Monthly thinking room
How much control do agents need so that companies can hand work over to them, and how does real impact show? This edition shows how pioneers build review, limits and measurement and what leaders can adopt from them.
33 conversations added · none removed
2 new · 3 confirmed · 5 sharpened · 2 dropped
With the new episodes, individual observations have become a pattern: NGM Group, ZIRNGIBL, T-Mobile, Microsoft Deutschland, HV Capital, Instagram and Herbalife independently describe how they explicitly ask AI for follow-up questions, disagreement and counterviews, and the AI leadership monitor of FOM Hochschule shows the second opinion to be a widespread use among leaders.
View finding →With the new episodes, the technical limitation of agents is broadly documented: Statista, GO at Deutsche Bahn and the hosts Marc and Jens set hard usage limits, NewStore and EnBW restrict permitted actions and data spaces, Scott Duncan separates read access from intervention in the plant, GitLab enforces controls on the server side, Stefan Brock (HPE) calls for one-click actions for stop and rollback, and Kevin Surace and Antonio Krüger (DFKI) explain why rules in the prompt are not enough for this.
View finding →Ten conversations from medicine, veterinary medicine, energy, fashion retail, software and product management describe the same setup: generation and review sit with separate instances, because the generating model does not reliably find its own errors.
View finding →The new episodes start one step earlier: Google DeepMind and “Der KI-Podcast” first check for each step whether a language model is needed at all. The separation of understanding and calculating is thus the result of this breakdown.
View finding →The new episodes show the same movement in the work of individual leaders: Weleda, sipgate, snipKI and 60 Kollektiv maintain a context base of project and Markdown files that the AI tool accesses. Shared skills are thus one instance of a broader pattern.
View finding →Ten conversations from software companies, platforms, consulting and investment, as well as from adidas and the pharmaceutical data provider Norstella, describe the same restructuring: where AI takes over execution work, companies merge specialist roles and shrink teams so that handoffs and coordination disappear.
View finding →The new episodes show the forced model change alongside the planned one: all of Uwe Franke's overnight jobs stopped when a model was switched off for non-US users, sipgate bridged a regionally restricted Claude version with a rented machine, and Parloa automatically switches to Codex when Claude Code has an outage.
View finding →Eight conversations from retail, software, consulting and training describe the same lever: leaders and employees show others what they themselves have built or achieved with AI, in all-hands meetings, show-and-tell calls, open Slack channels or lighthouse projects, instead of merely mandating its use.
View finding →Several episodes show the step before selection: Schwarz Digits classifies around 98 percent of the collected cases as not AI topics, Grupo Diesco puts 100 to 150 use cases on hold, and Drees & Sommer filters eight ideas from an ideathon down to three use cases. The pattern therefore now describes condensing many ideas into a few processes, not just calculating in time and money.
View finding →The new episodes show how companies make impact robust: Erik Brynjolfsson scales a productivity gain back from 41 to 11 percent after proper controls, Staffing Referrals relies on a holdout group, and Drees & Sommer launches first use cases only where the baseline is already being measured. The competency levels per role from the previous version are no longer in the headline, because among the new episodes only Mainsail Partners describes them.
View finding →No longer among the ten most broadly evidenced patterns in the revision.
No longer among the ten most broadly evidenced patterns in the revision.
Unchanged: the same five podcast episodes in the same order as in version 1.
Compared with V1 · 03 Aug 2026. Changes are documented in this edition; the full previous edition is not available.
The curator’s table
Selected conversations to start with, including context and their most relevant insights.
Organization before technology
AI transformation is 70 percent organizational and people work. Anyone who treats AI as a highly gifted new team member rather than as a tool has to redesign roles, maturity levels and development conversations, not just buy licenses.
Enable rather than distribute
Canisius shows that AI in retail only takes effect when it is not run as a list of projects but managed in three clearly separated clusters and built on the company's own data. For leadership this means: data foundation and enablement first, then the core process.
Depth over breadth
Tesa shows that AI in finance builds on harmonized processes and a central data store without waiting for perfect data. Three to four deep use cases with a seven-figure impact beat many small experiments.
Sorting ideas
In this case AI is not managed as a pilot but as a profit calculation along traffic, conversion and costs. The breakthrough only came in the third test round after 15 months, which corrects expectations about realistic time horizons.
Roles in transition
Generative AI dissolves the role boundaries between product, design, data and engineering, but only four guardrails turn this into speed rather than chaos. For role design, competency models and recruiting profiles, this is currently the clearest template from a large corporation.
What keeps recurring
Drawn from this edition’s conversations. Open a finding to read the context and explore the supporting episodes.
Patrick McQuaid, Group Head of Data, AI and Analytics at NGM Group, presents his AI risk framework to the company's own model and has it look specifically for weaknesses and gaps.
Lawyer Dennis Engbrink (ZIRNGIBL) has the same set of facts assessed one after another from a legal perspective and from the perspective of the producer, organizer or intermediary, and treats the outputs like advice from consultants.
