Skip to main content

The Future of the Firm Is a Learning Loop

The mistake most companies are making about AI is treating it as a tool. That sounds sensible. It is also too small.

For thirty years, enterprise technology has largely helped organisations coordinate, store, search, measure and communicate. Software improved the firm, but it did not usually learn the firm. The intelligence remained mostly in people: in judgment, relationships, routines, institutional memory, and the tacit knowledge that lives in the heads of experienced operators.

AI changes that. For the first time, companies can create a live cognitive loop between human work and digital systems. Work can generate traces. Traces can become feedback. Feedback can improve agents, workflows, retrieval systems, evaluations, and decision support. The next piece of work can then be performed by a system that has learned from the last one.

That is the real transition. Not "AI adoption." Not "which model should we use?" The strategic question is: who owns the learning loop?

The answer will determine which companies merely get productivity gains and which ones build a new kind of organisational capital.

Human capital and token capital​

Every firm now has to build two forms of capital at once.

The first is human capital: the judgment, taste, experience, relationships and imagination of its people.

The second is token capital: the reusable AI-native capability a company accumulates from its workflows, private knowledge, decision traces, expert corrections, evaluations, tools, agents, and institutional memory.

Token capital is not a prompt library. It is not a chatbot. It is not even a fine-tuned model. It is the firm's ability to turn work into learning signal, and learning signal into better future work.

This distinction matters because the public conversation is still far too model-centric. Frontier models are extraordinary, and they will keep improving. But for most companies, the model itself will not be the moat. A competitor can buy access to the same model. It cannot easily buy years of proprietary workflow traces, private evaluations, expert corrections, customer context, risk decisions, failed attempts, and accumulated organisational judgment.

The moat is the loop, not the model​

The real opportunity is not picking the best general-purpose model. Models matter, but they are not the durable moat for most companies. A firm that merely rents intelligence from a frontier provider receives productivity. A firm that captures learning from each use begins to build compounding advantage.

You can offload a task. You may even be able to offload a job. But you cannot offload your learning.

That distinction is everything.

The research is already pointing in this direction. McKinsey's 2025 State of AI survey finds that adoption is widespread but scaled impact remains uneven; the companies getting value are not simply those "using AI," but those changing operating models, governance, workflows, and human validation practices around it. OpenAI's 2025 enterprise AI report makes a similar point from another angle: economic value appears after firms translate model capability into scaled use cases and core infrastructure. Reporting on the MIT NANDA GenAI Divide was blunter: most enterprise AI pilots are failing to produce measurable returns, not because models are useless, but because organisations are failing to integrate systems that learn and adapt inside real workflows.

That is the gap. Companies are renting intelligence, but not yet compounding it.

The firms that win will behave less like software buyers and more like learning architects. Their AI systems will capture how work gets done: prompts, tool calls, documents consulted, decisions made, edits applied, exceptions escalated, outcomes achieved, and failures corrected. They will build private evaluations that measure what matters inside their own business, not just public benchmarks. They will know whether an AI underwriting assistant, sales agent, legal reviewer, software engineer, claims processor, clinical documentation tool or procurement copilot is improving against the firm's standards.

Without private evaluations, companies are flying blind. They may produce more output, but they will not know whether they are producing more excellence. They may deploy more agents, but not know whether the agents are becoming trustworthy. They may automate activity without compounding capability.

What the enterprise architecture needs to become​

This is why the future enterprise architecture needs to be built around the loop.

  • Knowledge layer: documents, decisions, policies, customer context, code, designs, meeting notes, incidents, and domain reference material.
  • Trace layer: the actual record of how work gets done, including prompts, tool calls, agent actions, human edits, exceptions, approvals, and outcomes.
  • Evaluation layer: private tests that measure whether AI systems improve against business outcomes, not just public benchmarks.
  • Memory layer: institutional context that can be retrieved efficiently and governed properly.
  • Agent layer: workflow-specific systems that act with bounded autonomy, not one generic assistant expected to do everything.
  • Reinforcement layer: environments where systems can improve from real traces, expert corrections, preference signals, and outcome data.
  • Governance layer: permissions, auditability, provenance, retention, security, IP boundaries, and escalation paths.
  • Model layer: frontier and specialist models that can be swapped as the market changes.

The power is not in any one layer. It is in the loop between them.

Private evaluations are especially important. External benchmarks are useful for understanding general capability, but they do not tell a bank, retailer, manufacturer, government agency, hospital, or software company whether an AI system is getting better at the work that actually matters inside that organisation.

The firm needs its own evals: tests grounded in its own workflows, standards, risk appetite, customer promises, regulatory obligations, and definition of excellence.

