Inside the Company Building AI That Explains Itself

Cohere has been the quietest of the major AI companies for the past two years — overshadowed by OpenAI’s consumer dominance, Anthropic’s safety narrative, and Google’s infrastructure advantages. Its strategy, by design, has been different from all three: build AI that enterprises can actually audit.

The product is called Command R+, and the distinguishing feature is not performance on benchmarks but a retrieval and citation architecture that traces every claim in a model output back to a specific source document. When Command R+ tells you something, it tells you where it got the information. If the source is wrong, you can find it and fix it. If the model has hallucinated, the absence of a citation makes the hallucination visible rather than plausible.

“The killer feature of LLMs for enterprise is not capability,” says Cohere’s chief product officer, speaking to prompt/power from the company’s Toronto offices. “It’s that enterprises need to be able to trust what the model says. Trust requires verification. Verification requires transparency. We built for transparency from the start.”

The business case is proving out. Cohere’s enterprise revenue grew 340% in 2025, according to figures the company shared with prompt/power — the company remains private and does not publish full financials. Its customer list includes several of the largest financial institutions in North America and Europe, for whom the ability to audit AI outputs is not a nice-to-have but a regulatory requirement.

The question is whether the approach scales as capabilities advance. Transparent retrieval works well for factual question-answering on curated document sets. It becomes more complex for tasks that require synthesis, reasoning across sources, or the generation of novel content. Cohere’s bet is that enterprises will prefer the explainable model even at some capability cost. The market is still deciding whether they’re right.