The Agentic Revolution
Most organisations are making AI decisions based on capabilities they evaluated twelve to eighteen months ago. Based on the keynote by Ralph Ramos, Member of Technical Staff at Anthropic, this paper sets out why the model you evaluated and the model you are about to deploy are not the same product, and what that means for governance.
Introduction
At the 2026 Zartis AI Summit, Ralph Ramos, Member of Technical Staff at Anthropic, argued that most organisations are making AI decisions based on capabilities they evaluated 12 to 18 months ago. The model they assessed and the model they are about to deploy are not the same product, and the gap between them, in terms of what AI can autonomously do, is larger than most decision-makers appreciate.
This paper is based on that keynote. It examines what changed inside AI capability in the space of 1.5 years, why evaluation cycles built for incremental software releases break down against a categorical shift rather than an incremental one, and what is already happening inside engineering teams while most governance frameworks still assume a chatbot from legacy timelines.
The central argument: organisations must not build governance around what a specific model version can do today, as those limits are constantly shifting. Instead, governance should be built around durable principles:
- Action Boundaries: Defining clear categories of actions an agent is permitted to execute.
- Escalation Pathways: Establishing protocols for when an agent encounters out-of-scope tasks.
- Reviewable Audit Trails: Maintaining clear records to keep agentic behaviour transparent and accountable as underlying capabilities advance.
Govern for the AI you'll deploy, not the AI you piloted
Get the full paper on why the evaluation cycle most organisations run is built for a product that no longer exists.
Closing the eighteen-month gap between the AI you evaluated and the AI you are about to deploy takes more than awareness of the problem.
Follow these steps for a clear path forward:
Design governance around principles and mechanisms, not around one model version's specific behaviour
Define the categories of action an agent is permitted to take, and the escalation path for when it exceeds them
Treat AI tools already in use across your organisation as a current state to manage, not a future risk to prevent
Make the sanctioned AI path easier to use than the unsanctioned one, rather than trying to restrict access outright
Prepare the wider organisation conceptually, not just technically, before agentic tools reach non-technical workflows
Survey your engineering leadership on what agentic tools are already in daily use, across sanctioned and unsanctioned channels, within the next 30 days
Download the whitepaper
