Monthly thinking room
Who keeps judgment, control and costs in hand when agents act and business units build for themselves? This edition shows how companies define permissions, checkpoints and metrics before AI has a say in day-to-day decisions.
For working with your own AI assistant. How it works
We handpicked 106 podcast episodes released in September 2026 and distilled 318 insights from them. These ten patterns show how companies decide what agents may handle on their own, measure AI by its effect on the process and let employees build their own solutions without losing control.
According to CTO Michael Carrel, 172 teams with almost 900 employees took part in Nationwide's hackathon, including 120 participants from business units; many of the business cases are to be funded from the technology budget through 2028.
According to lawyer Michael Roth, KPMG Law no longer relies on lecture-style training: employees rebuild the use cases they are shown themselves, and at a recent promptathon Roth gave participants 15 minutes to build an AI agent.
According to CEO Tobi Lütke, Shopify lets its agent River work only in open Slack channels so that around 7,000 employees see in passing how others request summaries, tickets or prototypes.
Status: confirmed, 2 new supporting episodes. Moody's has employees show in AI Cafés which problem they wanted to solve with Claude, which prompts they used and what came out of it, and in Anthropic's people team, someone regularly presents a technique of their own with Claude in open coffee rounds.
Supported by these conversations
For its chatbot Penny, Peapack Private Bank & Trust does not measure the questions answered, according to CTO John Kowal, but the tickets of the branch support team, which fell by 60 percent in four months.
Thorsten Mühl of cosnova notes that an adoption rate of around 95 percent with many prompts and custom GPTs does not on its own create business value, and wants to connect AI to core systems.
In Sebastian Winkler's small company, every experiment records how many hours it took to build and how many minutes it saves per week; Advantage Solutions prioritizes by service quality and time to complete a task.
Status: confirmed, 3 new supporting episodes. Julie Bedard of BCG advises rewarding the productive use of tokens rather than usage; Covestro prioritizes AI initiatives solely by their contribution to the profit and loss statement, and Leidos CIO Alexandra Guenther names revenue, costs, risk profile and employees who become productive faster as the measure of CIO success.
Supported by these conversations
According to Dominic von Proeck, Leaders of AI lets an accounting agent pass on receipts without review, has three independent AI models check the output of another AI and releases brand-relevant marketing content only after human review.
According to CEO Christian Klein, SAP ships agents for financial close, inventory management, shipping and invoicing only once they reach 95 percent accuracy; for customer briefings or headcount reports, 90 to 94 percent is enough because they still save preparation time.
Viren Lall of ChangeSchool recommends three levels defined in advance, from human decision to routine with spot checks; Daniela Hermle of Schunk considers human in the loop worthless if reviewers lack the expertise to assess AI outputs.
Status: sharpened. Leaders of AI places every task on a spectrum ranging from no review through a consensus process with three models to human approval, and SAP ties the release of agents to accuracy thresholds tiered by the severity of consequences of the task; this makes the gradation between “alone” and “human reviews” more precise.
Supported by these conversations
With a subscription
With a subscription you read all 10 with their sources and download them as a text file.
The complete collection
106 conversations, ordered from newest to oldest.
Breadth first, then depth
TRUMPF first spreads AI broadly with a GenAI platform and training, then redesigns a few core processes end to end instead of layering AI on top of old processes. Head of AI Sarah Engel names what this step requires: business units with 50 to 100 percent capacity instead of fragmented time, and tiered security requirements from the POC onward.
With a subscription
With a subscription you open all 106 conversations of this room with their insights, search the catalogue and download the files.
15 conversations added · none removed
7 confirmed · 2 sharpened · 1 unchanged
With a subscription you see which conversations, findings and recommendations changed.
Compared with V1 · 03 Oct 2026.
