Cohere North 2: The Enterprise AI Control Room That Works With Any Model

Most enterprise AI platforms make a quiet bet that customers won’t care which model sits underneath. Cohere’s North 2 is betting the opposite — and it’s betting on a real gap.

What Happened

Cohere has released North 2, a platform designed to orchestrate autonomous agents and multi-step workflows across enterprise environments without requiring organizations to commit to a single AI model. The platform runs on Cohere’s own Command A+ models, on custom organization-trained models, or on third-party models — deployed on-premises, in the cloud, or fully air-gapped. Production deployments are already live at LG CNS and Bell Cyber.

The platform’s mechanics center on a shared knowledge library and reusable skills catalog that agents draw from across sessions, with integrations into Slack, SharePoint, and Jira. A North Admin interface handles token usage controls and per-user quotas. When agents are asked to take consequential actions — modifying records, sending messages — human-in-the-loop approval gates can be inserted at any step. Cohere also claims North 2 requires fewer tokens on Nvidia Blackwell and Hopper hardware than competing platforms, a claim that matters at enterprise inference volumes.

Why It Matters

The acquisition of Aleph Alpha in April 2026 clarifies what North 2 is actually targeting. Aleph Alpha had been serving European governments on data sovereignty grounds — the argument that sensitive state and enterprise data cannot flow through a US hyperscaler’s API. North 2 appears to be the combined platform built for that buyer profile: organizations that need autonomous agentic workflows but cannot accept standard cloud model inference terms.

That buyer segment is larger than it appears. Regulated industries — financial services, healthcare, government contracting — have been largely absent from enterprise agentic AI deployments not because they lack AI budgets, but because most agentic platforms are architecturally incompatible with their data residency requirements. A model-agnostic, air-gapped-capable orchestration layer removes that blocker.

The model-agnostic design also speaks to a different kind of enterprise risk management. Organizations running multi-year AI programs are increasingly wary of hard dependencies on a single model vendor’s API pricing and availability. A control plane that can swap models without restructuring workflows hedges against that exposure.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.