Part 1 of 5: The RAG Series. The Economics of Knowledge in the Firm

Why do firms keep reinventing the wheel? The pattern is familiar: a team spends weeks on a problem, only to discover afterwards that a colleague solved something nearly identical two years ago. The prior work exists—in a folder somewhere, in someone’s email, or in a departed employee’s head. But it might as well not exist, because nobody could find it.

This is not a technology problem, or not only one. It is an economic problem. Understanding when it pays to capture and reuse organisational knowledge—and why most attempts to do so fail—requires thinking carefully about costs, benefits, and incentives.

The Basic Proposition

The useful starting point is Paul Romer’s work on economic growth (Romer, 1990). The core insight: growth comes not just from accumulating resources but from finding better ways to combine them. Ideas—recipes for arranging inputs—are non-rival. If one person uses an idea, others can use it simultaneously. This property creates the theoretical possibility of increasing returns: a firm that develops a useful method can deploy it repeatedly without using it up.

But most firms fail to capture these returns. The reason is that while knowledge itself is non-rival, everything required to use knowledge is rival. Deploying prior work requires someone’s time to find it, attention to understand it, judgment to assess whether it fits, and effort to adapt it. These are all scarce. A framework document does not execute itself.

The firm only achieves increasing returns if it can spread knowledge-creation costs across many uses while keeping deployment costs low. If every reuse requires significant adaptation or senior review, deployment costs scale linearly with use. The creation cost gets amortised, but the economics don't fundamentally change.

A Decision Framework

When does it make sense to codify knowledge? The decision can be framed as a cost-benefit calculation:

(Frequency × Value × Extraction Efficiency × Trust) > (Codification Cost + Maintenance Cost + System Cost)

Frequency is how often similar situations arise. Value is what each reuse is worth—typically the cost of recreating the knowledge, plus any quality or speed premium. Extraction efficiency is the fraction of potential value actually captured, given how knowledge is stored and accessed. Trust is whether people will use what they find; if the knowledge base has a reputation for being outdated, people won't search.

Extraction efficiency deserves particular attention. Research suggests knowledge workers spend approximately 1.8 hours daily—around 20% of the workweek—searching for information (Panopto, 2018). For a 100-person professional services firm, this represents substantial opportunity cost. The difference between 25% extraction efficiency (manual search, read, adapt) and 70% (system-assisted retrieval and adaptation) is not marginal.

The Incentive Problem

The equation above describes what benefits the firm. But codification decisions are made by individuals, and their calculation differs.

Consider a consultant who has developed an effective approach to a recurring problem. Writing it up costs her time immediately and with certainty. It requires effort to sanitize the data, format the insight, and tag it correctly—hours that could be spent on billable work or rest. The benefits—recognition, perhaps time saved on future projects—are uncertain and flow partly to others. Worse, codifying expertise can reduce bargaining power: if her knowledge is now available to everyone, she becomes more replaceable.

This is a standard collective action problem. The cost (time and effort) falls on the individual; the benefit flows mainly to the firm. It explains why exhortation ("share your knowledge!") rarely works. The incentives are misaligned.

What Went Wrong Before

Knowledge management had a significant wave in the 1990s. Firms hired Chief Knowledge Officers, built intranets, and invested in systems like Lotus Notes (Davenport and Prusak, 1998). By the early 2000s, most initiatives had failed. Gartner removed knowledge management from its Hype Cycle in 2007, declaring it "mature"—which in practice meant the excitement had collapsed (O'Leary, 2016).

The failure wasn't necessarily that the technology didn't work at all—SharePoint and email search became ubiquitous utilities. The failure was that the economic promise was overstated. These tools succeeded as digital filing cabinets (we stopped losing physical paper), but failed as synthesis engines. They did not dramatically lower the cost of finding and adapting ideas, so the "reinvention of the wheel" continued.

Where AI Changes the Economics

Given this history, where does AI fit? The honest answer: it changes some terms in the equation significantly, others barely at all.

Extraction efficiency is where AI has clearest impact. Traditional keyword search fails when users cannot guess the right terms—the relevant document exists but cannot be found. Semantic search, which matches by meaning rather than keywords, substantially reduces this problem.

But semantic search alone is insufficient. Research on retrieval-augmented generation (RAG) systems reveals persistent challenges. Without careful architecture, legal and technical RAG tools exhibit hallucination rates of 17-33%. These are often "sycophantic" errors: the system retrieves text that looks relevant but is factually distinct.

The engineering responses to these problems are increasingly well-understood:

• Hybrid search combining semantic and keyword retrieval reduces "vector drift" where queries pull toward generic content.

