Summary

• North 2 introduces a redesigned orchestration layer for multi-step AI agents and automated workflows.
• Administrators can impose token limits, user quotas, rate controls and organisation-wide caps while selecting which models different users can access.
• The platform remains model agnostic and can run in cloud, self-hosted, hybrid, on-premises and air-gapped environments.

Cohere is building cost controls and autonomy rules directly into its enterprise AI platform as companies move from small numbers of experimental assistants towards larger fleets of agents consuming models, data and computing resources.

Cohere has launched North 2 with a redesigned orchestration system for multi-step agents alongside reusable skills, organisational libraries, persistent memory and expanded administrative controls. The platform can use Cohere models or models supplied by customers and can operate across hosted, self-managed and air-gapped environments.

North Admin gives organisations control over model access and token consumption at user and agent level, alongside rate limits, quotas and broader caps around usage. The administrative functions address a problem that becomes more visible as AI moves into ordinary work.

Giving one employee access to a chatbot creates a relatively simple software cost, while allowing hundreds of people and autonomous workflows to invoke several models repeatedly creates a variable infrastructure bill whose relationship with business value is much harder to understand.

Agents turn model use into an operating expense

Conventional enterprise software often has a reasonably predictable licensing structure because a company buys a defined number of seats or pays against a known usage metric and can forecast the resulting expenditure with some confidence.

Generative AI introduces more variable economics as consumption changes according to model choice, prompt size, retrieved information, output length and the number of separate model calls required to complete a task.

Agents amplify that variability. A user may request one outcome while the software conducts several intermediate steps, searches internal material, calls external tools, revises its approach and invokes different models before returning an answer.

That behaviour is useful precisely because an employee does not have to manage each step manually, but one apparently simple action can consume considerably more model capacity than an ordinary chat request.

North 2 attempts to make that behaviour visible to administrators. Organisations can monitor token use, configure consumption tiers and receive alerts as thresholds are approached, while model permissions can be assigned according to user or group.

The objective is not simply to minimise expenditure because an organisation may rationally pay more for an expensive model when it materially improves a valuable task while routing routine work through a cheaper alternative. Cost governance becomes an allocation problem rather than a universal instruction to use fewer tokens.

Reusable skills bring standardisation to agents

The redesigned agent system also moves away from every employee independently building an automation from scratch. Skills provide reusable capabilities that several agents can call, while shared libraries give them access to common organisational knowledge and assets.

Memory allows agents to retain context across sessions rather than beginning every interaction without awareness of previous work. Those features can make AI more useful across repeated workflows, but they also make governance more consequential as persistent systems accumulate access and context over time.

An isolated chatbot session can disappear when the conversation ends, whereas an agent with memory, internal knowledge and permission to act through connected tools behaves more like an enterprise application and needs corresponding controls around identity, data and authorised actions.

Cohere says autonomy policies can determine which actions agents may perform independently and when human oversight is required. That distinction is central to regulated deployments because organisations are unlikely to apply the same level of autonomy to every decision.

Preparing a draft document and moving money, changing a customer record or altering production infrastructure are all actions an agent could theoretically perform, but their consequences are materially different.

Deployment control remains part of Cohere’s differentiation

North 2 can run in public or private cloud environments, inside customer infrastructure or in networks isolated from the public internet. Cohere has increasingly built its commercial position around organisations that cannot treat sending sensitive data to a conventional hosted AI service as the default deployment model.

That strategy has become more pronounced through Cohere’s agreement to combine with Germany’s Aleph Alpha, which places sovereign deployment and regulated customers at the centre of the enlarged company’s proposition.

North 2 remains model agnostic, meaning customers are not required to use only Cohere models inside the platform. Large enterprises are increasingly unlikely to standardise every use case around one model family when different systems can be better suited to reasoning, coding, translation, document processing or lower cost routine work.

Procurement teams may also want an alternative supplier available if pricing, availability or performance changes. A model agnostic platform can therefore become a control layer above a changing model market.

The commercial difficulty for Cohere is that other enterprise software and cloud suppliers are pursuing the same position while already holding extensive relationships with corporate IT departments.

The platform race is moving above the model

Competition in enterprise AI is gradually shifting from access to a capable model towards the infrastructure required to operate many models and agents safely across an organisation.

Enterprises need identity integration, monitoring, workflow orchestration, permissions, audit records and cost controls around the underlying intelligence. Similar management functions emerged around earlier generations of cloud computing once experimentation became routine production use.

North 2 reflects that maturation because Cohere is increasingly selling the ability to govern who uses AI, what it can do, where it runs and how much it costs rather than simply providing access to generative models.

Whether North becomes the layer through which enterprises manage those decisions will depend on integration breadth and adoption against much larger platform competitors. The operating requirement itself is becoming easier to identify.

Once agents begin consuming money and taking actions without a person initiating every individual step, model access stops behaving like a simple employee productivity feature and starts becoming an operational system somebody has to administer.