Cohere: From Enterprise To Everywhere
By: Tej Sharma & Raymond Wang
The Ivey Business Review is a student publication conceived, designed and managed by Honors Business Administration students at the Ivey Business School.
Covering Cohere
Cohere is a Canadian technology firm focused on developing language artificial intelligence (AI) enterprise solutions by implementing advanced large language models (LLMs). In 2019, Cohere was founded to explore natural language processing (NLP), a branch of artificial intelligence that allows computers to understand and manipulate language like humans. Since then, Cohere has grown considerably to become a major competitor and rival to OpenAI and Microsoft in the enterprise segment of the generative AI market.
Before Cohere, companies utilized a sequential approach to analyzing data. Hence, many chatbots for companies had little usability in solving complex customer problems as they ignored the context from previous datasets. However, Cohere’s innovation was its creation of the transformer model, a deep-learning model that revolutionized the way machines comprehend language by allowing them to process vast amounts of data simultaneously, rather than sequentially. Drawing upon increasingly diverse and extensive training data, Cohere’s model was greatly improved at executing a wide range of tasks, such as text classification and generation.
In June 2023, Cohere announced a $270 million Series C and a partnership with Oracle to further explore use cases for its product. With a current valuation of $2.2 billion, Cohere is actively seeking opportunities to further advance its technology and raise additional capital to continue competing with industry giants. With the additional funding, they are actively seeking new B2B opportunities and partnerships to “get their models out into the real world.”
Multi-Language as the Next Frontier
With the recent intense competition from rivals like OpenAI, Cohere’s LLMs are currently losing on sheer performance. According to standardized tests like the Knowledge-oriented LLM Assessment benchmark (KoLA), Cohere’s models are known to hallucinate (present false or misleading information as fact) the most and score worse on general performance and reasoning. However, the major difference between Cohere and OpenAI is Cohere’s market dominance in the B2B space. Therefore, to keep up in the race of speed, innovation, and market share against OpenAI and other competitors, Cohere needs to differentiate through new service offerings and prove a unique use case for enterprises.
So, where could Cohere not only compete but maintain a consistent advantage? Cohere should focus its efforts on offering multi-language support, which would become more important in an increasingly globalized business world. Whereas OpenAI continues to cement itself as a market leader in generative AI for English and other major languages, hundreds of lesser-known languages, or low-resource languages (LRLs), remain overlooked. For example, Swahili has 150 million speakers and is one of the most widely spoken languages in Africa, yet LLMs including ChatGPT struggle to speak it fluently. In late 2022, Cohere was the first provider to launch a multilingual language understanding model with support for over 100 languages, including Swahili. Although it outperformed existing models by 230 percent, it did not support the text generation functionality necessary for intelligent assistants—a much larger and more difficult undertaking requiring the collection and validation of millions of sample multilingual text conversations. Cohere currently lacks this data for LRLs.
To increase its access to such data, Cohere initiated its Open Science Aya Project—a 3,000 volunteer-driven effort to create high-quality multilingual datasets for training new international models. As of early December 2023, it has gathered over 250,000 general examples across dozens of LRLs. However, the existing dataset falls significantly short of bridging the gap to create an effective model. For instance, Swahili has only 300 examples. This pales in comparison to the millions of examples used by previous machine translation models to achieve baseline accuracy for intelligent assistants.
To improve the functionality of Cohere’s core command model for LRLs, such as Swahili, without terabytes of data, two data requirements must be met: firstly, the data has to be high quality, and secondly, the data has to be related to enterprise activities. Most important is a substantial body of high-quality, enterprise-related Swahili language data, ranging from general text to exchanges between customers and service agents to financial-related inquiries. Furthermore, collected data needs to be verified and validated for accuracy before it can be used to train a model. Cohere’s strategy should thus focus on obtaining such troves of this type of data.
Learning the Language
Partnerships with major industry players to jointly develop multilingual models could further enhance the training of Cohere’s command models for LRLs. As COO Martin Kon said: “We are focused on enabling our customers to create proprietary LLM capabilities leveraging their data and creating strategic differentiation and business value.” Thus, Cohere’s strategy should extend beyond increasing client consumption and include rearchitecting organizations’ analytics functions by partnering with enterprises to co-develop multilingual models that set the standard for LRLs.
Cohere is uniquely well-positioned to consult for international clients who aim to improve operations by implementing advanced NLP capabilities. By supplementing existing multilingual models with global language support such as millions of transcribed phone calls and chat message logs, Cohere could further develop its expertise. Drawing from its own internal expertise in designing, implementing, and scaling data management processes such as annotation for anonymizing and securing client data, Cohere could help organizations create rich multilingual datasets that catalyze the creation and adoption of bilingual LLMs and beyond. In doing so, Cohere itself would gain access to valuable regional data and contextual expertise, while partners would have access to a highly tailored, bespoke model built for their operational needs.
