Do Wolters Kluwer’s divisions withstand AI disruption?

In the early nineties, I was responsible for building OPmaat, an alerting service for the Official Publications of the Dutch Central Government, the first online electronic publishing product of (at the time) Sdu Uitgevers. After thirty years, OPmaat is still running, despite many technological waves. This fact colors my view of disruption: curated information services are tougher than stories of disruption suggest. I wanted to test whether my idea about this is realistic. How that unfolded with the help of AI is what I want to show here.

The impetus for this research was not an investor question, but a commercial one. Suppose you wanted to build an integration layer between a client’s AI and the curated content of a professional content publisher; what would you want to know from such a publisher? At the very least, whether it will still have a right to exist in the future. That answer leads to an investment decision. Also on the part of the provider of such an integration service.

So I dove into a current case: Wolters Kluwer (WKL). On February 3 of this year, Anthropic announced a legal AI application. The share price fell sharply on that day (12.7%), in a market where Anthropic’s announcement was interpreted by investors and analysts as a potential threat to legal information services. Industry peer Thomson Reuters also lost 15.7% that day, and RELX 14.1%. On May 6, Wolters Kluwer released a quarterly update that met expectations. The stock dropped another 17.2 percent, this time after JPMorgan warned of erosion of WKL’s workflow products.

A drop of 58% from a peak of €178.70 (Feb 14, 2025) to €70.08 now (Sep 17, 2026) is certainly food for thought. Especially when earnings per share are actually rising. Could the market be anticipating a possible decline in Wolters Kluwer’s earning capacity? The term “AI disruption” was frequently seen in the reports of analysts covering WKL.
I decided to investigate this using AI and expected that the benefit of AI for me personally would primarily lie in gathering information. That was indeed correct. Within a few weeks, there was more material on the table than I could have gathered in such a short time without AI. But then what? The really interesting part only happened after that.

You can query a language model (LLM) endlessly based on the collected information. And the thing really does reason, if you would call it that. And to my surprise, often astonishingly sharply too. But did those arguments hold water? My initial (implicit) working hypothesis was that AI disruption risk is a property of a company, in this case Wolters Kluwer. What led me astray from that track was a contradiction.

The Legal & Regulatory division was the primary target of Anthropic’s announcements, which caused at least part of the share price drop. If the disruption narrative was true, that very division should have been the first to fall back. But in the first half of 2026, it actually grew five percent organically. A single risk score for the entire company could not explain that. Wolters Kluwer reports five business units with distinctly different markets, products, and customer relationships. I soon discovered that a holding company with five different divisions is not exposed to just one disruption risk. There are several, and they are totally different in nature. Legal & Regulatory, the challenged division, represents only 16.4 percent of revenue and generates an adjusted operating margin of approximately 18.2% (financial year 2025). It sells workflow software to professional customers that allows them to consult a large body of legislation and case law. A significant portion of the source material is publicly available, making the pure search and summary function vulnerable. This does not necessarily apply to editorial enrichment, annotation, metadata, and integration.

Much more important for WKL is Health, accounting for 26.1% of revenue and approximately 30% adjusted operating margin. Its main product is UpToDate for decision support. The use of UpToDate can contribute to the traceability and substantiation of a clinical decision, thereby arguing compliance with a professional standard. A generic model cannot offer this as long as UpToDate remains a de facto standard, even if that standard is not universal.

The Tax & Accounting and Financial & Corporate Compliance divisions also rely on workflow products (CCH Axcess and CT Corporation, respectively) that are deeply embedded in the client’s processes. Together, these represent approximately 47% of revenue and generate a margin of over 30%. Finally, the small Corporate Performance & ESG division has also established itself within the client environment with its platforms Enablon and CCH Tagetik.

After much ‘discussion’ with my AI, the following analytical framework emerged: AI disruption risk depends on the protective layers that shield a product or service from competition. A single company (or division) often has multiple layers simultaneously. These characterize the barrier to entry.

