What it is
Cohere North is a company-wide AI workspace (chat, search over internal knowledge, no-code agents and scheduled automations) that you run inside your own environment rather than in the vendor’s cloud. Cohere installs it in your VPC, on your own servers, in a hybrid setup or fully air-gapped. There is no multi-tenant SaaS version of North. The closest option is Model Vault, where Cohere hosts dedicated, isolated inference while the North application stays on your side. It runs Cohere’s Command models by default, and model, search and embedding components can be swapped for models you already license.
North went generally available in August 2025. North 2 shipped on October 5, 2026, adding agent memory that carries across sessions, shared agents, reusable skills and libraries, and admin controls for token spend: quotas per user and group, organization-wide caps, alert thresholds before a cap bites, and usage exports by model, user and department (VentureBeat). Cohere cites SOC 2 Type 2, ISO 27001 and ISO 42001.
The company behind it is changing shape. Cohere signed a merger agreement with Germany’s Aleph Alpha on September 16, 2026, after announcing the plan in April. The combined company will keep the Cohere name, with headquarters in Toronto and Berlin, Schwarz Group committing €500 million in financing, and a reported combined valuation of $20 billion. The deal still needs regulatory approval (SiliconANGLE). Cohere’s revenue was about $240 million last year, mostly from model API access rather than North.
Why it shows up in ops stacks
- The data cannot go to a US SaaS tenant. Banks, insurers, public-sector bodies and defense suppliers often fail ChatGPT Enterprise or Copilot at the data-residency review, not the feature review. North is the rare full workspace that passes because nothing leaves your network. RBC co-developed a banking configuration of North, Dell sells North inside its Dell AI Factory stack, and Bell Canada resells it as sovereign AI for Canadian government and enterprise buyers.
- Agents for back-office queues. The use case Cohere documents best is support triage. In its CoreWeave case study, North agents route and draft responses to tickets, an engineer approves every one, and mean time to resolution fell from 4–8 days to 2–5 days after a 90-day rollout. HR policy lookups, finance close checklists and contract-clause search follow the same pattern.
- Connectors plus MCP. Built-in connectors cover Gmail, Outlook, Slack, Salesforce and SharePoint, and North 2 added OneDrive, Exchange, Jira, Linear, Notion and GitHub. Anything else goes through a custom MCP server or the SDK.
Pricing reality
Cohere publishes no list price for North, offers no self-serve trial and quotes per deployment, based on infrastructure, usage, support and customization. Procurement guides report minimum commitments and annual true-ups in North contracts. Plan the budget in two lines:
- The license, negotiated. Treat 500 Microsoft 365 Copilot seats at $30 per user per month, about $180,000 a year, as the yardstick you negotiate against. North is built and sold as a company-wide program, and below a few hundred users the commitment rarely pencils out.
- The infrastructure, which you own if you self-host. Cohere’s docs size Command A+ at one B200 or two H100 GPUs. Two rented H100s at $2–3 per GPU-hour come to roughly $35,000–$53,000 a year before redundancy (estimate). Model Vault moves inference to Cohere: its published rates start at $2,000 a month for a small embedding instance, and a generative-model price is quoted, not listed.
Best for
CIOs and ops leaders at regulated companies of 1,000+ employees — banking, insurance, healthcare, public sector, defense supply chains — who must keep prompts, documents and agent actions inside their own network or jurisdiction, and who have a platform team to run it.
Do not buy it for a single team’s AI project, or if your security review already accepts a US-hosted tenant.
Versus the alternatives
Microsoft 365 Copilot and ChatGPT Enterprise hold the largest installed bases. Pick either when a vendor-hosted tenant passes your data review. Both roll out in weeks, and Copilot is already in your Microsoft 365 contract. Pick North when the answer to “can this data leave our network?” is no.
Claude Enterprise is the fastest-growing rival for enterprise agent work. Pick it when agent and model quality are the deciding factor and cloud hosting is acceptable.
Langdock and Mistral are the European sovereignty picks. Pick Langdock when an EU-domiciled vendor with per-seat pricing is enough. Pick Mistral when the model maker itself must be European. Pick North when you need air-gapped or customer-run infrastructure plus a negotiated enterprise program.
If even North is too heavy, self-hosted open-source Onyx covers chat and search on your own hardware at a fraction of the cost, without North’s governance layer or vendor support.
Watch-outs
- The merger is not closed. Roadmap, support org and contracting entity may change once Aleph Alpha folds in. Guard: add a change-of-control clause with a termination right, and get named roadmap commitments in the order form rather than on slides.
- Opaque pricing with true-ups. Minimum commitments plus annual true-ups can turn a pilot into an unplanned second-year bill. Guard: cap true-ups at a fixed percentage, tie the year-one commitment to a scoped pilot with exit criteria, and get the Model Vault rates in writing if you will not self-host.
- Spend-cap behavior is undocumented. Cohere does not say publicly what happens to an agent that hits its token cap mid-task. Guard: in the pilot, deliberately run a scheduled automation into its cap, and set alert thresholds well below hard caps for anything that writes to a system of record.
- You own the operations. Self-hosting means GPU provisioning, upgrades and uptime are your platform team’s job. Guard: confirm a named owner and GPU capacity before signing, or price Model Vault as the default and self-host only the application.
- The connector list is shorter than Copilot’s or Glean’s, and no helpdesk connector (Zendesk, ServiceNow) is named publicly. Guard: list the five systems your first agents must read and write, and get a working connector or MCP server for each in the pilot. If enterprise search across many permissioned sources is the real job, look at Glean.