Pre-trained language models (foundation models) understand text, tables, documents and images, and can process very large amounts of data in one go. Instead of training a separate model for every task, you describe the task in plain language. The model analyses, summarises, classifies or writes code.
As of October 2026
AI in 2026: From Forecasting Models to AI Agents
The last time we wrote about artificial intelligence here, the topic was Python libraries such as TensorFlow, Keras and scikit-learn. You used them to build and train models yourself. These tools still exist. But the way companies use AI has changed fundamentally since then.
Large language models as general-purpose tools
Reasoning models
Newer models “think” in several steps before answering. They break problems down, check intermediate results and correct themselves. This noticeably improves results wherever accuracy matters: calculations, data analysis and programming.
AI agents
The biggest step: AI no longer just answers, it acts. Agents use tools. They query databases, run code, read and write files and operate web applications. This lets them complete multi-step tasks largely on their own. Open standards such as the Model Context Protocol (MCP) connect agents securely to a company's systems.
Foundation models for time series
Pre-trained models now exist for forecasting too, such as Chronos (Amazon) and TimesFM (Google). Without extensive training, they deliver useful predictions for sales, inventory or freight rates. In practice, we combine them with classical statistical methods and domain knowledge. That way they clearly outperform gut feeling.
Knowledge from your own data (RAG)
With retrieval-augmented generation, a model looks up a company's own documents, contracts or databases before answering. Its answers are then based on current, verifiable information rather than general training knowledge. This is the prerequisite for reliable assistants in day-to-day business.
In practice: this website
An example from our own work: to move this website from WordPress to our own server, we used Claude, the AI model by Anthropic, as an agent with access to the project folder and the command line.
Claude read the existing site in both languages and analysed its structure. From that it built a lean static website without a content management system, including a build script, server configuration and redirects for all old addresses. Along the way it found and fixed errors in the old site, such as an endless loop in the links. The first draft of this article was also written by Claude.
We made the decisions: what stays, what goes and what the server looks like. That is exactly how we see the sensible use of AI: it takes over the legwork, while people stay in control of goals and results.
What hasn't changed
For all the progress, the fundamentals of our methodology still apply. It starts with a clearly defined goal, then the right data. Results have to be measurable. A language model does not replace a clean data foundation. But it makes that foundation usable faster than ever before.