StandpointKnowledge Management
AI gives answers. Your company needs the right ones.
Language models are becoming interchangeable. Your company’s legal knowledge is not. Why knowledge management is the infrastructure question for every business that works with contracts.
In the mid-nineties, Lew Platt, then CEO of Hewlett-Packard, summed up the problem in one sentence: “If HP knew what HP knows, we would be three times as profitable.”1 Three decades later the sentence still holds, but for a different reason. Answers have become cheap, now that any company can rent a capable language model. What has become scarce is the knowledge Platt meant. Only your own company has it.
Hence my thesis. Knowledge management, for many years a long-term topic for corporate staff functions, is becoming an infrastructure question for every company that earns its money with contracts. Not although AI exists, but because it does.
What searching costs
The finding has been stable for twenty years. The McKinsey Global Institute put it at roughly a fifth of the workweek in 2012: time knowledge workers spend searching for internal information or tracking down the colleagues who have it. A YouGov survey commissioned by Panopto (2018, n = 1,001) arrived at 5.3 hours per employee per week lost to waiting for knowledge or reconstructing it. Extrapolated to a large US business, that is 47 million dollars of lost productivity per year. And the most recent large survey by a KM vendor (Bloomfire, 2025, over 10,000 respondents across 115 companies) puts search time at 8.5 hours per week without structured knowledge management against 4.6 with it. That is almost half the searching gone, arithmetically the capacity of 98 full-time employees per thousand staff.
Two of the three studies come from vendors selling software for precisely this problem. I cite them as what they are, namely vendor studies. What remains remarkable is that independent and interested sources have pointed in the same direction for two decades, and that nobody measures the opposite.
In contract work, the finding shows particularly clearly. The legal department answers the same question about a penalty clause three times a quarter, because the answer sits in three inboxes. Procurement renegotiates a liability clause for which an agreed position existed long ago; nobody can find it. And when the contract manager leaves, her playbook leaves with her, because it was never written down anywhere. That the gap begins at the fundamentals was recently shown by an empirical study among employees of commercial enterprises.2 On core questions from the duty to give notice of defects to Incoterms®, only a minority answered correctly. I discussed the study in the article “Commercial law as a knowledge gap in companies”. Training closes the gap in what employees know. The second half of the task is the knowledge that has to stay in the company when people leave.
Answers become a commodity, context does not
What changed in 2025 and 2026 is not that companies suddenly document better. What changed is who reads the documents, and what happens without them.
What it costs when the context is missing was measured in July 2025 by a study from the MIT project NANDA (“The GenAI Divide: State of AI in Business 2025”). Against 30 to 40 billion dollars of enterprise investment in generative AI, 95 percent of organizations see no return at all. The difference lies not in model quality or regulation but in approach. The authors explicitly name missing context and systems that do not learn as the bottleneck, not infrastructure, regulation, or talent. The figures are labeled preliminary and rest on 300 publicly documented initiatives, 52 interviews, and 153 surveyed executives. They are no basis for precise figures, but they are one for the direction.
Microsoft’s CEO Satya Nadella, of all people the largest landlord of AI models, classified the models themselves as interchangeable on his India tour in December 2025: there are “lots and lots of capable models”, models alone are “not sufficient”, and the strategic value sits in the data and experience layer above them (report). When someone who sells models tells you the model is not the point, you may believe him on this one. The capital side puts it even more directly. Insight Partners, one of the large growth investors, calls access to specific, often messy company data “one of the strongest moats in AI”. The same shift is putting the legal profession’s own pricing model under pressure. That is the subject of the article “The billable hour is out of time”.
In parallel, the audience is changing. GitBook measures on its own platform that in 2025, AI read access to documentation rose by over 500 percent. In the accompanying State of Docs Report, a quarter of the documentation professionals surveyed expect documentation to be written primarily for machines in the future. Both figures come from a platform vendor, but the direction is unambiguous all the same. Machines are the fastest-growing readership.
