The compounding organization: how GeMarkt plans to learn, automate, and expand

Most companies describe growth in quantities: more people, more products, more markets, more revenue. Those measures matter, but they miss the kind of growth GeMarkt is trying to build first.

We want each completed project to make the next project easier to begin, safer to execute, and cheaper to verify. A research decision should become a reusable record. A failure should become a test. A manual judgement should become a clearer review rule. A successful automation should become a bounded capability that can be recombined with others. A new model should be replaceable without forcing the company to relearn what good work looks like.

We call this the compounding organization: an organization whose capacity grows not only because it acquires more resources, but because it retains and recombines what it learns.

The phrase is our own synthesis, not an established scientific category. Its foundations are well established, however, across research on organizational learning, memory, dynamic capabilities, knowledge transfer, and technological change. That research also supplies an important warning: experience does not compound automatically. Organizations forget, routines preserve mistakes as easily as they preserve wisdom, and new technology often disappoints when it is added without the complementary work needed to use it well.

This essay is therefore a vision, but not a prediction of frictionless growth. It sets out what GeMarkt is trying to become, which parts exist today, and what evidence would be required to say that the vision is working.

Experience is not yet institutional memory

An organization can repeat an activity for years and still fail to become much better at it.

Linda Argote and Ella Miron-Spektor describe organizational learning as a process in which experience interacts with context to create knowledge. The context matters because knowledge can become embedded in people, routines, tools, and relationships rather than living in one document or one mind. (Argote and Miron-Spektor, Organization Science, 2011)

But experience is perishable. In a study across production organizations, Argote, Sara Beckman, and Dennis Epple found that knowledge acquired through production depreciated rapidly; cumulative output overstated how much learning actually persisted. (Argote, Beckman and Epple, Management Science, 1990) A company does not retain a lesson merely because somebody once learned it.

That distinction defines the first requirement of a compounding organization. Work must leave a retrievable change behind.

James Walsh and Gerardo Rivera Ungson’s framework for organizational memory separates the acquisition, retention, and retrieval of information. (Walsh and Ungson, Academy of Management Review, 1991) Maurizio Zollo and Sidney Winter go further: durable capabilities develop through a combination of accumulated experience, deliberate discussion, and knowledge codification. Simply doing more is not the same as learning deliberately. (Zollo and Winter, Organization Science, 2002)

For GeMarkt, this means a completed task should produce some combination of:

• a decision record explaining what changed and why;
• structured data with an identified source of truth;
• a repeatable workflow with a named owner and boundary;
• a test, comparison, or review artifact that defines acceptable quality;
• a recorded failure mode and the condition that exposed it;
• a permission rule stating what automation may propose and what still requires a person;
• a rollback or replacement path.

The valuable output is not a long agent transcript. It is the better system left after the conversation ends.

GeMarkt already has early examples of this principle. A blind restoration comparison did more than select a default tool: it created a habit of logged decisions and revealed what the next evaluation should record. Persistent project memory carries product and design decisions into later work. Schemas, migrations, checksums, tests, and deployment guards turn one successful implementation into a constraint that later implementations inherit.

These artifacts are not proof of a self-improving company. They are the beginnings of an organizational memory that does not depend entirely on the founder remembering every detail at the right moment.

Technology creates options only when the organization can absorb it

Following technology closely is necessary for GeMarkt, but chasing every release would be the opposite of a strategy.

Wesley Cohen and Daniel Levinthal called a firm’s ability to recognize, assimilate, and apply valuable external knowledge its absorptive capacity. Their central insight was that this capacity depends heavily on prior related knowledge and is path-dependent: what a firm has learned before influences which new knowledge it can use next. (Cohen and Levinthal, Administrative Science Quarterly, 1990)

This explains why the same AI model can be transformative for one organization and little more than a demo for another. The model is only one component. The organization also needs task definitions, trusted data, evaluation cases, integration points, permissions, cost controls, and people able to recognise a plausible failure.

Research on the AI productivity paradox makes the same point at a larger scale. Erik Brynjolfsson, Daniel Rock, and Chad Syverson argue that general-purpose technologies require complementary innovations, new skills, and organizational redesign before their full value appears. Those investments behave like intangible capital and can create an implementation lag between technical capability and measured productivity. (Brynjolfsson, Rock and Syverson, NBER, 2017)

So GeMarkt’s technology practice should follow a repeatable cycle:

1. Sense: monitor new models, agent methods, standards, infrastructure, interfaces, and cost changes.
2. Test: compare them on real GeMarkt tasks against a recorded baseline, including failure and rework.
3. Bound: decide what the technology may read, propose, change, spend, or publish.
4. Integrate: place it behind stable contracts so the rest of the organization does not depend on one provider’s behaviour.
5. Retain: record the result, the limitation, the evaluation date, and why the decision was made.
6. Retire: remove or replace a tool when a better option wins under the same evidence standard.

