How Cohere's North 2 tackles the enterprise AI agent control problem

Enterprise teams deploying AI agents face two recurring problems: costs that are hard to predict and platforms that tie them to one vendor’s models. Cohere’s North 2 release, published on October 5, 2026, is pitched at both. Cohere says it adds granular cost controls and a model-agnostic design. This piece separates what Cohere states, what The Logic reports, and what we infer from them.

What Cohere says it shipped

Cohere describes North 2 as its “biggest upgrade yet” and says it has a redesigned agent orchestration system for multi-step automations, with humans kept in the loop where it matters. The post lists Skills (reusable capabilities agents can call), Libraries (shared organizational knowledge), Memory (context kept across sessions), Applications (prototyping decks, dashboards, documents and lightweight apps from natural language) and Automations (templates, a drag-and-drop workflow builder, and real-time performance monitoring).

On connectors, Cohere names Slack, SharePoint, OneDrive, Microsoft Outlook, Microsoft Exchange, Jira, Linear, Notion and GitHub as new, and lists PitchBook, Crunchbase, Daloopa, FiscalAI, S&P Global and FactSet as planned financial data providers.

Cost and governance controls

This is the headline claim. According to Cohere, North Admin lets administrators set roles and permissions, control which models are used where and by whom, and monitor activity in real time, including token spend and limits. Cohere says admins can define consumption tiers based on requests and token rates for users and groups, and that alert thresholds can notify teams before limits are set. The post’s summary also mentions rate limits, user quotas and org-wide caps.

The Logic quotes Paul Teyssier, Cohere’s vice-president of product for AI, saying cost control is “probably the number one rising concern for AI leaders.” The Logic also reports that North 2 lets businesses set budgets, limit spending per employee and team, and automatically downgrade to cheaper tools for simpler tasks. It adds an important caveat: Cohere has not yet built its own router into the platform, though Teyssier said the company is seeing demand for one. So automatic downgrading, as described, should not be read as a finished routing product.

Model choice

Cohere states that North 2 is model agnostic: customers can use Cohere’s models or bring their own. The Logic gives specifics, reporting that customers can plug in systems from Anthropic, OpenAI and Google, or cheaper open-source models from Chinese firms such as Alibaba, DeepSeek and Z.ai. It contrasts this with Claude Code and Codex, which are tied to their makers’ models.

Teyssier told The Logic: “A whole bunch of players [are] trying to take the customers deeper and deeper into their ecosystem, with no way out. We are intentionally taking the other approach.” That is a vendor’s positioning statement, not an independent finding.

Deployment and security

Cohere says North deploys self-hosted, in a VPC, hybrid, on-premises or fully air-gapped, and lists SOC 2 Type 2, ISO 27001 and ISO 42001 certifications. It also lists guardrails, autonomy policies that limit agents to authorized actions, and observability tools. Cohere says it has rolled North out in finance, healthcare, telecommunications, manufacturing, energy and the public sector. The Logic separately describes Cohere’s focus as tools for highly regulated industries such as financial services, telecom, defence and the public sector.

Named customers and partners

In the Cohere post, LG CNS says it uses North internally and brings it to customers across Korea, and Bell Cyber says it is bringing North into its security operations. NVIDIA’s Kari Briski is quoted saying Blackwell and Hopper accelerate inference for Cohere models, and Cohere says that pairing yields more tokens per second per node. These are vendor and partner statements; the post gives no figures. The Logic reports that North launched in January 2025 with RBC as its flagship customer and became widely available that August, and lists Bell, STC Group, the federal Innovation Department and the Aston Martin F1 team among other clients.

The trade-off Cohere acknowledges

The Logic reports that CEO Aidan Gomez acknowledged, in an interview last month, that Cohere’s technology is not as good at certain coding and cybersecurity capabilities as Anthropic’s or OpenAI’s, and that the company is working to close the gap with more compute and larger models. Teyssier said opening North 2 to third-party models should not be taken as an admission that Cohere’s models are lagging, and expects customers to keep using Cohere’s models for most routine applications while choosing alternatives for niche tasks or experiments.

What is not disclosed

The Cohere post gives no pricing, benchmarks or cost-savings figures, and directs readers to contact sales for a demo. We found no independent test results in the sources we could read. We could not read VentureBeat’s coverage, so this piece does not use its survey data.

Our read

This section is our inference, not a reported fact. If the model-agnostic design works as described, it would let a customer pair Cohere’s orchestration with a stronger outside model for coding-heavy work, which fits Gomez’s acknowledgment of gaps. Whether the cost controls reduce spend in practice cannot be judged from public material, and the missing router suggests the cheaper-model downgrade is still limited. Buyers weighing North 2 against other agent platforms should ask Cohere for pricing, a list of supported third-party models, and evidence of how usage limits behave under load.

The piece treats the model-agnostic claim and the missing router as separate threads, but they intersect in a way I would want pinned down before a pilot. If North runs air-gapped or inside a customer VPC, real-time token-spend monitoring and org-wide caps have to reconcile against a control plane the customer may not let send anything back out to Cohere. So when an org-wide cap is actually reached, does the agent stop mid-workflow, or is that only surfaced afterwards in reporting? The description of alert thresholds reads like reporting, and whether there is enforcement behind it is a different question.

The Logic’s note that Cohere has not built its own router into North yet is the detail that matters most for an architecture decision. If the cheaper-model downgrade has to come from a routing layer you build or buy, then two systems are picking models: North Admin’s per-user and per-group model controls, and your router. Which one wins when they disagree, and can a request that North Admin has pinned to a specific model still be routed elsewhere downstream? Without a documented precedence, the cost-control story and the model-agnostic story can pull against each other.

It is worth separating the three certifications Cohere lists, because they are not interchangeable. ISO 27001 covers information security management, while ISO 42001 is the AI management-system standard, so the scope, the auditor and the boundary of the certificate differ, and either can cover only part of an organization. Since the article flags these as Cohere’s own statements, a buyer in a regulated sector would want certificate numbers and the stated scope rather than the standard names, and specifically whether the deployment mode in play sits inside that scope. The post lists the certifications without those details.

The word to watch in the connector list is “planned”. FactSet, S&P Global, PitchBook and the others are entitlement-based products, so connecting an agent to them is not only an integration backlog, because the customer’s existing licence has to permit machine access at that volume. For a team evaluating North specifically for research or advisory workflows, the deciding question is whether those planned connectors ride on entitlements the firm already holds or require renegotiation. That can set the timeline more than the engineering does, and the announcement gives no dates for them.