Can I Offer You Corrupted Data Gruel?

Imagine walking into a new restaurant. A digital AI chef in a spotless apron approaches you, filled with confidence and a convincing smile. He radiates such professionalism that you trust him immediately – you think to yourself that you might have found a hidden Michelin-starred spot that no one knows about yet.

He places a plate in front of you and says with great passion: "Here you go, this is exactly the information you ordered, bon appétit."

The presentation looks flawless, almost like a work of art. But as soon as you examine the plate more closely, you know something is not quite right. The ingredients seem completely out of context with the order, and it is clear that no recipe was used. When you then taste it, your suspicion is confirmed – the flavor is distorted.

Indeed, what it takes for the chef to serve good food varies. In fact, it all depends on his access to quality ingredients and clear recipes – or in other words: context and understanding of the data he works with.

This is exactly what happens when AI is unleashed on your data without context. It is like this smug chef: it is nothing but confidence and always sounds like it knows what it is doing. It can generate answers that look professional and are presented with great conviction, but without guidance, the content is nothing but beautifully presented digital data mush.

Data is raw material - not a finished meal

We regularly hear that data is valuable and crucial, but owning a lot of it achieves nothing on its own. In traditional business, all kinds of data accumulate here and there. We store them in databases, documents, and the various systems we use day-to-day. We do our best to manage and categorize them to make them accessible later.

However, that is like having a lot of raw ingredients for cooking; they do not make a good meal unless they are prepared with expertise. No matter how fancy the databases or archives are, they often lack this context to be able to cook up good information. Data storage is, after all, primarily a technical problem, but giving it meaning and creating context is what turns the raw material into real value.

Recipe for success: Infrastructure that supports decision-making

If we want to ensure our AI chef lives up to expectations, we need to look at technology as the tools in the kitchen. Many new and unfamiliar words accompany the AI wave, but fundamentally they are all about the same thing: ensuring the chef understands the order, finds the right ingredient, and uses his knowledge to prepare good food.

But a good restaurant is of little use if there are no customers. It is, after all, the guests who judge the quality of the food. They have expectations, and they are the ones who decide whether the order was delivered correctly to the plate. If the guest does not get what they ordered, it does not matter how fancy the presentation is.

In order to make the guest experience as good as possible, we need to introduce some new tools to the kitchen's organization:

• RAG (The chef's lookup book): Think of RAG (Retrieval-Augmented Generation) as the difference between two chefs:

  • Chef without RAG: He learned everything in school five years ago. If you ask him: "How much salmon is left?" he is forced to guess, because he hasn't checked today's inventory status. He might say with great confidence: "There is plenty of salmon," even if it is sold out.

  • Chef with RAG: He has a tablet in his hand that is connected directly to the inventory system. When you ask him about the salmon, he starts by glancing at the screen, retrieves the results, and then answers: "We have exactly 4.2 kilograms of salmon".

• System Prompts: Think of these as the chef's established operating procedures. They tell him exactly how to behave towards guests and how to interpret orders so that the experience is always the same.

Example: "We never use cilantro in this dish and all fish must be served with lemon." This ensures the chef always follows the restaurant's rules and doesn't start spinning his own rules every time someone orders food.

• Data Models and Organization (The Pantry itself): For the chef to work systematically, access to ingredients must be good. If access is time-consuming and labeling is incorrect, it can significantly affect efficiency and increase the likelihood of distortion. The technicians working with the AI can do a lot to pave the way to the right ingredient, but sometimes data specialists simply need to be called in to make improvements.

• Pre-preparation of ingredients (Staging and ETL): Just like in fine restaurants, it is important to prepare ingredients before the order comes in. This is what we call ETL (Extract, Transform, Load). We retrieve data from core systems, such as enterprise resource planning (ERP) systems, transfer it to a "staging" layer, and clean it there.

  • Example: This is like washing the sand off the lettuce and peeling the potatoes in advance. This way, the ingredient is ready for the pot when the chef needs it, which reduces the load on core systems and ensures the AI works with correctly handled data.

• Organized ingredient storage (The Warehouse): Instead of the chef having to search through a huge and chaotic warehouse where items are stored based on when they arrived, data specialists reorganize the storage into a data warehouse. A query layer is created there, modeled after the menu (the company's business model).

  • Example: We arrange the ingredients on shelves (dimensions). One shelf is for meat products, another for vegetables, and the third for dairy products. Instead of searching through thousands of boxes for tomatoes, the chef knows exactly which shelf to go to to access the right ingredient.

• Semantic Layer: This is where technology, business intelligence, and context meet. This layer ensures that everyone understands terms in the same way, regardless of how the question is phrased.