This is how staying close to technology becomes a durable capability rather than a series of migrations driven by excitement.

The frontier is advancing, but it remains jagged

There is good reason to keep testing. AI capability is moving quickly. METR’s task-horizon research estimates that, across its suite of software and reasoning tasks, the human-equivalent length of tasks that frontier agents can complete with 50% reliability roughly doubled every seven months between 2019 and late 2025. METR is explicit that this measure does not represent the whole economy or all real work; its value is as a tracked signal of progress in multi-step task completion. (METR, updated 2026)

Progress is not uniform. In a preregistered experiment with 758 consultants, AI improved speed and output on tasks inside its capability frontier, but made participants less likely to reach the correct answer on a task outside that frontier. (Dell’Acqua et al., Organization Science, 2026) In another field experiment, 791 Procter & Gamble professionals worked on product-innovation problems: individuals using AI matched the performance of two-person teams without AI, while human judgement retained value in selecting the best ideas. (Dell’Acqua et al., Organization Science, 2026)

Other settings produce less flattering results. Sixteen experienced open-source developers completed 246 tasks in repositories they knew well and took 19% longer when using early-2025 AI tools, despite expecting to become faster. (METR, 2025) A July 2026 study of self-organizing multi-agent teams found that they consistently failed to match their best individual expert on the evaluated benchmarks, with losses reaching 41.1% on some machine-learning tasks. The teams often found the expert and then diluted its answer through compromise. (Pappu et al., 2026)

The conclusion is not that AI works or that it does not. The conclusion is operational: capability must be tested at the level of the actual task, workflow, and user. A fast-moving frontier rewards organizations that can evaluate and integrate quickly. A jagged frontier punishes organizations that confuse access to a model with possession of a capability.

From isolated automations to an operating system

An automation performs a task. An organizational capability keeps performing that class of task as inputs, tools, and circumstances change.

The difference is the surrounding system. A useful capability needs memory, interfaces, evidence, permissions, monitoring, and an accountable owner. It must be possible to tell when it is working, when it has drifted, and when a person should intervene.

Over time, GeMarkt’s operating layer should contain five connected forms of capital:

1. Knowledge capital

Rights decisions, restoration experience, catalogue structure, customer questions, product performance, operational failures, and design principles should become searchable and reusable without being stripped of source, date, or uncertainty.

2. Automation capital

Repeated work should move into modular workflows with deterministic checks around probabilistic output. Each module should do a bounded job and expose a stable interface, making it possible to improve or replace one part without rebuilding the whole chain.

3. Evaluation capital

Benchmarks should be drawn from real work. A new model should not be judged only by a public leaderboard; it should face the failure cases, edge conditions, costs, and quality thresholds that matter to GeMarkt. The evaluation set itself becomes more valuable as it accumulates representative cases.

4. Trust capital

Customers, partners, and future team members should be able to see what GeMarkt verifies, what it does not know, where human approval remains, and which promises the system is structurally unable to make without evidence.

5. Identity capital

As the product range and technology change, the organization needs a stable answer to a deeper question: what kind of company are we building? Research on organizational identity suggests that identity need not be frozen to remain coherent; some instability can help an organization adapt as its image and environment change. (Gioia, Schultz and Corley, Academy of Management Review, 2000)

For GeMarkt, the enduring identity is not a particular model, marketplace, product format, or infrastructure provider. It is a method: source carefully, verify claims, improve the work without falsifying it, make the customer path honest, measure outcomes, and keep a person accountable for consequential decisions.

Technology can change quickly around that method without making the company unrecognisable.

Capabilities should expand before sectors do

The long-term vision is larger than today’s art catalogue. But expansion should follow capability adjacency, not fashion.

Bruce Kogut and Udo Zander argue that firms create new knowledge by recombining existing capabilities; what a firm has done before shapes the options it can pursue next. (Kogut and Zander, Organization Science, 1992) Constance Helfat and Kathleen Eisenhardt describe how firms can obtain economies of scope over time by redeploying capabilities between related businesses. (Helfat and Eisenhardt, Strategic Management Journal, 2004)

That is the model for GeMarkt’s expansion.