Because AI assistants tend to agree by default, Anna Ott, Vice President People at HV Capital, built a Claude skill that responds in the style of a Berlin bus driver and thereby deliberately generates pushback.
Supported by these conversations
A Statista customer burned through a quarter of a million credits in the first three days with an agent that had gotten out of control; Statista works with stop-loss limits beyond which usage is stopped.
Ben Hoskins (CTO, NewStore) blocks his coding agent's bash actions and instead has refactorings carried out through IntelliJ functions connected via MCP.
Manu Narayan (CIO, GitLab) anchors guardrails and governance on the repo side, because no reliable control can be enforced on developers' machines, whether they use Claude Code, Codex or open models.
Supported by these conversations
321 MED has agents that know nothing of each other check one another, so that a second AI stops an incorrect output to patients before it goes out.
The CPTO of felmo embeds quality rules as hooks that enforce a security check and an architecture check before the push to GitHub, because skills get forgotten in the model's context.
Parloa has every pull request reviewed by several specialized agents; an LLM judge evaluates their consensus and sends risky pull requests to a person for review.
Supported by these conversations
BARMER has agents understand and break down scanned documents for co-payment exemptions, while the eligibility check runs on deterministic logic, because an age comparison by an agent, for example, would be expensive and carry a risk of hallucination.
Philipp Schmid (Google DeepMind) estimates that about 50 percent of agent projects could be solved more accurately with a deterministic workflow, because with an agent the model itself decides on tool selection and when to stop.
Marie Kilg and Gregor Schmalzried (“Der KI-Podcast”) advise checking before every workflow whether a language model is needed, and automating if-then steps such as a claims threshold in insurance deterministically.
Supported by these conversations
At GitLab, all employees may submit skills to a GitLab repo; a review process decides on inclusion in the official repo, from where they are integrated into Claude and distributed by persona.
Stefan Lange-Hegermann, Lead for AI and APIs at sipgate, stores documents and strategy materials as Markdown in folders, uses Obsidian as a viewing layer and has Claude Code sort collected meeting transcripts.
Andrea Lederer, Chief Digital Officer at Weleda, rarely writes more than three sentences per prompt but works with project files and an OS file, and in this way developed a three-year plan in about three hours instead of roughly a week.
Supported by these conversations
At GitLab, six to seven roles, from business system analyst to QA, were involved in internal CRM development; with AI, individuals cover several of them and internal handoffs disappear.
In 2026, Instagram introduced pods of four to six fairly generalist engineers and one product staff member who also takes on parts of design, data science and research.
Nick Olsen (Mainsail Partners) describes how product managers, design and engineering work together on the customer problem from the start instead of passing documents along; some product managers commit code themselves.
Supported by these conversations
Gebr. Heinemann uses a model-agnostic, EU-hosted platform with Claude, ChatGPT and Mistral on which assistants, agents and workflows remain in place when the model changes.
According to Uwe Franke (founder of ByteWorx), all of his overnight jobs stopped at the same time when the Anthropic model Fable was switched off for non-US users by a US government decision; he calls for keeping control of one's own stack.
When a new Claude version was temporarily available only to US customers, sipgate rented a machine in Finland for one week for around 7,000 euros, running up to about 80 sessions at peak with an open coding model.
Supported by these conversations
Ben Hoskins, CTO of NewStore, built a point-of-sale system in four days and showed it at the all-hands of product engineering, design and support; only this demo noticeably raised the level of use.
Deichmann supplements the pilot training for Copilot licenses with show-and-tell calls several times a week, in which colleagues show from their day-to-day work how they use the AI tool.
A client of Christa Hill (Learnit) brings teams together across functions, and each person has to show something they successfully built with AI in the past one to two weeks.
Supported by these conversations
Schwarz Digits collected a mid three-digit number of AI cases in administration across countries; around 98 percent were not AI topics, and the few remaining cases move into a proof-of-value phase.
Drees & Sommer assessed eight ideas from an ideathon for technical feasibility and a measured baseline, and launched the three use cases that met both criteria with 150 test users.
Johannes Foertsch (OpenAI) describes how top users, instead of lists such as 89 use cases in one department, choose at most five processes with ownership, KPIs and a baseline, and decide after about 100 days whether to scale or stop.
Supported by these conversations
Erik Brynjolfsson reports on a company whose users of AI tools for programming were around 41 percent more productive than non-users; after controlling for selection bias from power users, around 11 percent remained.
David Folwell (Staffing Referrals) calls logins and tool usage adoption rather than ROI, asks about fill rate, time to fill and placements per recruiter per month, and considers a holdout group without the AI tool the gold standard.
GitLab rejects “token maxing”, tracks tokens only for cost control, and measures release cadence, review time and pipeline failures in development, and time to first response in customer service.
Supported by these conversations
The complete collection
118 conversations, ordered from newest to oldest.
Your selection
Your selection from this thinking room is saved in this browser.
No saved episodes yet.
Browse the catalogue