M-PESA’s Mobile Money Platform
One example is M-PESA, whose clientele predominantly speaks Swahili. Operating as the de-facto “un-financial institution” of choice for the unbanked and two-thirds of the entire population of Kenya, M-PESA leads the world as one of the most profitable and successful mobile-based fintechs founded in a developing nation. Serving users in both English and Swahili, M-PESA empowers Kenyans to store and transfer money through basic text messaging with comparable functionality to Canadian chequing accounts of Interac E-Transfer—bridging the infrastructural gap by offering alternative access to key financial services. Owned and operated by Safaricom, Kenya’s largest telecommunications provider wielding over 60% of the nation’s mobile subscribers, M-PESA has been a significant driver of its parent company’s growth.
The Case for a Language-Specific AI
As an essential monetary medium that millions of Africans rely on daily, M-PESA struggles to adequately serve the rapidly growing population using only human customer support agents. M-PESA’s reliance on human agents increases the time-to-response and financial cost while limiting its hours of availability. M-PESA’s current answer to this is a rudimentary, English-only Facebook Messenger chatbot named Zuri, which operates using rudimentary if-else branching and requires internet access. However, primary banking services operate through SMS, meaning customers without data plans or sufficient technical knowledge to operate applications beyond basic text messaging often have to call and wait in long queues to receive assistance. Furthermore, considering that Zuri operates only in English when 97% of Kenyans do not speak it as their primary language, over 17 million speakers of Swahili—Kenya’s true lingua franca—are left behind when it comes to basic AI.
Implementation
A Cohere-M-PESA partnership would revolve around the joint development of a Swahili-English multilingual model, with M-PESA’s rich customer interaction data transformed into a comprehensive training dataset by Cohere’s team of machine learning engineers. This would in turn enable the creation of a model fine-tuned for bilingual financial interactions, ideal for automating customer support over SMS and a revamped Zuri chatbot. Through continuous performance monitoring and user feedback collection, Cohere and M-PESA could ensure that the new Swahili LLM operates safely while constantly improving its performance. This also further strengthens their lead in targeting high-demand business use cases, especially that of customer support, where 73 percent of leaders predict that conversational AI-powered customer service would be the new standard in less than 5 years.
Partnerships in Third-Party Model Usage
This partnership would open the door for organizations to transform their dormant client data into active, revenue-generating multilingual models. For example, once a partnered institution develops a language-specific model together with Cohere, they could allow third parties to access and use their model and earn revenue through a profit-sharing scheme. This would serve as a wholly new and unexplored income stream that enables non-traditional players to take their piece of the $45B generative AI market.
Partnership Structuring
Data security is a key concern for many organizations and has led to many large companies banning the use of LLMs. Cohere should thus structure the partnership by separating the consumer data and the enterprise data. The consumer data is used to improve the capabilities of the LLM and in turn, the enterprise is able to benefit from having a better model to improve the operations of the business through enterprise data. This ensures sensitive data does not leak while allowing Cohere to continue improving its multi-language capabilities.
From Enterprise to Everywhere
Equipped with a truly localized AI model to support its 57M existing customers and growing user base, M-PESA’s partnership with Cohere would allow it to activate next-generation natural language processing technology in significantly enhancing its customer service function. M-PESA could apply Cohere’s experience in deploying a high-performing conversational service agent for all users. In doing so, M-PESA would not only help users onboard faster and more efficiently to their platform but open the door to a new era of financial services for a community that lacks infrastructure; essential in achieving its goal of lifting a region with a 37% financial literacy rate out of poverty. Furthermore, M-PESA could capitalize on its 99% market share domination of B2B and B2C money transfers to sell existing clients on using its high-performing, jointly-developed Swahili model for their services—providing yet another diversified income stream and opportunity to enter the generative AI market.
For Cohere, a pioneering partnership with M-PESA could pave the path for its domination of LRL markets around the world. This way, Cohere could differentiate itself from performance powerhouses like OpenAI while gaining valuable expertise in becoming the LLM provider of choice in developing regions with high linguistic diversity. Enabling not only technological innovation, such collaborations also offer the opportunity for new unexplored revenue streams and business models. Through such key partnerships, Cohere could achieve its mission of empowering every developer and enterprise to build products and capture true business value—regardless of whether their first language is Java or Javanese.
