Cohere says its North 2 platform, which lets a company’s staff build AI assistants that take actions inside internal tools, can be installed on a closed network and can swap which AI model it runs on. For a regulated buyer, that combination is the notable part. Everything below comes from Cohere’s own launch post, which has no publication date, no benchmarks, no pricing, and no availability date.
The post names several places North 2 can live: the customer’s own servers, a private cloud, a hybrid arrangement, or fully on premises. It also says North can run air-gapped, meaning on a network with no connection to the outside world, so company knowledge never leaves the building. A disconnected install and connectors to hosted services such as Slack and Notion pull in opposite directions. The post does not say how the two coexist.
Cohere calls North “model agnostic,” meaning it can swap which AI model does the work. Customers can use Cohere’s own models or bring their own, and administrators can decide which models get used where, and by whom. The post names no third-party model tested inside North and offers no performance comparison. Whether a swapped-in model works as well as Cohere’s own is left open.
Control is the other half of the pitch. A console called North Admin sets roles and permissions, ties into a company’s existing identity systems, and shows usage down to the individual user and agent. Spending is the concrete lever: administrators define consumption tiers by request volume and token rate for users and groups, and can set alerts before a limit is hit. Autonomy policies are meant to restrict an agent to actions it is authorized to take, with a human sign-off for critical ones. The post gives no example of that working in practice.
On integrations, the post lists Microsoft’s Outlook, Exchange, and OneDrive alongside SharePoint, plus Jira, Linear, GitHub, and Notion. Financial data feeds such as PitchBook and FactSet are described as planned, not shipped. On compliance, Cohere cites ISO 42001, a standard for managing AI systems, along with ISO 27001 and SOC 2 Type 2. It does not say what those audits cover or link any reports.
Cohere also says that with Nvidia, its models produce more insights for fewer tokens on Blackwell and Hopper chips, which would cut what customers spend. No figure comes with the claim. Nvidia’s Kari Briski, vice president of generative AI, is quoted saying the chips “accelerate inference” for Cohere models; that is a hardware partner’s endorsement, not a measurement.
Two customers speak in the post. Yohan Jin, head of the AI Center at LG CNS, says his company uses North and is bringing it to customers across Korea. Jawed Ahmad, chief technology and AI officer at Bell Cyber, says North supplies “control, privacy, and sovereignty” for security operations. Both are testimonials Cohere chose to publish. Neither states a deployment size, a cost, or a measured result. Cohere adds that it has rolled North out in healthcare, finance, and government work, naming no customer beyond these two.
The post argues that earlier platforms forced a choice between security, intelligence, control, and cost, and says “North 2 is where that tradeoff ends.” Its title promises “Enterprise AI Without Compromises.” Those are marketing claims in Cohere’s voice. The post includes no test, customer result, or rival comparison that would let anyone check them.
A regulated buyer can test the headline combination directly. Ask Cohere to demonstrate a fully disconnected install running a non-Cohere model with the connectors it needs switched on, and to put a price on it. Until that demo exists, North 2 is a spec sheet.
Cohere, “North 2: Enterprise AI Without Compromises,” a post on Cohere’s company blog credited to the Cohere Team, which carries no publication date.
