TORONTO — Cohere is betting that AI agents are the future of office work and of its business, as it makes its North assistant-builder system widely available.

The Toronto-headquartered AI firm has already sold customized versions of the product to a handful of large clients like RBC, Bell, the Saudi Arabian telecom company STC Group, and U.S. hospital software developer Ensemble Health Partners. Now, it’s opening North up to smaller firms hoping to use AI to boost efficiency and employee productivity.

Workers can use North to find information within their company’s databases and documents, or to create agents that can handle tasks on their behalf, then share them with colleagues. “It goes beyond just Q&A, and gets into doing work for you,” said Cohere co-founder Nick Frosst. 

North is powered by Cohere’s large language models (LLMs), as well as ones that can search and retrieve information from documents, images or a company’s other software systems.

Like many AI developers, the firm says its models can reason to come up with the right answer. While competitors like OpenAI and Anthropic tout their products’ scores on tests of math skills and advanced knowledge, Frosst said the reasoning capabilities of Cohere’s models are focused on figuring which tool to use for a specific task. 

For example, a hiring manager responding to a colleague’s requests to add to their team could use North to summarize the initial inquiry, then draft a job posting using information from the firm’s HR system. Or an investment banking analyst could build an agent that would scrape the firm’s financial databases to generate earnings summaries for companies it tracks, then automate the reports for future quarters.

The Canadian firm also touts the security and privacy credentials of its products. Clients can run North from their own servers, or within their existing cloud setups. That lets the system connect directly to internal data sources, and means Cohere isn’t reading the prompts workers are entering into the system or the responses and tasks they’re getting out. 

Cohere said it will work with clients to customize North, and has significantly reduced the time to roll out the system—large companies now take weeks to get online instead of months, according to Frosst. The firm started by targeting the biggest businesses because they have the highest security requirements and the most potential uses for the technology, he said. The company will now begin working with smaller firms, he added.

Cohere will charge clients a fee per user of North tied to the scope of capabilities they’re employing, rather than based on their usage. Developers have typically priced LLMs by the token, the single words or fragments that the system processes in and out. That’s limited adoption of agents, according to Frosst, who claims some workers have avoided using the technology so as not to run up their companies’ AI bills.

Model makers, large software firms and AI startups have launched agents for coding, IT management, and customer service. North is a “superset of what else is available,” Frosst claimed, letting clients replace a portfolio of other AI tools with Cohere’s single system.

RBC tested technology from several developers before buying North as its main generative AI system, Foteini Agrafioti, senior vice-president of data and AI, said in an interview earlier this year. It chose Cohere because the firm focuses on businesses rather than consumers, and because its tools can be run on RBC’s own hardware.

The bank is rolling North out to its developers this summer, and other teams are keen to employ it, Agrafioti said. RBC staff have only “scraped the surface” of what they can do with generative AI, she said. “Now they want to get serious with it.”