The first layer remains the core: researched and restored art, physical and digital products, discovery, and a trustworthy path to purchase. The next adjacencies can reuse the same capabilities: more product formats for the home, better tools for choosing and arranging art, deeper cultural and rights-aware catalogues, and a commerce system able to carry a broader range of GeMarkt products.

Beyond that, the research, media, catalogue, workflow, and verification capabilities may create options in related product or service areas. Those are options, not present commitments. A new sector should qualify only if it passes a small set of tests:

• it serves a customer or problem GeMarkt can understand directly;
• it reuses several capabilities the organization has already proved;
• it strengthens rather than confuses the GeMarkt identity;
• it can begin as a bounded, measurable pilot;
• it contributes knowledge or infrastructure back to the core;
• it has evidence of customer and economic value before scale.

This rule matters because automation can make starting things deceptively cheap. The strategic cost appears later, when unrelated products compete for attention, data, support, capital, and trust. The compounding organization should accumulate coherence, not merely activity.

Compounding can preserve error as efficiently as knowledge

There is nothing automatically virtuous about institutional memory. A bad rule can become a test. A biased decision can become training data. A temporary workaround can harden into architecture. A fluent group of agents can produce a consistent explanation of something that is not true.

James March’s work on exploration and exploitation describes a related danger: systems that refine established knowledge can become efficient in the short run while undermining long-run adaptation if they stop exploring alternatives. (March, Organization Science, 1991)

GeMarkt therefore needs forgetting and revision as well as memory. Important records should carry dates. Tests should be reviewed when the environment changes. Models, vendors, prices, policies, customer behaviour, and legal interpretations should not be treated as timeless. Decisions need evidence, but they also need expiry conditions.

Other risks grow with the system:

• automation debt: workflows nobody understands but everyone depends on;
• provider dependence: organizational knowledge trapped inside one model or platform;
• false certainty: a large volume of consistent output mistaken for independent evidence;
• identity dilution: adjacent expansion that weakens the reason customers trust the brand;
• founder concentration: too much permission, context, and operational continuity resting on one person;
• measurement capture: optimizing the recorded proxy while the real customer outcome deteriorates.

A compounding organization must make these risks easier to see, not hide them behind higher output.

We should be able to measure whether capacity is compounding

The thesis becomes empty if every new automation is counted as progress. GeMarkt should call its capacity compounding only when the operating record improves.

The relevant measures include:

• elapsed human time from a defined objective to verified completion;
• defect, rework, rollback, and exception rates;
• cost per accepted outcome rather than cost per generated output;
• the share of new work that reuses an existing contract, dataset, evaluation, or workflow;
• time required to evaluate, integrate, replace, or retire a model;
• the size and age of human-review queues;
• the proportion of consequential claims with a current source and review date;
• customer, revenue, retention, and margin outcomes from adjacent experiments;
• key-person and provider concentration.

Some of these measures do not yet exist. That is itself useful information. The next stage is not to invent a success number; it is to build the instrumentation that would make success or failure visible.

The vision, separated by state

As of 1 August 2026, GeMarkt is still a human-managed, agent-assisted one-person studio. Its live customer experience includes art discovery, buyer tools, and first-party measurement, while purchases complete through external marketplaces. A private media path exists in a non-production AWS environment. Provider-neutral catalogue work, persistence, migrations, and automated checks exist locally or in CI as foundations for future commerce.

Those layers should not be collapsed into a claim that an autonomous retailer already exists.

The next planned stage is integration: connect reviewed catalogue, media, commerce, measurement, and operating memory without weakening the boundaries that keep draft work, non-production systems, and customer promises separate.

The longer-term direction is a company that can absorb new technology quickly, strengthen its automations through measured experience, extend its product and customer capabilities, and enter related sectors where its accumulated knowledge creates a real advantage.

None of this requires GeMarkt to remain one person forever. Small human headcount is a present condition and a design constraint, not an ideological ceiling. The point of organizational compression is to make human participation more deliberate: add people, partners, or specialists because their judgement, relationships, craft, and responsibility create value—not simply because the organization forgot how its own work operates.

A company that grows by remembering

The goal is not to build a company that needs no people. It is to build a company that forgets less, learns faster, and turns every completed project into greater future capacity.

If GeMarkt succeeds, its deepest advantage will not be access to a particular model. Competitors can buy the same model. The advantage will be the accumulated system around it: the cases, standards, decisions, interfaces, measurements, identity, and trust that determine how technology is used.

That is how automation becomes more than saved labour. It becomes institutional capability.

And that is how a small organization can grow into a much larger one without having to begin from zero each time the technology, product, or market changes.