Dreamforce 2024, San Francisco – Salesforce, a titan in customer relationship management, has ignited the annual Dreamforce conference with its most significant artificial intelligence announcement to date: Koa, a groundbreaking reasoning model designed to revolutionize how businesses leverage AI for sales, marketing, and customer support. This marks Salesforce’s inaugural foray into developing its own frontier reasoning model, a strategic move that leverages Nvidia’s cutting-edge, open-weight Nemotron architecture. The collaboration between the two tech giants signifies a pivotal moment, addressing the evolving and increasingly specialized needs of enterprise AI.
The unveiling of Koa at Dreamforce, Salesforce’s flagship annual event that typically draws tens of thousands of attendees, including customers, partners, and industry analysts, underscores the company’s commitment to pushing the boundaries of AI within the business landscape. Dreamforce, often a platform for major product launches and strategic announcements, has this year placed AI innovation at its absolute core, with Koa emerging as a central theme of the conference.
Koa represents a departure from the prevailing trend where proprietary AI labs often incentivize enterprises to upload vast quantities of sensitive data—including proprietary files, code, and user feedback—directly into their models. This approach, while enabling advanced capabilities, can incur substantial costs for businesses and raise significant concerns about data privacy and control. Salesforce, by contrast, is positioning Koa as a solution that empowers its enterprise customers with a powerful, internally-focused AI reasoning engine.
A Strategic Partnership: Nvidia’s Nemotron as the Foundation
The genesis of Koa lies in the strategic partnership with Nvidia, a leader in AI hardware and software. Koa is built upon Nvidia’s Nemotron, an open-weight model renowned for its advanced architecture and potential for customization. This open-weight nature is crucial, as it allows for greater transparency and adaptability compared to closed, proprietary models.
Nvidia’s Nemotron, announced in August 2026, was specifically designed to address the growing demand for sophisticated AI models that can be fine-tuned for specific enterprise applications. Its open-weight status has been a key differentiator, enabling companies like Salesforce to build upon a robust foundation without the restrictions often associated with proprietary systems. This approach aligns with a growing movement towards more accessible and controllable AI development in the enterprise sector.
The post-training phase, where Koa was meticulously refined, focused on imbuing the model with deep expertise in sales, marketing, and customer-support workflows. This specialized training distinguishes Koa from general-purpose AI models, allowing it to perform complex reasoning tasks with a nuanced understanding of business contexts.
Addressing the Enterprise AI Divergence
The development of Koa highlights a significant divergence in the artificial intelligence landscape. While frontier AI labs are increasingly focused on broad, general-purpose models that can ingest and process a wide array of data, the enterprise world is exhibiting a growing need for AI solutions that are not only powerful but also deeply integrated with specific business functions, data security, and cost-efficiency.
"We’ve built many small task-specific language models, which are part of Agentforce’s portfolio," explained Jayesh Govindarajan, EVP of Salesforce AI, in an interview with TechCrunch. "But reasoning has always been something that we’ve relied on the frontier model providers for. Until now."
This statement underscores a long-standing challenge for enterprises: the reliance on external AI providers for sophisticated reasoning capabilities. Before Koa, when a Salesforce agent within the Agentforce platform encountered a complex, multi-step task requiring advanced reasoning, the request would be routed through Salesforce’s AI gateway to external models such as Anthropic’s Claude or OpenAI’s ChatGPT. This reliance introduced potential bottlenecks, increased latency, and raised questions about data governance and cost.
Koa’s Unique Value Proposition for Businesses
Salesforce is offering Koa as a distinct alternative within its Agentforce platform, a robust environment where customers currently build AI-powered agents to automate routine tasks like answering customer inquiries, scheduling appointments, and managing lead follow-ups. The integration of Koa is poised to significantly enhance the capabilities of these agents.
The key advantages of Koa for Salesforce customers are multifaceted:
- Enhanced Reasoning Capabilities: Koa is engineered to tackle complex, multi-step reasoning tasks that were previously beyond the scope of many task-specific models. This allows AI agents to handle more sophisticated customer interactions and business processes.
- Domain-Specific Expertise: Through specialized post-training, Koa possesses a deep understanding of sales, marketing, and customer support scenarios, leading to more accurate and contextually relevant responses and actions.
- Cost Efficiency: Salesforce anticipates that Koa will offer a more cost-effective solution in terms of token usage compared to routing complex queries to external frontier models. This is a critical consideration for businesses looking to scale their AI initiatives without incurring prohibitive operational costs.
- Data Sovereignty and Security: By keeping reasoning tasks within Salesforce’s ecosystem, Koa enhances data control and security for its customers. This addresses growing concerns about proprietary data being processed by third-party AI providers.