With contract knowledge, the playbook that helps your procurement team today becomes tomorrow’s context that keeps your internal AI assistant from inventing a liability clause. Technically the pattern is called retrieval-augmented generation. The model answers not from memory but from your documents. A language model without that foundation answers anyway. It answers fluently and politely, but wrongly. With it, the same question returns a correct answer with a source.
How reliable the models have become in law is a measured question. Conrads and Schweitzer put 200 cases from contract, commercial, and employment law to the leading systems over two years.3 Hit rates rose from 50 to over 70 percent. In commercial law they remain the weakest at 71 percent. The reason is that commercial questions do not turn on standard formulas but on case-specific, context-dependent facts and institutional background knowledge. A real case also consists of several steps built on one another. The individual rates multiply. At 75 percent per step, three steps leave 42 percent and four leave 32 percent. Missing context therefore turns the tool into a risk.
That reverses the economics. As long as answers were expensive, you could afford unstructured knowledge. The bottleneck was the expert anyway. Now the bottleneck is the material. Whoever does not have their contract knowledge in structured form cannot use the cheapest resource of the decade.
Diagram: three layers of knowledge with decreasing retrievability. General law, standard clauses, and definitions are fully retrievable and come with the model. Your own templates, checklists, and process descriptions exist but are scattered. Clause positions, negotiation limits, and escalation thresholds are in people’s heads only.
Why knowledge repositories regularly fail
The obvious objection is that you already have an intranet for that. You do, and that is exactly why you know how it turns out. After two years the repository is an archive nobody trusts. That is not your team’s fault. Generic repositories fail at two points, always the same ones.
First, structure. A page meant to explain, instruct, and serve as a lookup at the same time does none of the three. The software industry has had an ordering principle for this problem for years, Diátaxis. Under it, every page answers exactly one kind of question. For contract knowledge, that gives four content types:
| Content type | The user’s question | Example in everyday contract work |
|---|---|---|
| Getting started | Guide me. | The first contract check with the playbook, step by step. |
| How-to guide | Help me with this task. | Defect discovered in a sale between merchants: give notice without undue delay under Section 377 of the German Commercial Code (Handelsgesetzbuch, HGB), with a model letter. |
| Reference | What applies? | Clause library: liability clause with agreed positions and escalation stages, German and English. |
| Background | Why is it this way? | Why excluding the CISG across the board is often the worse choice. |
This separation decides whether the sales manager under time pressure finds the how-to instead of working through a treatise. It also decides whether the machine finds an unambiguous, citable source instead of unstructured prose.
Second, ownership. Content without a professionally responsible curator goes stale, and stale contract knowledge is more dangerous than none, because somebody trusts it. A platform loses its users’ trust the moment the first wrong page is noticed. Maintenance is therefore the actual product. Case law, reforms, and new clauses have to be worked in, quarterly and event-driven, by someone who commands the field and stands behind the content.
What “machine-readable” means in practice
Machine readability cannot be added afterwards. It emerges from the same decisions that make knowledge usable for people. One page per question, so every answer has an address. Structured text instead of a PowerPoint archive and a PDF folder, so search and language model find the same source. Citations on every statement, so the answer stays verifiable, for the human and for the machine. Versioning, so it is traceable what applied when; for contracts not a formality but the question of which version governed at signing. And graded access rights, because not every playbook belongs in every chat.
That last point is not an IT question. Under Section 2 No. 1(b) of the German Trade Secrets Protection Act (Gesetz zum Schutz von Geschäftsgeheimnissen, GeschGehG), information remains a trade secret only as long as it is the subject of confidentiality measures appropriate to the circumstances. The Higher Regional Court of Hamm requires measures that are actually enforced.4 A platform concentrates precisely the knowledge that needs this protection: costing bases, negotiation limits, supplier terms. Run it without graded rights and you risk the protection not for one page but for the entire body of knowledge. How far the requirements reach I have worked through in the article “Effective protection of trade secrets”.