• Layer 1 is capability. The barrier is ability: what a generic model can do with material it has access to: finding, summarizing, establishing connections, generating meta-information. This is the fastest-decreasing barrier, and the only one that can be measured publicly: every model release from a major AI lab is a benchmarkable event.
• Layer 2 is access. The barrier is ability to access it. It concerns reliable and up-to-date source data that a competitor cannot access. For example, because part of Dutch case law is not made public, because it is not included in a searchable database, or because meta-information is legally protected, as Wolters Kluwer does through its terms of use. For certain forms of text and data mining, Wolters Kluwer can invoke Article 4 of the European Directive, which regulates the reservation of rights.
• Layer 3 is status. The barrier is recognition. Here, institutional trust, professional legitimacy, accountability, and potentially liability apply. Even with the same model capability and access, a language model cannot provide this, because the value lies in who is responsible for the answer. In Financial & Corporate Compliance, this is regulated by law: in many US states, companies must designate a registered agent to receive official documents. This gives the service a statutory function, although this does not mean that every associated information service is untouchable. In Legal & Regulatory, the status rests on recognized authority: the value of legal commentary can partly depend on the reputation of authors, editors, and publishers. Such sources are cited in proceedings, even though they generally do not have binding status.
• Layer 4 is lock-in. Switching costs are the barrier here. Because Wolters Kluwer offers its content to its customers via its own software, and because that software is now deeply embedded in its customers’ business operations (data formats, work processes, training, contracts, and compliance), migrating to another solution quickly leads to high switching costs.

Investors seem to react primarily to developments within Layer 1. That layer is in constant flux, and this is measurable: every release of a new language model delivers better performance. Layer 2 moves in bursts: a publisher’s content is no longer made public every quarter, but can become accessible in a single press release, for example through a licensing deal between a publisher and an AI lab. Layer 3 moves barely. Layer 4 could erode if the interface between publisher and customer changes. Layer 1 is the most visible and easiest to link to a news event, thereby influencing the daily price. The effects on the other layers are more difficult to measure and are likely factored into the price over a longer period.

Perhaps the most salient example of vulnerability to Layer 1 AI disruption is the American company Chegg (hence the title above), literally a textbook example. Chegg charged twenty dollars a month for homework answers and, at its peak, had over eight million paying subscribers and nearly 800 million dollars in annual revenue. No exclusive source, no authoritative authors, no regulatory role, and little technical or procedural lock-in. ChatGPT was released on November 30, 2022. Partly as a result of this, Chegg acknowledged the significant decline in revenue more than five months later (May 2, 2023), and the stock lost 48 percent in a single day. From a peak of 113.51 dollars in February 2021 to less than 1 dollar now: nearly fourteen billion dollars in market value evaporated.

Back to Wolters Kluwer. If you were to project Layer 1 AI disruption risk onto the company as a whole without further nuance, the share price drop is understandable. Legal & Regulatory is indeed vulnerable at Layer 1: summarizing legislative changes and case law is precisely what a generic language model does well.

The Layer 1 pressure on Health is understandable. Recently, this also seemed to be easily measurable. On June 26 of this year, Nature Medicine published a comparison in which generic models beat the specialized clinical instruments from Wolters Kluwer (UpToDate Expert AI) and the American product OpenEvidence on medical benchmarks: over 97 percent for the best generic model, compared to 88 to 90 percent for the two specialized products. I initially read that outcome as confirmation of pressure at Layer 1. Until I got to the rejection rate. UpToDate rejected nineteen percent of the questions, compared to only one to six percent for the generic models. That is where the power of Layer 3 manifests itself: status, in this case built on trust. A specialized system is likely only allowed to answer based on validated proprietary content; if a question falls outside of that, it refuses to provide an answer. A generic model always has other sources available and simply delivers, with or without validated substantiation. Wolters Kluwer therefore responded with: clinical AI instruments are not designed to win benchmarks. There may also be other factors explaining a high rejection rate.

Interestingly, protection at Layer 2 (access) and integration with client systems at Layer 4 (lock-in) can nevertheless lead to different outcomes. Thomson Reuters, a peer of WKL in the legal domain, announced on May 12, 2026, that it would connect its own legal system, CoCounsel Legal, to Anthropic’s product Claude. This was done via the Model Context Protocol (MCP), an open standard for communication between AI agents and external data sources, tools, and systems. More ‘open’, therefore, precisely in the domain where Wolters Kluwer remained ‘closed’. While Wolters Kluwer did announce an extensive collaboration with Anthropic’s competitor OpenAI, it did so within a more closed architecture. Models were brought in, integrated into their own products, and delivered to clients via their own WKL platforms. At the same time, that same Wolters Kluwer did open its databases in Health to external AI agents. Thus, there was no question of an open publisher versus a closed publisher. There was one company that gave a different answer per division.

