TTS

Cognitive Infrastructure for Artificial Collective Intelligence

An archival brief on the missing enterprise-AI layer

Olaf Witkowski and Ahmer Inam

---

Thesis. The scarce layer in enterprise AI is no longer access to a frontier model. It is access to governed institutional cognition: tacit expert judgment, community-specific context, provenance, consent, and the right to use that knowledge safely. Cognisee is building Cognitive Infrastructure: a governed layer for capturing, protecting, composing, and compounding that knowledge into artificial collective intelligence.

The Missing Company in the AI Stack

The AI stack has model providers, deployment platforms, agent frameworks, and application wrappers. What it still lacks is the layer that turns institutional expertise into a durable machine-reasoning asset. That missing layer matters because enterprise value does not come from generic model capability alone. It comes from applying capability to knowledge that is proprietary, situated, trusted, and operationally relevant.

The strongest AI companies of the next decade will not merely expose intelligence through an interface. They will own a compounding substrate: a system that improves as institutions use it, that becomes harder to replace as it accumulates expert context, and that is safe enough to operate in regulated, high-stakes environments. This brief argues that the defensible substrate is governed tacit knowledge infrastructure.

BCG’s 2025 study of more than 1,250 firms found that only about 5% were achieving AI value at scale, while 60% reported minimal material value despite substantial investment.¹ The interpretation is straightforward: the bottleneck is shifting from model access to institutional transformation, workflow redesign, data quality, and governance. In other words, the decisive question is not whether enterprises can buy AI. They can. The question is whether AI can reason over the knowledge that actually governs consequential work.

The Part Nobody Has Solved

Michael Polanyi observed in 1966 that we can know more than we can tell. This is not a romantic claim about intuition. It is a structural fact about expertise. A senior intensive-care physician recognizes that a patient’s vital signs are wrong in a way the monitors do not display. An intelligence analyst sees which combination of signals constitutes a pattern before she can fully explain it. A structural engineer reads a building site and calculates risk through trained perception, not just formulas.

None of this knowledge appears cleanly as text. It lives in trained dispositions, situated attention, tacit comparison, embodied routines, and community-specific judgment. It is transmitted imperfectly through apprenticeship, co-presence, repeated cases, correction, and trust. When experts document what they know, the result is a protocol, a memo, a guideline, or a case summary. Those artifacts are useful, but they are not the expertise itself. Documentation is often the residue of expertise, not its generative substrate.

Large language models are extraordinary at the documented layer. They compress and recombine explicit knowledge with breadth and speed that would have been implausible only a few years ago. But the knowledge that makes the most valuable decisions reliable in medicine, defense, law, science, finance, and industrial operations is often tacit, local, and socially governed. It is not absent from current training corpora by accident. It is absent because it was never fully textualized.

Five Architectural Requirements

The gap has five dimensions. They must be solved jointly; solving four while missing the fifth does not produce institutional intelligence.

1. Tacit knowledge capture. A viable system must represent and recover embodied, situated expertise that resists full documentation. Existing paradigms—LLMs, RAG systems, world models, knowledge-management platforms, and most agent frameworks—operate mainly on explicit or externally logged traces. They do not yet provide a principled architecture for tacit knowledge as a first-class computational object.

2. Collective grounding. Expert knowledge is not produced by isolated individuals. It is formed inside communities of practice, with roles, histories, norms, error-correction rituals, and shared standards of judgment. The unit of knowledge is often the expert community, not the single expert. AI systems that aggregate individual contributions without preserving the social conditions of expertise flatten the very structure that

gives knowledge its reliability.

3. Governed attribution. Provenance cannot be repaired after the fact. Longpre et al. (2024)Longpre and colleagues documented license omission rates above 70% across more than 1,800 AI training datasets, showing how easily attribution is lost when data pipelines are optimized for scale rather than governed lineage.² A serious institutional AI system must maintain contributor identity, consent scope, access rights, and downstream dependency tracking as architectural invariants from the beginning.

4. Local epistemic authority. The authority to judge, modify, and apply expert knowledge belongs to the communities and institutions that hold it. A clinical AI system should be governed by clinicians and clinical institutions. A defense system should respect command, classification, and operational context. A vendor should not become the de facto authority over knowledge it did not create and cannot responsibly interpret.

5. Compounding institutional memory. Institutions learn through use: difficult cases, corrections, near misses, practitioner judgment, and shared postmortems. Current AI deployments often remain frozen at training time or are periodically retrained in ways that erase local context. A cognitive infrastructure layer must instead compound: each authorized use should improve calibration, traceability, and institutional competence.

Cognisee’s Architecture

Cognisee’s architecture defines two primitives and one operating principle.

Cognitive Vaults are governed epistemic repositories, not databases. A database stores records and enforces access rules. A Cognitive Vault stores epistemic objects: knowledge with contributor identity, consent status, calibrated confidence, provenance record, and access policy enforced at query time. Each object carries who contributed it, when, in what context, under what consent scope, and for which authorized uses. Consent is revocable. Revocation propagates: if a contributor withdraws authorization, downstream inference outputs that depend on that contribution are flagged and excluded from subsequent operations. Vault contents remain institutionally sovereign and can be deployed on-premise or in air-gapped environments.

