The Context Problem

In 1934, T.S. Eliot wrote two lines in The Rock that have proved more durable than most management theory: “Where is the wisdom we have lost in knowledge? Where is the knowledge we have lost in information?”¹ He was not writing about knowledge management. He was writing, more broadly, about a civilization increasingly preoccupied with accumulation, measurement, and progress, yet in danger of forgetting the purposes those things were meant to serve. Information could become a substitute for knowledge, knowledge a substitute for wisdom, and the act of gathering obscure the reason for gathering at all.

Half a century later, management theorist Russell Ackoff would express a related idea in organizational terms. In a 1989 paper, From Data to Wisdom, he described a hierarchy through which raw facts become progressively more valuable: data, information, knowledge and, at the apex, wisdom.² The model’s central insight was not that organizations need more data, but that each step upward requires interpretation, context, and judgment. Data can reveal patterns. Information can provide meaning. Knowledge can guide action. Wisdom asks a different question altogether: whether the action is worth taking, what consequences it might create, and what trade-offs it conceals.

The challenge is that most organizations are designed to produce information, not wisdom. KPIs, dashboards, quarterly targets, utilization metrics, panel scorecards and procurement frameworks are highly effective at generating information. They are often far less effective at producing the judgment needed to decide what matters, what doesn’t, and what happens next.

The distinction matters because every major wave of legal technology over the past thirty years has promised to improve how information is captured, organised, retrieved, or analysed. Far fewer have addressed the question of how context and judgment are created, preserved, and transferred.

By the early 2000s, the Data-Information-Knowledge-Wisdom (DIKW) pyramid had become the dominant conceptual framework of the legal knowledge management movement. It appeared in conference presentations, practitioner papers, and CKO presentations at firms that were, for the first time, taking seriously the question of what they actually knew and how to make it findable. The technology was improving. Recommind, Autonomy, and their successors used AI — conceptual and natural language search — to make it easier to retrieve documents from sprawling firm repositories. Enterprise search was gaining traction. The mood was cautiously optimistic.

And yet the central problem refused to yield. At the Interwoven Legal IT Leadership Summit in May 2006, Ron Friedmann (then of Prism Legal Consulting) gave a presentation that identified what he called perhaps the biggest challenge in legal KM. “Documents without context,” he wrote in the published version, “— information about the case or the deal — may not be that useful.”³ The document could be found. What the document meant — why it was drafted the way it was, what had been negotiated away, what the client’s specific circumstances had required — remained invisible.

The legal KM movement, enabled by increasingly sophisticated search and retrieval technologies, had become very good at moving from Data to Information. It had made progress toward Knowledge. The move toward Wisdom remained largely unsolved.

Outside the United States, Practice Support Lawyers (PSL) were the human attempt at addressing that problem. Non-billable lawyers, typically with many years of practice experience, whose job was not just to find good documents but to annotate them, to capture the matter context that gave a precedent its meaning, to build (and update as legislation changed) playbooks, to distinguish between documents with precedential merit and those without. They were, in DIKW terms, one of the legal industry’s most deliberate attempts to move from Knowledge toward Wisdom — to add to the “what” and the “how” the “why” and the “why-not”.

Inside the United States, as Friedmann observed, few large firms had the appetite to employ them in significant numbers. The response was to automate instead: better search, better retrieval, better metadata; the bottom of the pyramid made more efficient, the top left alone.

By 2018, Friedmann noted that the field had effectively abandoned the attempt to capture context in documents and had shifted its attention to finding the people who held it.⁴ The expert could identify the useful document and explain what it meant. It was a pragmatic response to an unsolved problem — and it told you something important about the nature of the problem itself.

The context was never primarily in the documents. It was in the people who produced them. The documents were the output. The reasoning was elsewhere, in conversations that were never recorded, in email threads subsequently deleted by retention policies, in the accumulated judgment of partners who eventually retired and took it with them.

Eight years later the legal technology market has moved on — through document assembly, through contract lifecycle management, through the first wave of machine learning applied to legal documents, and now into generative AI and agentic systems. The tools have become dramatically more capable, yet the underlying problem has not changed. The context was never in the documents in 2018, and it is not in the documents now. And the current wave of investment — larger, faster, and more confident than anything that preceded it — is built on the same incomplete foundation that Friedmann identified nearly two decades ago, with considerably more sophisticated tools for retrieving from it.

A large language model learns from the documents. The Retrieval-Augmented Generation (RAG) system retrieves from the documents. The Model Context Protocol (MCP) connector links the model to the systems where the documents live. What MCP does not do is make the data inside those systems complete, consistent, or authoritative — it cannot resolve identities across systems that don’t share a common identity layer, correlate a counterparty’s actions across correspondence, court filings, and regulatory submissions, or tell you which version of a document reflects which negotiating position and why.

But fragmentation is only part of the problem. As Rory Sutherland observed in Alchemy, “all big data comes from the same place: the past. Yet a single change in context can change human behaviour significantly.”⁵ The corpus that these systems retrieve from is entirely historical. It contains what was captured. It cannot contain what wasn’t. And the most commercially significant context in legal work — the reasoning behind the clause, the judgment call on the deal structure, the pattern of how this counterparty behaves when they’re under pressure, or the exogenous business need that created that pressure — is precisely the context that tends not to survive into the final executed document.

