Monthly thinking room
When business units build their own solutions and agents run in production, who takes care of review, rules and direction? This edition shows which layers, levels and roles companies use to safeguard self-built solutions and which signals leaders should know now.
6 conversations added · none removed
4 unchanged · 5 confirmed · 1 sharpened
At KPMG, employees want to roll out around 4,000 agents they built themselves, and Kelle Fontenot sets up a checklist before every rollout: no duplicates, correct data references, the right security protocols and runtime environments, and no impact on upstream and downstream systems.
View finding →The Conference Board shows the sequence from at times 40 to 50 AI applications to targeted reduction combined with training. Kelle Fontenot (KPMG, the same person as in D2606-030) shows with three rollouts within about 18 months what happens without pacing: tool confusion, because no rollout reaches its target state. Indeed partly disagrees: more than half of its teams work with three or more coding tools, and training is deliberately designed to be independent of any single AI tool.
View finding →According to Leanne Allen, KPMG typically places an AI engineer alongside a person from the business unit, because people who have done a task for years can hardly redesign their process on their own, and only the interplay between the two drives automation rates up.
View finding →All supporting conversations in this edition
At KPMG, an agent checks every press release from corporate communications for whether it tells a story, leaders additionally receive feedback from an agent, and Leanne Allen deliberately uses ChatGPT as a sparring partner that challenges her own way of thinking (both episodes come from KPMG).
View finding →All supporting conversations in this edition
Indeed measures the impact of AI coding tools through system data rather than usage rates and, with 40 to 50 percent more diffs per developer year over year, sees the bottleneck moving to code review, because more merge requests are opened than resolved.
View finding →All supporting conversations in this edition
Maureen Davenport (KPMG) considers writing skills and years of experience to be merely a basic requirement when hiring in communications, and expects entry-level employees to take on considerably higher-value work and to master AI tools in order to apply judgment even in entry-level roles.
View finding →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 · 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.
Orchestrating transformation
Miro shows that AI transformation can be driven from the people function: non-technical roles build their own workflows, and a tiered model ties AI proficiency firmly to role and seniority level. An appeal becomes a binding framework of expectations.
Broad-based enablement
Enablement does not scale through course catalogs but through curated learning journeys and a jointly managed unit of IT and HR. Anyone who wants to bring thousands of employees along invests in the sluggish middle rather than in the vanguard that is already enthusiastic.
Anchoring change
Core thesis: transformations do not fail because of technology but because leadership responsibility is not taken up and middle management is ignored. For anyone who wants to move AI pilots into regular operations, this is exactly where breakthrough or abandonment is decided.
Work and leadership 2030
Rolling out AI tools does not yet amount to change as long as processes and roles remain untouched. What matters is enabling leaders before the workforce, otherwise adoption tips into knowledge hoarding.
Agentic practice
mobile.de shows that agentic development in a long-established legacy team is not a tool purchase but a shift of work to the edges: specification and context at the front, human review at the back. That very review becomes the new bottleneck and therefore a leadership decision.
What keeps recurring
Drawn from this edition’s conversations. Open a finding to read the context and explore the supporting episodes.
Parloa uses a deterministic policy engine to check before every tool call whether the upstream tools have run, so that the confirmation text message is sent only after the reservation has been made.
Upvest binds internal coding and knowledge agents to the identity and permissions of the respective user through impersonation, and at Seibert Media each specialist agent has only the credentials it needs.
tk accelis only scaled Smart Order once it had a deterministic search that filters out nonsensical material combinations; ABN AMRO uses tested input and output filters against prompt injection in its voicebots.
Supported by these conversations
Parloa invests in the harness around the model rather than in its own fine-tuning and points to a meta-study according to which a different harness with the same model can deliver six times better agent performance.
mobile.de indexes its entire codebase in a vector database, makes it available via MCP as context for Claude Code, and has rolled out its own skills repository to all 300 engineers.
Stefan Kemp (adesso) considers the model interchangeable when the harness controls tools, context and the agent loop; Jörn Ossowski (tk accelis) sees the agents' memory, context and tool access as the reason AI coding holds up.
Supported by these conversations
FINN has increased the number of employees who build internal tools from around 60 to more than 250; platform engineers safeguard this with templates, guardrails, CI/CD checks and security scans.
