Foundation Paper · v1.4 · April 2026
Friederike von Waldenfels · Forestrock GmbH
Three findings
Force
Internet, cloud and mobile accelerated communication, infrastructure and access — none of them touched the act of thinking. AI does. Because cognitive work sits at the centre of how organisations decide and act, the change does not arrive through any single channel: it cascades simultaneously across hierarchies, roles, incentive systems, value creation and business models — all hanging on the same compressed feedback loop.
Mechanism
Three years into broad AI deployment, 89% of executives report no measurable productivity impact. Reading this as a failure of the technology is wrong. The gap sits in the leverage points that determine how people actually work together — incentives, goals and the unwritten assumptions about work and value — and those are precisely what most companies leave untouched while reorganising their tooling.
Conclusion
Because the change is simultaneous and the bottleneck is institutional, point solutions and pilots layered onto the existing system get neutralised by it. The right level of intervention is the operating model itself: how value is created, how the organisation works, how roles are defined, how governance holds, how technology connects, and what infrastructure enables.
The six-layer operating model
What changes
AI dramatically lowers the cost of standardisable work. Customers insource services they used to buy. AI agents take over search, comparison and purchasing — intermediaries whose value rested on information advantage lose their raison d'être. Products emerge that behave like bespoke services but scale at product cost ("Bespoke at Scale"). The standardised middle gets squeezed from both sides.
Design principle
Those who merely execute become replaceable; those who orchestrate, contribute domain knowledge and take ownership become more valuable. Evolve the billing model — AI-tier / human-tier, gain-sharing, outcome-based. Treat platform decisions as a board-level topic.
What changes
Hierarchies served as information filters — when AI knows the full context, that function disappears. Feedback loops compress from quarters to weeks. Operational knowledge work migrates to the AI layer; human work concentrates on four discrete role types. Bonus systems that reward team size, and career paths that only point upward, penalise efficiency.
Design principle
Gradual transformation from classical hierarchy to a networked target organisation. Define teams by outcomes. Develop the four role types. Shift incentives toward impact. Don't treat automation as a cost programme — reinvest savings into higher-value roles. Build experimentation speed as an organisational capability.
What changes
Jobs don't disappear, but the composition of tasks changes. The half-life of a role description shrinks. When AI takes over execution, skill degradation looms: review competence erodes when the underlying skill is no longer practised.
Design principle
Continuous reskilling as part of work. Deliberate competence preservation through rotation between AI-supported and independent work. Roles remain as containers for responsibility; skills become the currency for deployment and mobility.
What changes
AI delivers probabilities, not certainties — classical quality assurance (excluding errors) no longer works. 41% of employees already use AI without their employer's knowledge. When AI systems act and learn autonomously, new liability and IP questions arise.
Design principle
Manage variance instead of preventing errors: define operating ranges, set escalation points, monitor the system's learning process. Make autonomy decisions deliberately — where may AI act, and may it adapt its own rules? Accountability stays with humans. AI policy before pilots.
What changes
Legacy IT was built around individual systems (ERP, CRM as silos). AI needs an orchestration layer that connects systems and agents — without APIs, a system is invisible to AI. Processes shift from linear flows to learning loops: goal → action → feedback → adjustment. Physical AI (robotics) is the next wave.
Design principle
Build systems to be connectable and modular — that is the prerequisite for everything else. Redesign processes from the desired outcome; don't layer AI on top of old workflows. Treat workflows as products: owned, versioned, measured.
What changes
Data used to be a reporting instrument. AI now consumes data in real time — without clean, accessible data, every AI project is built on sand. AI infrastructure concentrates on few actors: ~90% of global compute sits in the US and China. Energy becomes the physical bottleneck. Cybersecurity becomes a new category: prompt injection, model poisoning and agent manipulation are attack classes classical IT security doesn't know — and AI lowers costs for attackers.
Design principle
Data-first, not AI-first: the data question comes before the AI question. Company context (rules, logic, tacit knowledge) must become accessible to AI. Secure sovereignty — the platform decision is as long-term as the ERP choice once was. Treat cybersecurity as part of sovereignty: new attack surfaces, supply-chain risks and identity management for autonomous agents belong in the architecture, not in a downstream audit.
What stays the same
Long-standing customer relationships, the judgement calls made in a difficult negotiation, the confidence a key account has in the person on the other end of the phone — none of this is absorbed by AI. When standardised work commoditises, the differentiator moves upward: who is in the room, who holds the relationship, who the customer calls when something goes wrong. Relationship Managers become more valuable, not less.
AI produces output at high speed, but it cannot tell whether the output is correct in a given context. That judgement requires deep knowledge of the business, the customers, the edge cases, the regulatory reality. Domain Experts who validate AI output, catch hallucinations and make edge-case calls become the gating function for AI at scale.
Everything that requires presence, dexterity and real-time judgement in a physical environment — care, crafts, plant operations, maintenance, construction supervision — keeps its weight. The better the machine, the more expensive the human error: physical specialists become the critical reliability layer.
AI systems act probabilistically and learn autonomously, but liability does not transfer with the inference. Someone — still human — owns the outcome, signs the decision and answers when the system gets it wrong. Designing where that accountability sits is itself an operating-model decision.
Four role types
As operational knowledge work migrates into the AI layer, human contribution concentrates on four discrete roles. They are the containers around which to redesign teams, incentives and career paths.
Hold deep, contextual knowledge of the business, customers, edge cases and regulatory reality. They validate AI output, catch hallucinations and make the edge-case calls — the gating function that separates a company that trusts its AI from one that ships errors.
Design, instruct and supervise systems of AI agents end-to-end: defining goals, wiring tools and data, setting operating ranges and escalation points. They own the workflow as a product — versioned, measured and improved.
Carry the long-term trust that AI cannot earn: key accounts, partners, regulators, talent. As standardised work commoditises, the differentiator moves upward to who is in the room and who the customer calls when something goes wrong.
Operate in the world AI cannot reach — care, crafts, plant operations, maintenance, field service. The better the surrounding machine, the more expensive a human error becomes; physical specialists become the critical reliability layer.
Foundation Paper
Three findings on why pilots and point solutions don't move the needle — and why the operating model itself is the right level of intervention.
Why this tracker exists
A paper on AI is out of date the moment it ships. So the paper stays — and the evidence around it moves in the open.
Rather than rewriting the paper every few weeks, I check its concrete claims against fresh Tier-1 sources on a weekly cadence and log every change here. You see what still holds, what has shifted, and what has been superseded — with the original quote next to the new evidence.
A small in-house skill drives the workflow: extract claims from the current paper version, run a weekly research pass, curate drafts, publish. The paper is the source of truth; the tracker is the honesty layer around it.
Live-Updates
Every week I check the paper's claims against fresh evidence from Tier-1 research, regulators and press. Each entry references a specific claim in the paper — confirmed, updated, challenged or superseded.
Loading…
Read the full paper.