Monthly thinking room
How much responsibility can be handed over to agents before control, knowledge and trust can no longer keep up? This edition shows how companies manage agents, review them and supply them with their knowledge, and where adoption gets stuck at interfaces and fears.
1 conversation added · none removed
Added
10 unchanged
All supporting conversations in this edition
All supporting conversations in this edition
All supporting conversations in this edition
Unchanged: the same five podcast episodes in the same order as in version 1.
Compared with V1 · 04 Sept 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.
Mandate instead of individual prompts
The leap lies not in better prompts but in the mandate: goal, a framework with prohibitions, tone, escalation thresholds and depth of oversight turn operating AI into genuinely leading it. This matters for decisions because responsibility and liability thereby remain explicitly with management.
Rethinking the business model
Statista turns the SEO business model around and makes its own data holdings a machine-readable product for language models. For anyone who has so far drawn their reach from search engines, this is where their own future viability is decided.
Making data quality measurable
The jump in quality does not come from a better extraction model but from a second AI that checks every extraction against a maintained gold standard. Anyone applying AI to messy data needs exactly this feedback loop rather than a bigger model.
Foundation before speed
Rösener ties adoption to leadership behavior: where the leader works with AI daily, the department's throughput is four times higher. Added to this is a robust operating model of usage-based billing, hackathons close to the business units, and in-house control over hosting.
Leading without a head start
Faecks's thesis: the pace of AI development and the linear human learning curve are diverging, so expertise no longer supports a stable claim to leadership. Anyone who continues to reward learning individually creates lone figures instead of a learning organization.
What keeps recurring
Drawn from this edition’s conversations. Open a finding to read the context and explore the supporting episodes.
Beam stores its entire company context in GitHub, one repository per customer with readme files and instructions, so that a colleague's coding agent can find the information.
Moody's has one person on the team maintain the agents' context layer each week on a rotating basis; at Statista, whoever creates a data series answers five review questions, which raises answer quality by 15 to 20 percentage points.
Instead of Figma designs, Thermondo's design team delivers finished components and skills that teach agents how to use them; this lets engineering and product management assemble interfaces in the design team's style.
Supported by these conversations
BNY gives each of its more than 150 digital employees its own login, a human manager, onboarding and a performance review, and also measures effectiveness by hallucinations and model drift.
Lexware describes the handover to AI as a mandate made up of five components (goal, scope, tone, escalation, control) and provides more than 40 ready-made mandates on a dedicated website.
In John Healy's client case, every agent has metrics and is kept, improved or shut down; Kate Smaje of McKinsey warns that agents without lifecycle rules keep running as abandonware.
Supported by these conversations
Beam achieves around 95 percent accuracy with a low-cost model and a second low-cost model as a checking layer, while the same prompt on the latest model generation lands at about 74 to 75 percent.
An ongoing film production stays with ByteDance's older model Seedance 2.0 because Seedance 2.5 was not better in testing, only longer, and thus produces at 20 percent of the cost.
Haus Cramer Gruppe moves the language model and ticketing system for its agentic expansion stage into its own hosting so that no provider can charge 30 percent more tomorrow; Libra lets law firms choose between models from Google, OpenAI and Anthropic for each task.
Supported by these conversations
Beam sets an error budget for each type of error in advance: where GDPR is involved, full accuracy is required; for a wrongly shipped order, five hundred to a thousand euros are acceptable, a hundred thousand euros are not.
b.steuern decides on a risk basis which receipts a human still looks at, and aims to review no more than five percent of client receipts manually within twelve months at the latest.
BLP builds a “temporary final check” by accounting into invoice processing, which is to be dropped after three months without changes; Moody's wants to reduce the weekly review of an agent to roughly monthly once trust is established.
Supported by these conversations
REWE checks every product data extraction with “Goldhamster”, a second Gemini-based AI, against a manually maintained gold standard, and raised data quality from 80 to 95 percent in four months this way.
Felix Schlenther of AI FIRST places a reviewing agent behind the first one that assigns a score, and approves automatically only above a threshold, 95 percent in his example; DKB checks agent results with different models than those used to generate them.
At thekey.academy, a second AI instance checks the first one's results in the shared GitHub repo, and a human also looks at key points; at FINN, AI checks 98 percent of the source code before it goes live.
Supported by these conversations
General practitioner Sebastian Alsleben names a lack of interoperability as the most common rollout problem: if the AI phone assistant does not talk to the appointment calendar and practice management software, appointments are still transferred manually.
According to WalkMe's State of Digital Adoption report, only 29 percent of workflows are accessible to models via APIs or MCP, and 37 percent of employees skip AI because supplying the context takes more effort than it delivers.
An AI coaching pilot for leaders reached ten minutes of total usage in five weeks because the AI tool sat outside the workflow; Logitech, by contrast, makes the feedback from its Board Advisor Gem a fixed step before board materials are submitted.
Supported by these conversations
According to Jonas Bernard, a large manufacturer has a voicebot call its field sales staff right after each appointment, which writes the answers into the CRM and triggers follow-up actions; this saves one to two hours of follow-up work per person per evening.
General practitioner Sebastian Alsleben describes ambient listening that automatically documents conversations and recognizes diagnoses; with 24 patients per consultation session, he estimates that around one sixth of the work remains.
Felix Schlenther has every meeting at AI FIRST transcribed; an AI extracts the action points, prepares follow-ups and creates the items in the project management tool.
Supported by these conversations
At Zapier, support employees write bug fixes themselves with coding agents and deliver a finished merge request that engineering reviews and merges instead of starting from scratch.
Arizona State University provides all employees with an internal vibe coding platform on which around 2,000 people, mainly from administration, build their own applications, for example for parking management.
Grant Thornton had its workforce in Ireland build agents for real business problems in a one-week campaign, and several are now going into regular operations; Logitech requires champions to deliver a working artifact rather than an idea.
Supported by these conversations
The managing director of a data platform provider initially let everyone experiment freely, mainly with ChatGPT and Claude, saw new silos emerging and introduced Langdock because agents and workflows can be shared there.
thekey.academy's GTM team replaced eight daily agent reports in various Slack channels, which no one read anymore, with a shared cockpit of around nine tabs.
Marco Morgenstern, the member of an IT company's executive management responsible for HR, built more than 20 web apps and now distributes apps and skills through a marketplace on a SharePoint page instead of sending skill files by email.
Supported by these conversations
Will England of Walleye Capital writes to the entire firm that he used ChatGPT for this very email, that the cheating argument does not apply in business, and that everyone should use ChatGPT and be proud of it.
The head of people development at SaaS provider Exclaimer raised AI usage from 45 to 97 percent in around four months by addressing in front of everyone the question of whether employees are training their own replacement, and answering it with no.
Samantha Gloede of KPMG derives two leadership tasks from the period when employees used AI secretly because it was considered cheating: provide equivalent company tools and signal that no one has to hide.
Supported by these conversations
The complete collection
95 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
