Enterprise AI Transformation Programmes (2025–2026)

Narrative

Frontier lab data in this window functions as the primary quantitative record of enterprise AI usage, even though it is self-reported. Anthropic's Economic Index, now in its fifth and sixth iterations through mid-2026, tracks Claude usage across millions of sampled conversations and finds that coding remains the dominant enterprise task, with enterprise API usage overwhelmingly automated rather than augmented. The January 2026 report noted that enterprise API customers concentrated usage even further than consumer users: the top ten tasks represented 32% of first-party API traffic, up from 28% in the prior report. Anthropic also documents an elasticity between input and output tokens of 0.38, meaning complex enterprise tasks are increasingly bottlenecked by available context rather than raw model capability, a direct data point for the token-cost-economics research thread.

OpenAI's own "State of Enterprise AI" report frames enterprise adoption as accelerating in depth rather than just breadth: weekly Enterprise messages grew roughly 8x since November 2024, with ChatGPT Enterprise seats up approximately 9x year-over-year, while the report explicitly states that the primary constraints for organisations are no longer model performance or tooling, but rather organisational readiness. This claim, that the bottleneck has shifted from technology to organisation, directly parallels the independent commentary around the MIT NANDA "GenAI Divide" report, which found that about 5% of AI pilot programmes achieve rapid revenue acceleration while the vast majority stall, delivering little to no measurable impact on P&L. Independent critiques of the MIT figure note the underlying evidence is preliminary, based on a modest sample of executive interviews and public deployments, and caution that the finding reflects organisations treating GenAI as a parallel tool rather than an embedded operating change rather than proving outright technology failure.

On the governance and platform side, Google's April 2026 launch of the Gemini Enterprise Agent Platform (the rebranded successor to Vertex AI) packaged explicit governance primitives, an Agent Registry, Agent Gateway, and Agent Identity with cryptographic signatures, directly targeting the "agent sprawl" and audit-trail problems enterprises report when scaling from pilot to production. Anthropic has meanwhile built out a formal compliance stack (SOC 2 Type II, ISO 27001:2022, and ISO/IEC 42001:2023, achieved in January 2025 as one of the first frontier labs to do so) and offers HIPAA-ready configurations and zero-data-retention options, evidence that sector-specific regulatory bite (healthcare, finance) is shaping product design at the model-provider layer, not just at the enterprise deployment layer.

METR's evaluation work provides the independent capability backbone underlying claims about agentic transformation. Its time-horizon methodology, first published in March 2025, found that the length of tasks frontier models can complete autonomously with 50% reliability has been doubling approximately every seven months since 2019, with signs the trend accelerated to roughly every four months in 2024-2025. METR's February-March 2026 Frontier Risk Report, produced with data shared by participating AI companies, found that AI agents within AI companies often worked autonomously on real research and engineering projects with permissions and oversight comparable to human employees, while also documenting that agents showed significantly weaker performance on tasks requiring strategic judgement, stealth, and adversarial modelling than on pure technical capability, a nuance relevant to enterprises weighing autonomy-heavy agentic rollouts against residual human-oversight requirements.