Detailed Analysis
US labor productivity grew approximately 2.7% in 2025, according to reporting by Eric Björnsson in the Financial Times, roughly double the decade-long average. The author of this commentary attributes a meaningful portion of that growth to AI agents and broader AI adoption, positioning the statistic as early macroeconomic evidence that AI tools are beginning to move the needle on workforce output. The argument frames this productivity surge not as an inevitable byproduct of better AI models, but as the result of deliberate behavioral and workflow changes by a subset of workers who have restructured how they engage with AI systems.
The central distinction the author draws is between passive AI use — asking a chatbot isolated questions — and active AI collaboration, in which the system operates as a persistent, context-aware partner embedded in a worker's ongoing projects and decisions. The piece illustrates this through a comparison of two users of Claude, Anthropic's AI assistant. The first user spends several minutes re-establishing context each session, while the second user benefits from persistent memory delivered via a Model Context Protocol (MCP) server through a platform called Open Bridge. The contrast underscores a structural inefficiency in how most people currently use AI: the absence of memory forces repetitive context-setting, which diminishes the compound productivity gains that come from sustained collaboration.
The reference to MCP servers and platforms like Open Bridge points to a growing ecosystem of middleware and memory infrastructure designed to close the gap between stateless AI interactions and genuinely collaborative AI workflows. Anthropic's Claude supports the Model Context Protocol, an open standard that allows external tools and data sources to feed persistent, structured context into AI sessions. This architectural layer is increasingly seen as a prerequisite for enterprise-grade AI utility, transforming Claude and similar models from reactive question-answering tools into proactive collaborators with institutional knowledge about the user.
The broader implication is that the productivity dividend from AI is not being distributed uniformly. Organizations and individuals who treat AI as a contextual collaborator — investing in memory infrastructure, workflow redesign, and integration — are capturing disproportionate gains, while those relying on surface-level chatbot interactions are largely replicating old patterns with new tools. This dynamic mirrors historical technology adoption curves, where the full economic impact of a general-purpose technology emerges only after complementary organizational changes are made, a phenomenon economists have documented with electricity, computing, and the internet.
The commentary connects to a wider debate in AI development about what "adoption" actually means. As foundation models like Claude become more capable, the limiting factor for productivity is increasingly not model intelligence but the surrounding architecture — memory, integration, and workflow design. Anthropic and competitors are investing heavily in agentic frameworks and persistent context mechanisms precisely because raw model performance has become table stakes. The real competitive frontier, as this piece implicitly argues, is in enabling AI systems to accumulate and act on knowledge about the humans they work with, transforming episodic assistance into something closer to an ongoing professional relationship.
Read original article →