Beyond thin AI wrappers: Stop Generating Text, Start Modeling Reality

While everyone chases the latest prompt engineering trick, a quiet revolution is happening in the businesses that treat AI as an engineering discipline, not a content tool. Real competitive advantage doesn’t come from generating text faster, it comes from modeling the complex systems that drive your business and optimizing the decisions that matter. Learn why time series forecasting, graph neural networks, and reinforcement learning deliver ROI that generative AI can’t touch, and why the data foundation you build today determines whether these techniques will work for you tomorrow. This article is the second part in our three-part series on AI and machine learning. Find part 1 here Beyond thin AI wrappers: Why AI engineering wins.

The Untapped ROI of Deep Learning

When we talk about AI right now, we are almost exclusively talking about Generative AI; creating text, images, and code. While valuable, this focus on creation ignores the most profitable application of artificial intelligence: optimization.

This is where that rigorous data foundation pays off. Once you have clean, structured, contextual data (the kind we described in Part 1)  you unlock an entirely different class of AI capabilities. Your business cannot be boiled down to a text prompt. It’s a complex web of relationships (who buys what) moving through time (seasonal trends). If you try to jam that complexity into a message passed to an LLM, you lose the signal.

To solve high-value problems, you need to invest in Decision Intelligence – tools and techniques that model reality as it actually is. We will explore three powerful approaches:

• Time series foundation models to forecast what is likely to happen next.
• Graph neural networks to understand and detect patterns in connected systems.
• Reinforcement learning and causal AI to choose actions and measure what truly drives outcomes.

Generative models are not always going to be the right tool for the job.

Time series forecasting: Foundation models for zero-shot prediction

Forecasting used to be a heavy lift. If you wanted to predict sales for 10 000 different products, you effectively had to train 10 000 little statistical models. It was slow, brittle, and expensive.

We are now seeing the rise of Time Series Foundation Models. Just like ChatGPT reads the whole internet to understand language, these models have read billions of data points – things like stock prices, weather patterns, energy usage and retail sales to understand time.

The result? You can now perform “Zero-Shot Forecasting.” You can show these models a brand new sales chart they’ve never seen before, and they can predict the next three months with startling accuracy, often beating the custom models you spent months building. This democratizes enterprise-grade logistics for companies that don’t have massive data science teams. If you have the necessary data to fine tune them to your business context, then you’re in a really great position.

Graph Theory: Connecting the Dots

Most companies store their data in tables, rows and columns. But the real world is always some kind of network. Customers are connected to households; credit cards are connected to devices; suppliers are connected to shipping routes and weather disrupts flight schedules.

Graph Neural Networks (GNNs) allow us to treat your data like a network, not a spreadsheet.

Instead of treating a fraudulent transaction as an isolated data point, a GNN sees it as part of a suspicious subgraph; a “fraud ring” where multiple accounts share a device ID or shipping address.

Reinforcement Learning & Causal AI

While Time Series models predict what might happen, Reinforcement Learning (RL) helps you determine what to do about it.

Reinforcement Learning (RL) is about learning through trial and error. Instead of guessing how much stock to order, an RL agent inside a digital twin of your supply chain can play through millions of scenarios to discover exactly which ordering strategy prevents stockouts without bloating inventory. RL agents can discover counter-intuitive strategies, like pre-positioning inventory in a specific hub before a weather event occurs, effectively solving the trade-off between holding costs and shortage risks.

Causal AI is about finding the truth. In marketing, we often target people who are going to buy anyway (“Sure Things”). Causal AI tells you who only bought it because they saw an ad. It helps you stop wasting money and focus purely on the people who can be persuaded.

This is where the real ROI lives. It’s not in writing faster emails; it’s in mathematically optimizing the decisions that drive your bottom line. But these sophisticated techniques only work when built on top of the engineering discipline we outlined in Part 1: Beyond thin AI wrappers: Why AI engineering wins. Without that foundation, you’re just building expensive experiments that never make it to production.

About the author

Paul dos Santos is an AI Engineering expert in the Insights & Analytics team at Curamando, with a background in applied AI from foundation models to AI at the edge.

Discover our AI case studies

Read more about how we helped a contractor save 1,200 hours annually with an AI chatbot

Read more about combining human expertise with AI speed in content creation

Read more about how KLM streamlined passenger transfers with conversational AI kiosks