Evan Armstrong: Why the Context Layer Is Reshaping the Economics of Enterprise Software and Hiring

Tech analyst Evan Armstrong suggests looking at enterprise AI not as a threat to software itself, but as a redistribution of value within the stack. In his view, code and interface-level SaaS products are getting cheaper, and a company's main asset becomes context: knowledge of processes, access rights and the order in which things get done.

Why SaaS slumped

In his article Context is King, published on 12 February 2026, Armstrong tries to explain why the market sharply reassessed classic cloud services. According to him, roughly $300 billion in software-company market capitalization vanished amid fears that generative AI will turn building applications into a commodity. The investors' logic looks blunt but is easy to follow: if a CRM, an analytics dashboard or an internal tool can be assembled from a set of instructions, why pay a SaaS premium the way they used to?

Armstrong thinks it is not only about the hype around code generation. He points to three deeper shifts: slowing growth at public SaaS companies, worsening gross margins in products built around AI from the start, and falling costs of switching between services. By his estimate, public SaaS growth rates have dropped from 36% to 17% since 2023, while new AI applications currently hold margins closer to 50–65% rather than the former 75–85%. In this model, code gets cheaper, and the old business valuations no longer look sustainable.

Three layers of the stack

Armstrong's key idea is that enterprise software is now better described not as a set of separate applications but as a stack of three layers. The first layer is systems of record and databases, which are not going anywhere, because agents need verifiable records to rely on. The second layer is interface applications, where a person clicks buttons, reads reports and performs individual tasks. In the author's view, this is the layer most at risk of becoming interchangeable.

The third layer, the context layer, is where attention lands. It is not merely agent orchestration but a company's accumulated knowledge of how work actually gets done day to day. What matters here is not just formal rules but practical dependencies: who approves exceptions, where a process stalls, which combinations of actions speed up a deal or help close an incident. According to Armstrong, this layer determines whether an agent proves useful in a working environment or remains a polished demo.

• who has access to which data
• which actions can be taken without approval
• the order in which sales, procurement, incidents and support proceed
• which steps more often lead to success and which end in error

Such a layer turns an agent from an executor into a system that picks the right action at the right moment. It also accumulates: every run of a workflow leaves traces that make the next run more accurate. So context becomes not a box-ticking document but an asset with a compounding effect. In essence, a company starts encoding its own way of working so that not only people but also AI systems can use it.

"The context of how people work eventually becomes the context for how

AI agents work even better."

Who will get the layer

The most interesting question here is strategic rather than technical: who will own this layer. Armstrong lists several contenders. ServiceNow already has deep penetration into enterprise processes, Notion controls large bodies of internal documents and knowledge bases, and Glean is building a bridge from enterprise search to enterprise context. At the same time, OpenAI and Anthropic are approaching from the agent-platform side, trying to become a semantic operating system on top of CRMs, databases and internal applications.

An uncomfortable conclusion for the labor market follows from this logic. If the context layer really starts to replace coordination costs, money will flow out not only of IT budgets but also of payroll. The pressure will fall not only on developers of simple internal tools but also on roles that rest on emails, approvals, handoffs between departments and manual process support. This is about the gradual automation of office coordination, not just another round of developer optimization.

What this means

The main takeaway for business is simple: the winners will not be those who merely have more models or a lower token price, but those who fastest turn their processes, access rights and internal logic into agent-ready context. In the new enterprise race, the most valuable asset is not the interface but the knowledge of how a company really works.