At the Constellation group, 25 non-programmers built prototypes with Claude Code in a day and a half, and IT reviews each one from a GDPR perspective before it may go into production.
At Kantar, every decentrally built agent used by several people goes through a review of IP, governance and quality in the Agent Factory before it is rolled out across the company.
Supported by these conversations
The Conference Board at times had 40 to 50 AI applications in use, is now narrowing down to the few most useful ones and training the team on them; André Neubauer (CTPO at Trusted Shops) is consolidating toward Claude after broad experimentation.
KPMG rolled out Microsoft Copilot to all employees in the US, then Google Gemini six to eight months later, followed by the Anthropic deal; Kelle Fontenot calls the result tool confusion and plans a digital concierge as the single interface.
The Haufe Group plans AI announcements from IT, HR and Corporate Communication in a launch calendar three months in advance; LeverX recommends a change advisory board that decides when which project starts.
Supported by these conversations
At Control Expert, appraisers and data scientists sit on the same floor and improve the AI-based damage assessment through direct exchange, which Christian Temath (KI.NRW) cites as the reason it scaled.
Smart Order at tk accelis initially failed when IT simply delivered the finished solution, and only made progress once a business sponsor assigned their own people to the initiative.
The Haufe Group has merged IT and HR into Corporate Services; for AI competence, one responsible person each from IT and HR works jointly on the roadmap, metrics and measures.
Supported by these conversations
Lower Saxony's Minister of Education Julia Willie Hamburg puts finished speeches into an AI tool and gets feedback on argument structure and clarity, while she continues to develop the ideas herself.
Ethan Mollick (Wharton School) recommends having finished texts read by AI personas, such as a reader from outside the field or a critic looking for mistakes on social media, to get feedback from several perspectives.
The Group AI Officer of Erste Group explicitly asks ChatGPT or Copilot Chat to challenge an idea of their own and point out blind spots, and Stefan Dietz has the AI scrutinize a business idea in the role of a critical startup investor.
Supported by these conversations
tk accelis starts every new area with a low automation rate, systematically analyzes corrections and rejection reasons, and achieves more than 70 percent automation for structured documents.
Seibert Media measures agent operations in IT support over 30 days, with a 29 percent automation rate and a one percent error rate, and tracks the 26 percent of cases with follow-up questions as a backlog for the next automation stage.
At a mid-sized company whose sales Scalantec automated, the owner spent roughly a year first taking out the lowest-performing sales employees and only made a decision once first meetings, response rates and deals kept pace.
Supported by these conversations
Andreas Stryz (FINN) calculates that 10 initiatives per year used to produce one hit, while today 100 initiatives potentially yield 10 hits and 90 unusable projects, and he makes the hit ratio a shared metric for his teams.
At mobile.de, engineers working with Claude Code hardly write any source code themselves anymore; CTO David Gebhardt sees the additional effort in the specification before implementation and in the review of the pull request.
In the HMZE podcast, Sebastian Heide-Meyer zu Erpen and André Neubauer analyze Big Tech reports: Spotify reports 93 percent more pull requests, while Uber's CFO publicly sees no direct link to business metrics.
Supported by these conversations
In the GPRA-YP study with 107 respondents, 21 percent of young PR consultants say they are involved in ideas and concepts earlier, and 20 percent take on more quality assurance and revision.
Using the example of a law firm, Sinead Bovell describes how AI agents are eroding the paralegal path and how an agent director path is emerging, and she advises leaders to redesign junior positions instead of thinning them out through hiring freezes.
At Soundstripe, agents write all of the code; CTO Trevor Hinesley sees new responsibilities arising from this, such as acceptance criteria and harness maintenance, and expects that juniors will have to take on system responsibility earlier.
Supported by these conversations
Csaba Tamas (Parloa) pairs a code-building agent with a critiquing agent, because coding agents, when in doubt, adapt the unit test to the faulty logic instead of finding the bug.
At EnBW, the speaker triggers a code review by OpenAI's Codex model from within Claude Code and calls for at least a second instance in the loop, either a second human or a second model.
At mobile.de, AI-generated code has often already been pre-reviewed by a second AI before a human checks the pull request, and CTO David Gebhardt worries that the human will then merely click a button in GitHub.
Supported by these conversations
The complete collection
59 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
