Cohere's North 2 is a workplace platform for AI agents, and its main additions are limits on what each employee and team can spend, memory that lets an agent carry what it learned from one session to the next, and a shared library of agents and skills. A company can run it with Cohere's own models or with models from other makers. Cohere does not publish a price, and it sells to large organizations, especially banks, telecoms and government.
The AI bill is following the cloud bill
Cloud computing taught companies an expensive lesson in the 2010s. Engineers could start servers in seconds, nobody outside engineering saw the cost until the month-end invoice, and finance could only negotiate discounts on a bill it did not understand. The cure was to make spending visible by team and to stop it before it ran away, not afterwards.
Agents repeat the pattern with a different cause. A chat answer is a short exchange. An agent that reads a folder of contracts, checks each against a policy and drafts a summary does that work in many steps, and every step is billed. Large companies are already seeing sticker shock as ordinary staff start using agents, and many have started putting caps on usage.
Spending limits are not new in themselves. Microsoft, OpenAI and Anthropic already let administrators cap spending inside their own platforms. What North 2 adds is a cap and a meter that work across models from different makers, so a company can see one total.
Memory and shared skills are what a company ends up owning
Take a finance team that builds an agent to match supplier invoices against purchase orders. In the first month it learns the small things: one supplier always bills freight on a separate line, another sends credit notes with the wrong date. If the agent remembers those from session to session, and the team keeps them in a shared library, a new hire starts with an agent that already knows them. The administrator can also choose which model that team uses, so routine sorting does not need the most expensive one.
After a year, the model matters less than that pile of learned rules, saved skills and shared agents. It is the part that took the team's time to build, and it is the part that is hard to move.
That is why a platform that works with any model is a sensible pitch. If the knowledge lives in a layer above the model, the company can swap in a cheaper or better model without starting over. It also means the lock-in moves up a level: the company is now tied to whoever holds the layer. Cohere's offer is that this layer can run on the company's own servers, even with no internet connection, which matters most to the regulated customers it targets.
What follows for people at work
Within a couple of years, a team lead in a company that uses agents will own an AI budget the way an engineering lead owns a cloud budget. The useful conversation will be about return: what a given agent saved in hours, and what it cost to run. People who can show that, with a before and after for one named task, will get more allowance than people who cannot.
For anyone building agents at work, the practical question is where the instructions and memory live. If you can export them and point them at a different model, you own the work. If they only exist inside one vendor's chat window, the vendor owns it.