• Hierarchical retrieval (querying document summaries before chunk-level search) filters noise before it reaches the ranking stage.

• Semantic reranking improves precision but operates within hard limits—Azure's reranker, for instance, only processes the top 50 initial results.

• Data hygiene—deduplication, handling version conflicts, managing boilerplate—addresses the "hubness" problem where generic text clusters in vector space and crowds out specific answers.

None of this is magic. The phrase "garbage in, garbage out" applies with particular force: pointing a model at messy data does not extract intelligence. But with appropriate architecture and data curation, extraction efficiency gains of 2-3x over baseline keyword search are achievable with current technology.

Codification cost can be reduced through passive extraction—pulling knowledge from existing work products (emails, documents, meeting transcripts) rather than requiring separate documentation. This is promising but raises privacy concerns that are not trivial.

Maintenance gets partial help. AI can flag outdated content and prompt reviews. But detection is not the same as correction; someone must still update the material.

Incentives do not change. AI does not solve the collective action problem. If individuals bear the costs while the firm captures the benefits, the dynamic remains.

Trust can go either way. Effective retrieval builds trust; systems that surface wrong or outdated content destroy it. The hallucination problem is real and must be managed through confidence thresholds and epistemic safeguards—returning "I don't know" when retrieval confidence is low, rather than generating plausible-sounding errors7.

When Investment Makes Sense

The framework suggests conditions under which knowledge retrieval investment pays off:

• Extraction efficiency is the binding constraint. If the firm has valuable knowledge that people cannot find or adapt efficiently, improving retrieval directly increases the equation's value side. If the constraint is elsewhere—knowledge doesn't exist, or the domain changes so fast that maintenance costs dominate—retrieval technology helps less.

• A meaningful knowledge stock exists. AI search on an empty or poorly-maintained repository yields little. The precondition is that useful knowledge has been created and is at least minimally documented.

• The organisational problems are at least manageable. If incentive misalignment is severe, if there is no ownership of maintenance, if trust has already collapsed, technology investment may be premature. These issues do not need to be fully solved, but they need to be addressed alongside the technology.

• The firm can sustain the investment. Retrieval systems require ongoing maintenance, curation, and adaptation. A one-time implementation that degrades over time repeats the 1990s failure pattern.

Where these conditions hold—professional services firms with recurring problem types, substantial document stocks, and margin pressure driving interest in leverage—the investment case is strong. Where they do not, other interventions may be higher priority.

Open Questions

Several aspects remain unresolved:

• The incentive problem has no clean solution. The best approaches combine multiple partial fixes. A satisfying general answer has not emerged from either the literature or practice.

• The boundary between codifiable and tacit knowledge is unclear. Some expertise degrades when made explicit—checklists derived from expert judgment sometimes perform worse than the original intuition (Kahneman and Klein, 2009). Identifying in advance which knowledge falls into this category is difficult.

• Sector differences matter but are under-theorised. The economics differ substantially between professional services (high frequency, high value, project-based discontinuity) and manufacturing (operational knowledge, slower change, different retrieval needs). The framework applies generally, but optimal interventions differ.

References

• Barnett, S., et al. (2024) 'Seven Failure Points When Engineering a Retrieval Augmented Generation System', arXiv preprint.

• Davenport, T.H. and Prusak, L. (1998) Working Knowledge: How Organizations Manage What They Know. Boston: Harvard Business School Press.

• Kahneman, D. and Klein, G. (2009) 'Conditions for Intuitive Expertise: A Failure to Disagree', American Psychologist, 64(6), pp. 515-526.

• Luan, Y., et al. (2021) 'Sparse, Dense, and Attentional Representations for Text Retrieval', Transactions of the Association for Computational Linguistics, 9, pp. 329-345.

• Microsoft (2024) Service Limits for Tiers and SKUs - Azure AI Search. Available at: learn.microsoft.com [Accessed 4 Jan 2026].

• O'Leary, D.E. (2016) 'Is Knowledge Management Dead (or Dying)?', Journal of Decision Systems, 25(sup1), pp. 509-530.

• Panopto (2018) Workplace Knowledge and Productivity Report. Available at: panopto.com [Accessed 4 Jan 2026].

• Romer, P.M. (1990) 'Endogenous Technological Change', Journal of Political Economy, 98(5), pp. S71-S102.

• Storey, J. and Barnett, E. (2000) 'Knowledge Management Initiatives: Learning from Failure', Journal of Knowledge Management, 4(2), pp. 145-156.