North 2: Cohere Wants to Reconcile AI Agents, Security, and Cost Control
With North 2, Cohere promises regulated enterprises an agentic platform that runs where their data lives, without trading off intelligence, security, or budget.
In brief
Cohere is launching North 2, the new version of its enterprise AI agent platform, shaped by a year of deployments in finance, healthcare, telecom, and the public sector. On the menu: reusable agents, memory, automations, new connectors, air-gapped deployment, and fine-grained control over token consumption. The bet: sell sovereignty and governance rather than race for the most powerful model.
🍺 Bar-stool version
Cohere noticed that companies wanted AI agents, but not ones that would blab company secrets to the whole internet or drain the budget like an intern with the corporate card. So North 2 is agents that work in-house, on your own server, with a token meter and an alarm when things go sideways. Basically, AI with a built-in finance controller, which is either very reassuring or the most corporate thing ever invented. But it's exactly what banks and governments are waiting for before they'll finally scale this stuff up.
Key takeaways
- 1
North 2 is billed as the biggest update yet to Cohere's agentic platform, following a year of production deployments in heavily regulated sectors.
- 2
New features: reusable Skills, shared Libraries, persistent Memory across sessions, Applications generated via natural language, and Automations with a drag-and-drop visual editor.
- 3
New connectors to Slack, SharePoint, OneDrive, Outlook, Exchange, Jira, Linear, Notion, and GitHub, plus planned financial data providers (PitchBook, Crunchbase, S&P Global, FactSet…).
- 4
Deployment options include self-hosted, VPC, hybrid, on-premises, or fully air-gapped, with SOC 2 Type 2, ISO 27001, and ISO 42001 certifications.
- 5
North Admin enables guardrails and autonomy policies at the individual agent level, plus token spend tracking down to the user.
- 6
The platform is model agnostic: companies can use Cohere's models or bring their own.
- 7
Cohere highlights a partnership with NVIDIA (Blackwell, Hopper) for more tokens per second and per node, alongside clients like LG CNS and Bell Cyber.
The End of the Trade-Off, According to Cohere
The starting premise is simple: until now, companies deploying agentic AI had to pick a lane. The platform secure enough wasn't smart enough, the one smart enough wasn't controllable enough, and the one you could actually govern cost too much to scale.
North 2 positions itself as the answer to this impossible triangle. Cohere draws on a year of North deployments across finance, healthcare, telecom, manufacturing, energy, and the public sector to build a full-stack platform designed for organizations that can't afford to sacrifice anything.
The sales pitch boils down to one promise: make the ROI of AI more predictable, more transparent, and entirely owned by the enterprise.
Reusable Agents and Memory
At the core of North 2 is an agent orchestration system capable of handling complex multi-step processes, configurable via simple prompts, with humans in the loop where necessary.
The new building blocks aim to stop teams from rebuilding the same workflows over and over: Skills are reusable capabilities agents can call on, Libraries pool knowledge and resources, and Memory lets agents retain context from one session to the next.
On the production side, Applications generate presentations, dashboards, documents, and small apps from natural language, while Automations combine templates, a visual editor, and real-time monitoring. Connectors cover most of the core office and dev suite (Microsoft, Slack, Jira, Notion, GitHub), with planned financial sources like PitchBook or FactSet.
Security as the Selling Point
Cohere stresses the shift in risk calculus that comes with agents touching internal tools and exposed to prompt injection. North can be deployed privately, on-prem, in the cloud, or fully air-gapped.
North Admin lets teams set guardrails agent by agent, and autonomy policies restrict agents to approved actions, with human validation required for critical decisions. On top of that: red-teaming, third-party vulnerability scans, detailed logs, and SOC 2 Type 2, ISO 27001, and ISO 42001 certifications.
Bell Cyber illustrates the sovereignty angle: its CTO Jawed Ahmad says he wants to help Canadian companies and governments adopt agentic AI with control and confidentiality requirements built in from the start.
Governing the Token Bill
Third pillar: costs. Admins define roles and permissions, choose which models are used where and by whom, and track token spend in real time, from a system-wide view down to the individual user or agent.
'Flow control' lets teams set consumption tiers per user or group, with alert thresholds before limits are hit. North integrates with existing identity systems and remains model agnostic.
Finally, Cohere makes an efficiency argument: its models, trained and deployed on NVIDIA Blackwell and Hopper, reportedly produce more tokens per second and per node, and more insights for fewer tokens. Kari Briski, VP of Generative AI at NVIDIA, backs up the message.
“Every option demanded a sacrifice: the platform that's secure enough wasn't smart enough; the one that's smart enough wasn't controllable enough.”
“North was built to be owned, and it is increasingly built to be specialized.”
“NVIDIA Blackwell and Hopper accelerate inference for Cohere models, helping North 2.0 deliver more tokens for less. — Kari Briski, NVIDIA”
Why it matters
North 2 confirms Cohere's strategy: don't fight OpenAI or Anthropic on raw model power, fight on the turf where they're less comfortable — regulated enterprises that want to keep their data in-house. Air-gapped deployment, openness to other models, token governance: this is CIO and CISO talk, not viral demo talk. The battle for enterprise AI is shifting from benchmarks to plumbing (connectors, permissions, observability, chargeback), and Cohere has a real card to play there, especially on sovereignty versus the American giants. That said, this is a marketing piece: no efficiency gain figures, no customer ROI metrics, financial connectors still only 'planned,' and very generic testimonials. The announced features (memory, skills, automations) largely just catch up to what Microsoft Copilot, ChatGPT Enterprise, or Claude already offer. Differentiation will depend less on features than on Cohere's ability to prove, in production, that you can have intelligence without giving up control.
For you
Put it to work on your sources.
Free: a month of articles and three sources of your own. Pro: the whole archive and your sources, from €8/month.
For your team
The same machine, on your topics.
A space in your colours, your watch angles, your curators. Pilot open to three companies.
Read next
#safetyTodaySam Altman Defends His Middle Path Between AI Catastrophe and Monopoly
In the second part of his interview with Vanity Fair, the OpenAI chief justifies holding back a model, downplays the deal signed with Trump, and accepts that open source will cause serious incidents.
Source · Vanity Fair · Sam Altman Talks Donald Trump, Xi Jinping, and the Risk of Extinction
#safetyTodaySam Altman at Vanity Fair: the pause, the IPO, Musk, and the regret of not being rich
In a long video interview, the OpenAI boss talks about pausing training, the company's IPO, Elon Musk, suicide, and the film that portrays him.
Source · Vanity Fair · Sam Altman Talks Elon Musk, 'Artificial,' and the Future of Humanity
"No more AI": Elon Musk Wants to Rename AI "SI"
Three lines, two letters, and zero explanation: the xAI boss proposes dropping "artificial intelligence" from the vocabulary.
Source · X (Twitter) · No more AI