Why would you, as a publisher, open up your valuable curated content to an external AI? This has everything to do with the increasing proliferation of generic language models among customers.

In a closed architecture, customer processes are facilitated as much as possible by the publisher. The publisher purchases a language model itself, combines it with its own curated data, and delivers a ready-made product, or better yet: a service. The majority of the relevant interaction takes place within the environment that the publisher controls and manages.

However, today’s customer also has its own AI environment. And this is tailored to its own processes, procedures, and documents. It can go much further in this regard than a publisher ever could, especially because a publisher does not have that customer context.

The integration possibilities between the two environments are currently effectively dictated by the publisher. Access by a customer’s AI agents is usually not (yet) an option in this architecture. And the publisher has good reasons for this: it could undermine its revenue model or margin, damage its reputation, or blur the delineation of responsibilities and liabilities.

Influenced by language models deployed by the customer, pressure is building towards a more open architecture. In this architecture, the customer can request information from the publisher via AI agents from within its own AI environment. Billing between customer and publisher would then no longer be based predominantly on the number of actual users (‘seats’), but on other units, such as volume or tokens.

The publisher has quite a bit of work to do to be able to offer such an architecture responsibly. Ultimately, professionals and institutions want an accountable, citable source behind the answer: afterwards, it must be possible to determine which question was asked, on which version of which source the answer was based, and who vouches for it. Without guarantees from the publisher, that responsibility rests entirely with the client, and a reason to pay the publisher disappears.

For a supplier wishing to build such a link between client and publisher, exactly thát is the core of their offering. When providing data for professional use, therefore, it is equally about keeping the chain of accountability intact.

When it comes to architectural choices, Wolters Kluwer does not follow a single course. The clinical flagship UpToDate has been deliberately locked down — the model may only draw from its own content, and the terms of use explicitly prohibit customers from feeding that content into their own AI. Anyone wishing to access its content must go through a paid, controlled channel.

And at the same time: in October 2025, that same Health division opened a paid channel for access to its clinical content. And four months later, in February 2026, a connection for medication data that actually is open to external AI agents — metered, and for the time being, for a limited group of developers. Two different architectural choices within one division, four months apart.

Perhaps, and I am still not sure about this, it works as follows. Where a product has high vulnerability at Layer 1, you eventually lose out to generic language models there. If a generic model performs better anyway, keeping your product closed is primarily a reason to lose customers in the long run. With high barriers to entry at other levels, a publisher can afford that closed-mindedness (for now), and buys time.

The assessment of disruption risks, set against investments and the current revenue streams, explains the priorities of Wolters Kluwer’s divisions better than principled company-wide considerations.

My original question could now be answered more effectively. However, it was formulated differently than I initially posed it. “Does Wolters Kluwer still have a right to exist in the long term?” is too simplistic, because the answer differs per division and sometimes even per product within a single division. The real question is: how much resistance do the various layers of protection of the divisions and their products offer against AI disruption?

At layer 1, a publisher can be overtaken insofar as its value consists exclusively of information processing that a generic model can perform with comparable sources. As generic models become more reliable, up-to-date, and affordable for that task, the pressure on this layer will likely increase further. At layer 2, controlled access itself is the core of the service, and at layer 3, reputation and status determine that ‘resilience’. And even if those three layers were to erode over time, you still cannot simply rebuild your business due to the high costs at layer 4. Naturally, Wolters Kluwer retains control over these layers of protection for as long as possible. Anyone wishing to venture into the realm of the interfaces between WKL and its customers needs a particularly compelling narrative regarding the current and future revenue model of a specific division or product. And also regarding its controlled transition under market pressure, while maintaining the chain of responsibility. That is what I would convey to my client.

It took quite a bit of discussion with my AI (in this case, Claude from Anthropic and Perplexity from Perplexity AI) to reach this conclusion. With the advent of AI, gathering facts appears to have become virtually free. However, judging which steps of reasoning carry the most weight is not. That is where my work has shifted, at least in this case.

And OPmaat, which I built thirty years ago? It is still running. And on closer inspection, that demonstrates above all that resilience at the higher levels buys time.