Professional knowledge belongs in AI training

Painting contractor Tony Severino found a problem when Claude helped him prepare bids: the model estimated wall area by multiplying floor area rather than using the walls' actual dimensions. Anthropic recounts the example in its report on meetings with American business owners. Spotting the mistake took trade knowledge. Better wording in the chat box was only part of the answer.

This week's thesis: AI training needs more professional knowledge and fewer standalone prompting tricks. As vendors build more of the execution into their products, defining the task and explaining the result become more important. If that thesis is right, a course should be judged by the work participants can do afterward, not by the instructions they have collected.

This is Hammer Automation's interpretation of its research for September 7–13, 2026, not a measured finding from an education study. The English podcast covers OpenAI, Anthropic Claude, Google Gemini/DeepMind, Grok, and Mistral. Here, we look more closely at what the week means for learning at work.

Business owners wanted to build during the training

Anthropic's September 10 report offers a concrete reason to discuss course content. Across 635 evaluations after its business events, guided building sessions were the top highlight. Nearly two-thirds of participants wanted practical implementation support.

Those signals come from people who had already chosen to attend American AI events. They do not tell us how every Swedish business wants to learn, nor prove that a particular teaching method produces better results. But they point to demand that a lecture full of impressive demonstrations can miss: help me use this for my own work.

Severino's example makes the distinction clear. Someone learning to prepare bids with AI needs to understand which surfaces require painting. That knowledge can be taught alongside supplying Claude with information, asking follow-up questions, and correcting an assumption.

Source: Anthropic: What 1,000 small business owners taught us about AI

OpenAI makes the analysis tool easier to reach, not the business question easier

On September 10, OpenAI introduced a Data agent in ChatGPT Work. It can connect to approved company data sources, investigate changes, and create interactive reporting views. It also uses the organization's definitions of metrics, calculations, and relationships between data.

That last point matters for training. Asking why sales fell is easy. Knowing whether sales means booked deals, invoiced amounts, or paid invoices requires an understanding of the business. Shared definitions can help the tool, but someone still needs to explain the question and the conclusion.

The same day, OpenAI launched ChatGPT for Financial Services for eligible financial institutions, with selected built-in datasets and support for company document templates. That makes a course devoted only to producing an attractive presentation less interesting than one where participants can also explain the analysis's assumptions. This is an argument about course content, not a claim that finance professions have been automated away.

Source: OpenAI: Now everyone can put data to work

Source: OpenAI: Introducing ChatGPT for Financial Services

Google, Grok, and Mistral broaden the material for practical learning

Google's weekly research describes Gemini creating documents, spreadsheets, and presentations across Workspace apps. The feature was announced on September 9, but rollout had already begun on September 2, with English-only support initially. A Swedish educator therefore needs to distinguish what a product demonstrates from what participants can actually access.

On September 9, Grok gained outgoing message drafts inside the conversation, for approval before sending. That opens up a possible customer-communication lesson: participants can work on tone, commitments, and the recipient's needs in the same setting where they use the tool. We propose this as a course activity; we have not conducted that training experiment.

Mistral's September 7 report covers document extraction and local implementation partners. Its product news is background from the preceding week, not a set of fresh launches throughout September 7–13. Document extraction means retrieving text and structured information from documents. For finance administration, for example, that provides material to practice on without making the chat itself the main subject.

Beginners still need to learn the tool

The counterargument matters: professional knowledge offers little help if participants cannot access the service, add the right material, or ask a clear question. More features can make the basics more demanding, not less. Prompts are the instructions we give AI, and clear instructions still matter.

The proposal is therefore to tie those basics to real work. Have a participant draft a bid or explain a change in a report, using material permitted in the service. Then ask them to show which assumptions supported the result. That gives the educator something more concrete to discuss than whether the answer sounded professional.

For anyone planning staff development, the useful question is: what work should participants be able to perform and explain after the course? That is a good starting point for a conversation with Hammer Automation about Skill Forge and training for your organization.

The podcast is an AI-generated masterclass, created with NotebookLM from Hammer's in-depth daily research into AI providers' updates and features. The article draws on the weekly research and NotebookLM synthesis, not an audio transcript.