Author: Nick Scheffers
Date of publishing: 1 June 2026
It seems like every few days there is another news story about AI going wrong.
Really wrong.
Whether it’s a chatbot reminiscing to customers about its mother or the accidental leakage of sensitive data, these stories do more than just warn us of the hazards of bad AI implementation.
They also help to reassure us that human beings are, in fact, irreplaceable.
You cannot replace the knowledge or expertise of an experienced retail professional with AI. But it’s false to imagine that AI can’t generate incredible value in the hands of a seasoned merchandiser.
The key is to use the right technology in the right way.
In this newsletter, I want to give you a ‘peek behind the curtain’ by explaining exactly how our technology creates highly accurate (and reliable) inventory decisions.
Three layers of AI-driven inventory control
Thunderstock’s technology is a sophisticated system that is constantly fine-tuned to deliver the best performance for each individual retailer.
It consists of multiple technologies that work together to generate accurate insights and real-time stock decisions.
This is a complex system, but it can be divided into 3 distinct layers:
#1: Data layer
Retailer data is imported into the system using the most convenient method for them. Many retailers use bulk data import (e.g. CSV or XML), but data can also be synced using APIs.
Imported data includes both ‘universal’ types (stock levels, transactions, product information, product images, store information, and pricing) as well as optional, retailer-specific data.
This data is then converted into an internal data format, which is standardized and validated as it is imported, making it ready for ML.
#2: Intelligence Layer
The Intelligence Layer is like a workshop for Machine Learning algorithms. It’s a retailer-specific data lake where the extracted data is analyzed for historical trends and converted into insights and predictions.
Unlike many other solutions, our ML algorithms are trained exclusively on each retailer’s own data, instead of aggregating trends across multiple retailers. This means that predictions are specific (and accurate) for each retailer.
It also uses (optional) external data to inform better predictions with real-world contexts (weather, season, known trends, etc.).
With this solid foundation, ML algorithms operating in the Intelligence Layer can figure out potential sales for each SKU at each store, and across multiple timescales:
- Short-term forecasts for allocation, replenishment, and transfers.
- Long-term forecasts for reorders and assortment planning.
#3: Business Rules Layer
The Business Rules Layer gives each retailer granular control over how forecasts are interpreted. Using fine-grained business rules, you can create a realistic model of your own operational constraints, processes, and systems.
For example, you can set your own buffer stock levels, define maximum handling capacity, and include constraints like MOQs. Also, you can apply these across defined product groups, regions, or other parameters.
Depending on your tech stack and setup, inventory decisions are then sent to the ERP or used to generate pick tickets directly.
What about data quality?
You need high-quality data if you want ML to generate accurate forecasts and decisions.
In other words: rubbish in, rubbish out.
This is why we use other methods to validate and fix data, before it has a chance to cause problems.
This all happens behind the scenes, but it all helps make sure that each prediction and decision accurately reflects reality:
Import validation vets and validates all retailer-added data as it comes in.
Internal data checks are regularly performed by our dedicated team to detect data drift and accuracy.
AI feature safeguards ensure data quality and prevent confusion. For example, the Image model compares supplied product information with pictures of the product itself, to ensure accurate product content mapping.
Continuous ML validation is performed by a dedicated team. Their job is to continuously safeguard data quality and validate the ML algorithms.
Retailer-specific ML training is used to train Machine Learning algorithms for each retailer.
Regular ML retraining is performed every six months, or more often if needed.
In retail, context is everything
Consumers always buy based on specific contexts. These are something that an experienced merchandiser can generally pick out and use to optimize selling results.
But this task can only be done for a handful of key SKUs.
It is simply impossible to execute the same level of control across thousands (or hundred thousands) of SKUs with manual methods. There are not enough hours in the day.
This is why it makes good sense to use ML to achieve optimized sales at scale.
The goal is not to take over human jobs with an AI tool. Instead, Thunderstock exists solely to help retailers gain better control across their entire inventory.
By using the best ML technologies available, retailers can make decisions on the macro level instead of getting lost in the details.
This saves a lot of manual effort. It means that retailers can use simple rules to drive complex, granular decisions across numerous SKUs and stores.
No more trawling through data; just clearer insights, known contexts, and better inventory decisions.
And better inventory decisions means a better, more profitable retail business. Which is what we all want!
Feel free to send me a message if you have any questions or thoughts.
All the best,
Nick.
Quote: Start time travelling!