  • Example: Instead of the chef needing to know a complex calculation formula, we create a common context such as "Sales excluding VAT" or "Quantity sold". The AI simply asks for "Sales" and always gets the number that has been predetermined, e.g., "Sales excluding VAT".

• The Warehouse Manager at the Hatch (MCP and API): MCP (Model Context Protocol) is the latest technology in AI projects today when it comes to integrations. It works like a warehouse manager who is responsible for delivering what the chef needs through a service hatch (API). We put these service hatches on top of our data to make it more accessible, so the chef always gets the right ingredients based on his requests.

Example: Instead of the AI chef having to search everywhere himself for ingredients in the warehouse or other systems, he asks the warehouse manager (MCP) for the ingredient. The warehouse manager knows exactly where everything is, retrieves it, and hands it to the chef through the service hatch (API). This increases speed and security, in addition to creating context that prevents the chef from "winging it" with all kinds of ingredients he finds at any given moment (what is called hallucinations in AI).

• From Information to Execution (Agents): But the system can do more than just retrieve ingredients to answer queries. By adding Agents, the warehouse manager gains executive power. He can not only answer questions but also perform tasks autonomously, such as automatically ordering new ingredients when he sees that stock levels are running low.

Even though we are talking about many new concepts today, and even reviving old concepts that have taken on a new role and increased importance in the context of AI, one thing does not change: a thorough needs analysis is always a key prerequisite for success. Whether it involves RAG, MCP, data models, or traditional data processing, it is necessary to assess in each case which solutions are best suited. The choice of tools is not determined by what is newest or most prominent, but by the complexity of the project, its scope, and the decisions to be supported.

Is your restaurant ready to take orders?

As has been pointed out, it matters little how smart and convincing the "chef" is if he has neither clear recipes nor access to the right ingredients. Without context, definitions, and secure access to data, the experience will often be beautifully presented — but wrong.

AI does not solve problems that already exist in data. It simply makes them more visible. When it is given incorrect assumptions, it nevertheless answers with great confidence. There is a reason why a distorted data mush often looks better than it tastes.

The solution lies not in a single technical layer, but in the synergy of:

• AI specialists, who bring new tools and knowledge
  
  

• database and data specialists, who know the raw materials well
  
  

• the business side, which defines what a "correct answer" actually means
  
  

How do we open the kitchen?

A common misconception is that everything needs to be perfect before starting. The reality is different. Most companies can open the kitchen incrementally and purposefully, without reorganizing everything.

Often you can start with three simple steps:

Start where the need is real

Choose a use case that touches daily business decisions — for example, inventory levels, sales figures, or other key information. The goal is not to answer all questions, but those that really matter to people who need to get information quickly and reliably to make decisions.

Open a secure path to data

Instead of waiting for a perfect data model, you can open and have access control to the data that already exists. This can be through APIs, MCP, or other means.

The most important thing is not how the data is retrieved, but that both users and AI always have the same assumptions to work from. With such an approach, it is also possible to "stir the soup" before it is served: personally identifiable data is skimmed off, sensitive information is protected, and business secrets are prevented from accidentally becoming accessible. In this way, access is not only efficient, but also secure and in compliance with privacy and confidentiality regulations.

Bring operational knowledge along from day one

AI specialists and technical teams know how to connect to data and work with the tools. But it is the specialists in operations who know what the answer should mean and how it is actually used.
This knowledge resides with people who have worked with the processes for years — in finance, operations, sales, service, or IT. It does not necessarily reside in data models, but in experience, exceptions, and unwritten rules.

When these three steps are taken, the AI's answers will not only be technically correct, but also usable for operational purposes.

APRÓ supports your company on the data and AI journey

APRÓ is a leader in implementing the latest technology when it comes to data context and utilizing AI. We have one of the country's strongest teams spanning the entire process – from organizing the "ingredient pantry" to designing the "digital chef."

We employ specialists with excellent knowledge of:

• AI and Agents (AI Agents): We design and implement solutions like RAG and MCP to ensure that AI delivers the desired results,

• Databases and data processing: Our experienced data specialists and data engineers take care of cleaning, organizing, and modeling your data (ETL/ELT) so the AI can use it.

• Cloud solutions (AWS and Azure): We are specialized partners in both AWS and Azure. We leverage powerful infrastructure, whether it is Amazon Bedrock or Microsoft Foundry, to run AI projects securely and scalably in the cloud environment that suits you.

We help companies bridge the gap between complex technical solutions and business needs. Whether you need to bring order to large datasets or want to build your first "service hatch" in AWS or Azure, our team is ready to guide you through the process – from initial analysis to completed solution.

Are you ready to turn data mush into a great meal? Let's find the right recipe for success in your company together. Contact us and we will show you how you can utilize your data with the latest AI technology, regardless of which cloud solution you use.