That is why I build knowledge platforms as versioned, searchable text systems and not as slide collections on a shared drive. Structured text is the only format that serves both readers equally well: the buyer on Monday morning and the AI assistant he asks next year.
From answer-giver to architect
For my profession this means a change of role. When the individual answer becomes a commodity, the value of legal work shifts to where the answers come from. That is the maintained knowledge base from which your team draws correct answers today and your systems draw them tomorrow. The lawyer who only sells answers will be competing with a flat rate. The lawyer who structures, curates, and keeps current his client’s contract knowledge builds infrastructure that is useful between mandates as well.
The platform should hold playbooks for the most frequent negotiation situations and a clause library with agreed positions. Add to that how-to guides for processes with deadlines and form requirements as well as background pages for everyone who wants to understand. Plus operations that fit the subject matter: European infrastructure instead of a shared US cloud, processing on your behalf under Article 28 of the General Data Protection Regulation (GDPR), German and English, readable for people and machines.
The platform does not stand alone. It is the middle of three stages: advice generates the knowledge, the platform preserves it, training passes it on to the staff. Every mandate improves the playbooks; every training session shows which page is still missing. Knowledge that is only produced is an expense. Knowledge that is preserved and trained is an asset.
Since 2 February 2025, a legal duty has also applied to the third stage. Anyone deploying AI systems under their own authority is a deployer within the meaning of the AI Act. Under Article 4(1) of Regulation (EU) 2024/1689, they must take measures to support the development of AI literacy. This covers their staff and other persons dealing with the operation and use of AI systems on their behalf. In doing so, they must take into account, among other things, the context in which the AI systems are to be used. They do not have to guarantee any specific level of AI literacy. Until 26 July 2026 they had to ensure, to their best extent, a sufficient level of AI literacy among their staff and those other persons. That is what the provision required before Regulation (EU) 2026/1744 amended it. Chibanguza and Steege read the provision in that earlier version as a duty to instruct and train.5 For your own staff, I recommend material that is current, fits the deployment context, and sits where the teams actually work. That is exactly what the platform is.
When a platform pays off, and when it does not
Not every company needs this. If you conclude a handful of uniform contracts a year, a good template package and a lawyer you can reach quickly serve you better. A knowledge platform pays off when several factors come together: multiple teams working with contracts without being lawyers. Recurring questions to the legal department or outside counsel. Noticeable turnover in key roles. International contracts where the knowledge gap gets expensive. And the plan to deploy internal AI assistants seriously, instead of letting them hallucinate on an empty intranet.
If three of these apply, the question is no longer whether but where to start. The place to start is almost always the three topics where not knowing has recently cost the most.
Platt presumably meant the tripled profit rhetorically. The arithmetic behind it holds all the same. The distance between what a company knows and what it uses is one of the most expensive gaps in the organization. What is new is that this gap can be closed with less effort than ever. What is also new is that for the first time there are machines that turn structured knowledge directly into productivity. The companies that start now gain a head start. They use their knowledge while everyone else is still searching.
Frequently asked questions
What is legal knowledge management?
A maintained, searchable system that makes a company’s contract knowledge available in structured form: playbooks, clause libraries, how-to guides, and background pages, written for the teams that work with contracts every day, not for lawyers.
Does an AI assistant replace the knowledge platform?
No, it presupposes one. A language model answers every question, including wrongly. It only becomes reliable when it draws on curated, current company knowledge. The platform is exactly that foundation, readable for people and machines.
Where do you start?
With discovery: the contract landscape, the recurring questions, the three topics where not knowing caused the most damage. The result is a prioritized plan. A focused start with one area of law and one team is live within a few months.
Notes
Reference: Poleacov, P. (2026). AI gives answers. Your company needs the right ones.. INN.LAW. https://inn.law/en/perspectives/knowledge-management-in-the-age-of-ai/