Without private evals, the organisation cannot reliably know whether its token capital is compounding or simply producing more confident output.

The sovereignty test​

A company should be able to switch base models without losing its accumulated expertise. That is the practical test of AI sovereignty.

If changing the model destroys the capability, the company never owned the learning. It rented it.

If, however, the firm has built a learning architecture around the model, then the base model becomes more replaceable. The company retains its private evals, domain memory, tool interfaces, workflow traces, governance policies, examples of expert judgment, and reinforcement environments. It keeps the machinery that teaches any capable model how the organisation works.

That is the difference between adopting AI and owning an AI capability.

The model is replaceable. The learning loop is not.

Why the 2026 agent research matters​

The June 2026 AI research wave gives this argument a more concrete architecture. Agentic Software (arXiv:2606.05608) says the agent harness, tools, memory, evaluation, and governance layer are becoming the software system. Agent Memory (arXiv:2606.06448) shows that memory is not free context; it is a stateful workload with construction, retrieval, latency, quality, and staleness trade-offs. CORE-Bench (arXiv:2606.11864) shows that retrieving the right local repository context is its own hard capability. Agent Trajectories as Programs (arXiv:2606.16988) makes the path of an agent's work inspectable, not just the output.

That is exactly what token capital needs. A firm cannot compound learning if it only stores final answers. It needs the traces: what the agent saw, what it retrieved, what it ignored, where the human corrected it, which tool call failed, which risk decision was reversed, and which memory item later proved stale. The learning loop is not a folder of prompts. It is a governed evidence system.

The practical lesson is uncomfortable: companies that deploy agents without memory governance will accumulate cognitive sludge. They will have more generated output, but not necessarily better organisational knowledge. Companies that build retention, provenance, supersession, private evals, and trace review into the loop can turn daily work into compounding institutional capability.

Employees should be amplified, not extracted​

This also changes how we should think about workers. The best evidence does not support a simple story in which AI makes human expertise irrelevant. In the well-known study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work, access to AI assistance increased customer support productivity, especially for less experienced workers, partly by helping transfer the practices of more skilled workers. PwC's AI Jobs Barometer likewise finds that the most AI-exposed roles are placing rising value on judgment, creativity and empathy as routine tasks are absorbed by machines.

That should be the model: amplification, not extraction.

If employees experience enterprise AI as a machine that silently harvests their expertise and makes them disposable, they will withhold the very learning signal the organisation needs. Trust will collapse. People will stop correcting the system, stop sharing judgment, stop exposing ambiguity, and retreat into defensive work. The loop will starve.

But if employees see AI as a way to make their expertise more valuable, reusable and visible, the dynamic changes. Their judgment becomes institutional memory. Their corrections improve the system. Their taste becomes encoded in better workflows. Their experience stops being trapped in inboxes, meetings and turnover. Human capital and token capital compound together.

That is not automation as extraction. It is automation as organisational learning.

The political economy matters​

There is a broader political economy here too. An AI future in which every company uploads its expertise into a handful of external platforms and receives commoditised intelligence in return will not be socially stable.

We have seen the outline of this before. In the first phase of globalisation, aggregate efficiency often masked deep institutional and regional hollowing-out. GDP could rise while industrial communities lost their productive base.

We should not repeat that mistake with cognition.

The healthy equilibrium is not one in which a few frontier platforms absorb the value of every industry. It is one in which frontier models provide extraordinary general capability, while firms, workers, industries and countries own the learning loops that encode their specific knowledge. Platforms should enable more value above them than they capture inside themselves. That is the platform bargain worth defending.

This is also why "AI strategy" cannot be left to procurement or experimentation teams. It is a board-level question. The firm must decide what knowledge it is willing to externalise, what traces it must retain, what evaluations define excellence, what human expertise deserves compensation and recognition, and what parts of its operating system must remain sovereign.

The companies that move early will be hard to catch​

The companies that move early will be hard to catch. Not because they will choose one magical model, but because they will accumulate years of proprietary learning. They will know how their customers behave, where their experts intervene, what their risk teams reject, what their regulators require, what failure looks like, and which patterns have proved valuable over time. That compounding base will become increasingly difficult for late adopters to imitate.

A competitor can buy access to the same model. They cannot easily buy the accumulated judgment of how another organisation works, what its customers value, what its experts reject, what its risk teams approve, what failure looks like in its environment, and which patterns have proven to matter over time.

That is why the learning loop is the new moat.

The companies that treat AI as a tool will become more efficient. The companies that treat AI as a learning loop will become different kinds of organisations.

The model is not the moat.

The owned learning loop is.

Sources​