Tom Baldwin — founder of Entegrata, and formerly CIO/CKO at Cadwalader, Reed Smith and Sheppard Mullin, with thirty years inside the data and knowledge management functions of major law firms — made the technical version of this point precisely in response to Anthropic’s Claude for Legal announcement. The MCP connectors that generated so much enthusiasm don’t solve the underlying problem, he wrote, because “the data inside each system is still fragmented, and I’d wager that in most cases the work to make that data contextually relevant hasn’t been done.”⁶ Connecting systems to a language model creates what he calls “a tidal wave of partial context” — output that may or may not be right depending on the complexity of the query and the completeness of what’s being retrieved. In an industry where, as he puts it, certainty prevails, “may or may not” is simply not good enough.

The attempt to solve this through correlation of internal systems is not new. Matter-centric document management — the architecture that iManage pioneered, using the client and matter identifier as a correlating key across emails, document versions, edits, and time entries — was the most serious earlier attempt. In principle, if everything was tagged correctly against the same matter, you could reconstruct something close to the full history: who communicated what, when, alongside which document version, against what billing entry. In practice, tagging discipline was inconsistent, retention policies deleted the most contextually valuable records before they could be correlated, and the conversations that mattered most — the phone call, the corridor exchange, the judgment call made before the email was sent — were never in the system to begin with.

The most sophisticated current efforts to solve the correlation problem focus on connecting more data sources and surfacing more context. Platforms such as Palantir’s Foundry can correlate internal records with external signals, creating a richer picture of events and relationships than traditional document repositories ever could.⁹ It is a meaningful advance. But it remains bounded by the same constraint: it can only work with what was captured. The judgment that existed only between the people involved, the reasoning that never made it into a system, and the conversation that happened before the email was sent remain largely invisible.

This is not an argument that GenAI cannot assist with the context problem — it can, in ways that weren’t available even five years ago. Jenn McCarron, co-founder and CEO of Contracts.ai who led legal operations and technology at Netflix and legal technology at Spotify before that, has built her company on precisely this capability: using AI to extract knowledge from post-signature contracts at scale — the obligations, rights, termination triggers, and commercial terms that are embedded in executed agreements but rarely surfaced in a form that the business can act on.⁷ That is material and moves AI’s reach meaningfully up the DIKW pyramid from the data and information layers where it has operated most comfortably. It represents one of the most promising applications of GenAI to the context problem.

But it operates on what exists. The final executed agreement, the structured commercial terms, the patterns visible across a portfolio of signed contracts; these are the inputs, and from them, knowledge can now be extracted at a scale and speed that no PSL team could match. What it cannot supply is the context that was never in the document in the first place: the warranty that moved to a side letter or the negotiating history that unfolded across the boardroom table or email threads subsequently deleted by retention policy. The older the document, the more acute this problem becomes — the people who held the context have moved on, the email threads that explained the drafting choices have been deleted on schedule, and what remains in the corpus is the output of decisions whose reasoning has been systematically erased.

McCarron’s broader market observation takes this further. When Anthropic made the underlying model a commodity and Microsoft put contract drafting inside Word, “the differentiation that legal tech companies were selling at that layer collapses.”⁸ The application layer — the drafting interfaces, the workflow automation, the AI chat panels — is where most legal tech companies have been building and selling. It is also the layer that has just been commoditised from above and below simultaneously.

What remains, she argues, is the data layer: how contracts get turned into structured, queryable, vectorised data that AI systems can actually act on. But the data layer she identifies as the new source of value is the same layer Friedmann identified as the biggest KM challenge in 2006. It is the layer PSLs were positioned to build, the work that, in the US especially, few firms ever resourced at the scale the problem required. It is the layer that document retention policies have been systematically eroding by deleting the email trails that might have contained the context the documents themselves never held.

The firms that will win with AI, Baldwin argues, won’t be the ones who adopted the best AI operator or legal assistant tool first. They’ll be the ones who did “the strategically necessary but unglamorous work of getting their data platform in order. That work doesn’t tend to make the headlines, but it’s what everything else is built on.”

That work is, in DIKW terms, the construction of the Knowledge layer — the conversion of the raw data in firm document repositories into structured, contextualised, annotated information that a model can actually reason from, rather than merely retrieve. It is the work the PSLs were doing, at human speed, with human judgment. It is the work that thirty years of enterprise search, document assembly, and KM platforms promised to automate — and didn’t, because the automation could move data around but couldn’t supply the judgment about what the data meant.

There is a practical way to test this. Before signing the next enterprise AI contract, ask yourself what a model would actually know if it queried your systems tomorrow. Would it understand why the clause was drafted that way? Why the negotiation took the turn it did? Why this precedent mattered and three others didn’t? Or would it simply retrieve the artefacts left behind by decisions whose reasoning has long since disappeared? If the latter, no connector will retrieve what isn’t there.

There is a further dimension to the context problem that no data layer, however well constructed, can reach — one that goes beyond what was captured and into what could never have been captured at all. That is the subject of the fourth and final piece in this series.