The conversation at VB Transform 2026 zeroed in on a question that should be keeping every enterprise leader up at night: not whether to adopt AI agents, but whether you can trust the vendor you adopt them from. Rachad Alao, Cohere's VP of product engineering, made the case that sovereignty isn't a feature you bolt on. It's the entire stack, from the GPUs in your private cloud to the governance layer that decides which model touches your most sensitive data. For anyone who has spent the past two years watching enterprises bolt generative AI onto existing workflows, this is the hard truth: a chatbot that answers email is not a strategy. An agent that can reason, call tools, and act on your systems without you losing control of the data it touches, that is the actual prize.
Alao's sharpest point cut against the prevailing market logic. As token prices plummet, the assumption has been that bigger and cheaper models make local deployment less necessary. But Alao argued that total consumption is climbing faster than prices are falling, because agents don't just generate text, they reason, search, and take multiple steps. That means the cost curve isn't flattening; it's steepening. If you're paying per token, your bill scales with the complexity of the task, not the size of the model. Cohere's alternative is refreshingly direct: charge for outcomes, not consumption. That flips the incentive structure. Instead of maximizing your usage, they want to minimize it, routing each request to the smallest model that can handle it. For enterprises drowning in cloud bills, that is not a minor detail. It is the difference between a tool that saves you money and one that quietly becomes your largest line item.
The practical takeaway for anyone building agentic systems is to stop treating the model as the product. The product is the control plane. Alao's example of a Canadian bank routing regulated workloads to on-premises models while sending less sensitive tasks to frontier models through Cohere's North platform shows what maturity looks like. It is not about picking a winner. It is about building a routing layer that understands the sensitivity of your data and the intelligence required for each task. That is why the open-source releases matter. Cohere's North Mini Code runs on a single H100 and handles the bulk of developer tasks effectively, while Command A+ offers a compressed, Apache-licensed alternative for private deployment. For the 80% of use cases where a frontier model is overkill, these options give teams a real choice. And choice, not raw capability, is what breaks vendor lock-in.
Here is what we would tell a reader who asked us what to do with this: start auditing your agent stack today, not for performance, but for control. Ask who owns the governance layer, where your data physically resides, and whether you can swap a model mid-project without rewriting your entire application. The rise of AI projects in portfolios and the push for practical adoption guides suggests the market is maturing, but maturity without sovereignty is just dependence with better marketing. The specific question to watch is whether Cohere's model-routing approach forces the hyperscalers to change their pricing models, or whether they simply absorb the pressure by bundling agents into their existing enterprise agreements. If Alao is right that consumption will outpace price declines, then the vendor who charges per outcome, not per token, will be the one enterprises trust with their crown jewels. That is the bet. And it is a smart one.