Cohere Put a Hard Ceiling on What an Enterprise Agent Is Allowed to Spend

Cohere unveiled North 2 on October 5, an overhaul of the platform it sells for running AI agents inside large organisations. The headline feature is a cap on spending.

Administrators can now track token use down to an individual user or agent, define consumption tiers for groups, and set a ceiling on what any of them is allowed to spend. Alerts fire before a limit is reached. The controls live in North Admin, the platform's management console.

That sounds like an accounting feature. It is better understood as a statement about why enterprise agent deployments stall. Cohere's framing is that companies have been forced to give up one of three things, security, capability or cost control, to get the other two. The new release is an attempt to stop treating cost as the sacrificial one.

The orchestration layer is the other half

The spending caps matter because of what sits on top of them.

North 2 replaces the orchestration system that runs multi-step agent work. The new version works through a task on its own and pulls a human in for the steps that need sign-off, rather than asking for approval at every turn. Agents built from a single prompt can be shared across a company, so a workflow one team assembled does not have to be rebuilt by another. Memory persists between sessions. There is a skills feature for packaging logic an agent calls repeatedly, and shared libraries for documents every agent should be able to reach.

From a natural-language request, the platform can produce a slide deck, a dashboard, or a lightweight application that a user prototypes inside the chat window. For recurring processes there are prebuilt templates and a drag-and-drop workflow builder with reported performance.

The connector library covers Slack, GitHub, SharePoint and OneDrive, with planned additions for financial data providers including PitchBook, S&P Global and FactSet. North has been generally available since August 2025, and Cohere says a year of production deployments in finance and the public sector shaped what went into the upgrade.

Why spend controls are the interesting part

An agent that runs unattended is an agent that can run up a bill unattended. This is the operational reality that has kept a lot of enterprise deployments in pilot.

A multi-step agent loop makes many model calls. Some of those calls are cheap classifications. Some are expensive reasoning passes. Without per-agent accounting, the cost of a single workflow is invisible until the invoice arrives, and by then nobody can attribute it to the decision that caused it.

Cohere's controls address that directly. Per-user and per-agent tracking gives an operator the data to answer which workflows are worth their cost. Consumption tiers let a company hand an agent a budget rather than a permission. Caps convert an open-ended commitment into a bounded one.

There is a second effect that is less about money. A cap on token use is also a cap on how far an agent can go wrong before somebody intervenes. The same mechanism that stops a runaway bill stops a runaway loop, which makes the spending limit a safety control as much as a financial one.

Model-agnostic as a negotiation position

North 2 lets an administrator assign models to teams, including models the customer brings from outside Cohere. That matters in procurement, where the alternative is a decision that locks a department into one vendor's pricing.

Cohere is leaning on its work with Nvidia to make the economics work, claiming its models produce more tokens per second per node on Blackwell and Hopper hardware. Kari Briski, Nvidia's vice president of generative AI, framed the result as more tokens for less.

Security controls sit in the same console. Guardrails can be applied to individual agents. The screening covers personally identifiable information in prompts and responses, and attempted prompt injection. Autonomy policies confine each agent to actions it is authorised to take. North runs on-premises or in a customer's virtual private cloud, including air-gapped installations for organisations that permit no external connection.

The customers show what the product is for

Bell Canada's Bell Cyber unit runs North inside its own security operations. Jawed Ahmad, the unit's chief technology and AI officer, said the platform lets it run agentic AI with the control and sovereignty requirements its customers demand already built in. LG CNS in South Korea runs North internally and resells it. Radical Ventures, an early Cohere investor, built an automation that prepares the briefing its team reads before a founder call, replacing hours of manual context gathering.

Those three examples share a shape. Each is a workflow that already existed, staffed by people, with a known cost and a known risk. The agent is being slotted into an established process rather than inventing one, which is where the spending caps become usable. You cannot budget an agent doing something new. You can budget an agent doing what a team already does.

Why the release landed three weeks after a merger

North 2 arrives less than three weeks after Cohere signed a merger agreement with Germany's Aleph Alpha, a deal that keeps the Cohere name on the combined company. Aleph Alpha's Kolibri, a 78-billion-parameter open-weight model tuned for German and English, points at the same sovereignty-minded buyer that North is courting.

The two moves read as a single argument. Aleph Alpha supplies models that can be trained and run in a specific jurisdiction and tuned for a specific language. North 2 supplies the platform those models run on, with the cost accounting and permissioning that a regulated buyer needs before it will deploy anything. Neither half is sufficient on its own, and the merger is what makes the combination sellable.

That explains the emphasis on on-premises and air-gapped deployment. The customers who care most about where their data sits are the ones who cannot use a purely cloud-hosted agent platform, and they are also the ones with the budget to pay for a separate deployment. The customer list, a Canadian security unit and a Korean IT services firm, fits the pattern.

What the controls do not solve

A spending cap is a blunt instrument, and there are things it cannot see.

The hardest problem in agent cost accounting is attribution. If three agents share a document library, a skills package and a model endpoint, and one of them runs a workflow that costs a thousand dollars, the cap fires on the agent that made the calls. Whether that agent caused the cost or merely inherited it is a separate question that no per-agent meter answers.

There is also the question of what a user does when the cap is reached. An agent that stops mid-workflow leaves work in an unknown state, and a system that resumes after a limit is lifted has to reconcile what happened in between. Cohere has not described how North 2 handles that, and it is the operational detail that determines whether the control is usable day to day.

The more testable claim is that an agent with a budget is more likely to be deployed than an agent without one. That is what North 2 is betting on, and the next twelve months of enterprise deployments will say whether the bet was right.

What the release sits next to

North 2 arrives less than three weeks after Cohere signed a merger agreement with Germany's Aleph Alpha, a deal that keeps the Cohere name on the combined company. Aleph Alpha's Kolibri, a 78-billion-parameter open-weight model tuned for German and English, points at the same sovereignty-minded buyer that North is courting.

That is the market North 2 is built for. Not the company that wants the most capable agent, but the one that has to justify the agent to an auditor. Cost control, per-agent permissions and on-premises deployment are the features that let that conversation happen.

The unresolved question is whether the controls hold up in use. A token cap is only as good as the accounting behind it, and attributing cost across shared libraries, shared agents and shared context is not a solved problem. Cohere's answer is that administrators can see use at the user and agent level, which is the necessary starting point rather than the finished article.

The more testable claim is that an agent with a budget is more likely to be deployed than an agent without one. That is what North 2 is betting on, and the next twelve months of enterprise deployments will say whether the bet was right.