for operational improvements, product insights, and customer experience initiatives. But there are some recurring challenges that make it difficult to harness the full potential:
- Reviews are written in multiple languages;
- Most of the data is unstructured and hard to analyse at scale;
- Extracting sentiment and categorising feedback often requires dedicated ML teams;
- Even with these insights, operationalising the insights (e.g. triaging issues or auto-generating responses) is an added layer of complexity.
Snowflake Cortex can be key in handling this end-to-end workflow directly within the Snowflake ecosystem and without moving data to external services or building custom ML pipelines.
Cortex-Powered NLP with PySpark in Snowflake
Here are the key implementations we showcased using Snowflake Cortex Python UDFs via PySpark:
Translating multilingual reviews
We used the CORTEX.TRANSLATE() function to convert multilingual reviews into English:
Value
This implementation made it seamless to normalise reviews into a common language for downstream analysis. This capability is especially useful in global businesses, as it removes the need for third-party services.
Limitations
While good for general translation, nuances such as sarcasm, regional slang, abbreviations, or industry-specific terms can be challenging and mistranslations could potentially distort the original sentiment or the text.
Categorising unstructured review text
We used the CORTEX.COMPLETE() function with prompt-based classification to tag reviews into a spectrum.
Value
Using one-shot learning in the prompt yielded reasonably accurate classifications without training a model, which would be ideal for small to mid-sized datasets.
Limitations
Some edge cases, such as mixed or sarcastic reviews, can be misrated, and the approach depends heavily on well-crafted prompts.
Understanding customer sentiment
We applied the CORTEX.SENTIMENT() function to assign a sentiment score and a label (positive, negative, neutral) to each review.
Value
For strategic monitoring, you can aggregate rapid sentiment tagging by product, region, or time.
Limitations
Sentiment output is scalar (positive/negative/neutral) and may oversimplify complex or nuanced feedback. Again, some borderline or sarcastic statements may be misclassified due to model generalisation.
Diving deeper with aspect-based sentiment
Sentiment analysis is taken a step further by identifying sentiment tied to specific product aspect, such as, food quality, service, pricing, hygiene, etc.
Value
This capability provides fine-grained insight into what exactly is going well or poorly. It is a powerful tool for targeted product or service improvement.
Limitation
The JSON structure is sensitive to prompt design. Invalid or inconsistent JSON output can complicate parsing or downstream automation. Additionally, the prompt-based approach can miss subtle aspect expressions or infer unrelated ones if phrasing is ambiguous.
Identifying the issues
We enabled a quick analysis of the most negative reviews for the most negatively reviewed item to understand the main issues that the customers complain about by leveraging CORTEX.COMPLETE().
Value
This way, we turned the noisy review data into focused operational insights, ideal for prioritsing fixes and improving satisfaction.
Limitations
The summarisation can occasionally miss recurring but subtle themes. Also, performance may drop if review volume exceeds token limits.
Generating email responses
Last but not least, we used GenAI to auto-generated sample customer support responses based on review content.
Value
This is a great starting point for internal or external communication, as it saves time and ensures consistency in tone and format.
Limitations
Generated emails may still require human revision for accuracy, tone, or business-specific language, especially for legal, escalated, or highly specific cases.
Using AI to make the most of customer reviews
From a data engineering point of view, the appeal of Snowflake Cortex lies in:
- Unified platform: All computation happens within Snowflake, reducing ETL overhead and risk;
- Python and SQL support: Whether teams work in Python or SQL, Cortex adapts to common workflows;
- Prompt-based flexibility: Many NLP use cases can be addressed quickly without training custom models;
- Scalability: Cortex can scale with the data, handling spikes in usage behind the scenes.
Where caution is advised
Cortex is a valuable tool, but it is worth noting that:
- Prompt tuning is a skill: The success of many functions hinges on well-crafted prompts. These may need iteration or manual tuning per domain;
- No fine-tuning or domain training: Cortex models are general-purpose. Domain-specific accuracy might lag;
Output validation is crucial: You will often need to validate and clean outputs, especially when they feed into customer-facing or production systems; - Cost consideration: Frequent or large-scale calls to LLMs can drive up compute costs;
- Explainability: Often, it is unclear why the model has made a certain classification or output, which can be a concern in heavily regulated industries.
Final word on using Snowflake Cortex for customer reviews analytics
Snowflake Cortex enables organisations to tap into the value of unstructured customer reviews using familiar tools like PySpark and SQL, with no MLOps burden. It is well-suited for use cases like translation, sentiment analysis, categorisation, and communication automation.
While this approach does not replace fine-tuned, domain-specific ML pipelines, it does offer a strong return on speed and flexibility, particularly for teams already working in Snowflake.
Visit the Infinite Lambda blog for more insights into leveraging AI at enterprise level.