Tacit reasoners are governed, locally grounded inference modules. Formally, each reasoner is a tuple $T=(K,C,G,U,\phi )$T, defined as the tuple K, C, G, U, and phi. $K$K is a three-tier knowledge base covering explicit records, implicit behavioral traces, and structured tacit traces: narrated decisions, contrastive case analyses, demonstrations, annotations, and expert corrections. $C$C is a context envelope that bounds competence within a community, role, domain, and permission scope. $G$G is the governance structure that maintains provenance, authorization, update rights, and audit trails. $U$U is a calibrated uncertainty model. $\phi$phi is the composability interface through which reasoners exchange typed, attributed belief outputs without exposing raw tacit content.

The operating principle is AI-in-the-loop, not human-in-the-loop. In human-in-the-loop systems, human oversight is a guardrail around an otherwise autonomous system. In Cognitive Infrastructure, humans and institutions retain primary epistemic authority by design. Expert escalation, refusal, disagreement, and uncertainty are not failure modes. They are first-class operations.

When tacit reasoners compose, they exchange calibrated claims with provenance. They do not merge their knowledge bases. Epistemic sovereignty is preserved. When two reasoners disagree, the disagreement remains visible and attributed. In high-stakes decisions, the disagreement itself may be the most valuable output.

Why This Can Be Venture-Scale

The opportunity is not another enterprise chatbot. It is the control layer for institutional cognition. A platform that can govern tacit expertise has three compounding loops.

• Data loop: each deployment captures knowledge that competitors cannot scrape from the public internet and customers cannot easily document on their own.
• Trust loop: provenance, consent, revocation, and local governance make adoption possible in domains where generic AI tools remain blocked by risk.
• Reasoner loop: specialist reasoners improve with local use and become more valuable when composed into collective systems.

This creates a credible path from vertical wedges to horizontal infrastructure. Initial deployments can focus on domains where expert judgment is scarce, expensive, consequential, and poorly documented: clinical decision support, regulatory compliance, defense analysis, scientific R&D, financial diligence, and high-precision industrial operations. Over time, the same primitives become an enterprise cognitive operating layer: the place where an institution stores, governs, queries, and compounds what it knows.

For a frontier investor, the key question is defensibility. Cognitive Infrastructure has defensibility where generic model wrappers do not: proprietary zero-party expert data, consented institutional relationships, governance primitives embedded at the object level, compounding usage histories, and multi-reasoner networks whose value increases with breadth and depth of deployment.

What Remains to Be Proven

The central claims are research hypotheses, not completed facts. That is a strength if the program is built with explicit falsification. Three hypotheses matter most.

First, structured tacit elicitation should produce knowledge objects that expert panels judge more faithful than ordinary self-authored documentation. Second, governed institutional

memory should compound through use, improving calibration and traceability without catastrophic forgetting. Third, heterogeneous tacit reasoners should compose into collective capabilities that exceed what individual reasoners can provide while preserving uncertainty, provenance, and authority boundaries.

The first evidence program is deliberately minimal. One experiment compares Cognitive Task Analysis with standard documentation in a within-subjects study of senior emergency physicians. A second tests whether a Cognitive Vault blocks unauthorized access across prompt injection, role escalation, indirect leakage, and policy-ambiguity attacks. A third evaluates whether a specialist reasoner over a curated expert vault can match a frontier generalist on a 100-question regulatory-compliance battery at lower compute cost, with full provenance tracking.

Each experiment produces useful information whether the primary hypothesis is confirmed or rejected. The point is not to assert inevitability. It is to turn a major enterprise-AI bottleneck into a measurable scientific and commercial program.

Why Now

The model layer is powerful, increasingly accessible, and increasingly protocol-mediated. Anthropic’s Model Context Protocol introduced a standard way to connect AI systems to external tools and data sources, and Google’s Agent2Agent protocol defined a communication fabric for agents to exchange information and coordinate across enterprise platforms.³ The transport layer is emerging. What remains missing is the governed knowledge layer that makes transport valuable.

The timing is therefore precise. Enterprises are moving beyond pilots. Agentic systems are entering production. Boards and executives are asking where durable value will come from. The answer cannot be model access alone. It must be the conversion of private, tacit, institutionally governed knowledge into systems that reason, remember, and improve under local authority.

That is the ambition of Cognitive Infrastructure for Artificial Collective Intelligence: not to replace expert communities, but to make their knowledge computable, protected, composable, and compounding.

What Frontier-Scale Capital Builds

A financing at frontier-lab scale is justified only if it builds infrastructure, not a project. For Cognisee, the capital formation logic is to assemble four assets that reinforce one another.

• The expert-data network: long-term institutional partnerships that generate consented, high-value tacit traces unavailable to public model providers.
• The governed platform: Cognitive Vaults, policy engines, audit systems, deployment controls, and security evaluations strong enough for regulated institutions.
• The reasoning laboratory: a research group building specialist tacit reasoners, uncertainty calibration, multi-reasoner composition, and compute-efficient expert inference.
• The vertical proof engine: early wedges in domains where better judgment has immediate economic and strategic value, then expansion into horizontal cognitive infrastructure.

The company-defining asset is the compounding relationship between proprietary tacit knowledge, institutional trust, and governed reasoning. If this loop is demonstrated, Cognisee is not a services company and not a wrapper around a model. It becomes a candidate control plane for enterprise cognition.

Footnotes

1. Boston Consulting Group, The Widening AI Value Gap: Build for the Future 2025, September 2025. ↩

2. S. Longpre et al., “A large-scale audit of dataset licensing and attribution in AI,” Nature Machine Intelligence, 2024. ↩

3. Anthropic, “Introducing the Model Context Protocol,” November 2024; Google Developers Blog, “Announcing the Agent2Agent Protocol,” April 2025. ↩