- Reduced Latency: Processing reasoning tasks internally is expected to lead to faster response times for AI agents, improving the overall customer and employee experience.
The Genesis of a Sovereign Enterprise Model
The journey to developing an enterprise-grade frontier model like Koa has been a long-standing aspiration for Salesforce. However, a significant hurdle has been the lack of a suitable pre-trained base model.
"One of the reasons we hadn’t done this before, trained our own enterprise-grade frontier model – we always wanted to – but the challenge has always been the lack of a pre-trained base model to start with," Govindarajan elaborated. "Until Nemotron came along, there was no sovereign American pre-trained model that was available, one, and two, that was state of the art, and, three, that had clear data provenance. We have no idea what Qwen trains on," he added, referencing Alibaba’s Qwen model as an example of a popular but less transparent alternative.
The availability of Nvidia’s Nemotron, with its emphasis on clear data provenance and its American origin, provided Salesforce with the essential foundation. Data provenance is a critical factor for enterprises, as it ensures an understanding of the data sources used for training, which is vital for regulatory compliance, ethical AI development, and mitigating risks associated with biased or unverified training data.
Crafting Expertise with Synthetic Data
A significant aspect of Koa’s development involved a novel approach to data training. To avoid using actual customer data from Salesforce’s clients, which could raise privacy and security concerns, the companies opted to create synthetic data. This involved meticulously crafting datasets that mimicked the patterns, complexities, and nuances of real-world customer interactions and sales scenarios.
"We actually simulated a customer service environment with a persona customer service professional, including irate customers that call into the customer service center, all the way to a sales professional who’s trying to close a deal," Govindarajan described. This innovative use of synthetic data allowed Salesforce and Nvidia to train Koa to excel in its target domains without compromising customer privacy or proprietary information. The synthetic data generation process likely involved sophisticated algorithms to create realistic dialogues, problem-solving scenarios, and sales pitches.
Nvidia’s Role in Efficiency and Tokenomics
Nvidia’s contribution extends beyond providing the base Nemotron model. Kari Ann Briski, Nvidia’s VP of Generative AI Software for Enterprise, highlighted the architectural advantages of Nemotron that contribute to Koa’s efficiency.
"With Nemotron, we have a unique architecture for inference to be token efficient," Briski stated. "It’s kind of the trifecta of things that you need to have: sovereign AI, time to first token, efficient reasoning, for the tokenomics of it all."
This focus on token efficiency is crucial for managing the operational costs of AI models. Tokenomics refers to the economic principles governing the usage and cost of AI models, particularly large language models, where computational resources are often measured and billed in terms of tokens processed. By optimizing for token efficiency, Koa aims to deliver high-performance reasoning at a more predictable and manageable cost for businesses. The emphasis on "time to first token" also points to improved user experience, as it signifies how quickly a model can begin generating a response, reducing perceived latency.
A Balanced Approach: Continued Partnerships with Anthropic and OpenAI
Despite the significant investment in developing its own frontier model, Salesforce is not entirely abandoning its existing relationships with other leading AI providers. The company recently announced a strategic partnership with Anthropic, dubbed "ClaudeForce." This initiative allows companies to utilize Claude as their AI interface while ensuring that their sensitive data remains securely within Salesforce’s system of records, protected by its robust infrastructure.
This dual approach suggests a pragmatic strategy. Salesforce recognizes the value of having its own specialized AI capabilities for core enterprise functions while acknowledging the continued utility of broader, more general-purpose models from partners for certain applications. The ClaudeForce partnership, in particular, addresses a key concern for many enterprises: how to leverage powerful external AI models without compromising data privacy and control. By enabling Claude to operate within the Salesforce ecosystem, the company offers a secure and integrated experience.
Broader Implications for the Enterprise AI Market
The introduction of Koa has several far-reaching implications for the enterprise AI market:
- Democratization of Frontier AI: By building on an open-weight model and offering a specialized, yet accessible, solution, Salesforce is contributing to the broader trend of making advanced AI capabilities more accessible to businesses of all sizes.
- Rise of Sovereign AI Solutions: The emphasis on data provenance and the development of models trained on readily understandable data sources signifies a growing demand for "sovereign AI" – AI that is developed, controlled, and operated within specific national or regional boundaries, adhering to local regulations and ethical standards.
- Increased Competition and Innovation: Salesforce’s move into developing its own reasoning models will likely spur further innovation and competition among enterprise software providers, pushing them to enhance their AI offerings and explore new avenues for AI integration.
- Shift Towards Specialized AI: Koa’s success could signal a broader shift away from a one-size-fits-all approach to AI, towards more specialized models tailored for specific industries, functions, and business needs. This could lead to more efficient and effective AI deployments.
- Focus on Cost and Efficiency: The emphasis on token efficiency and cost-effectiveness by both Salesforce and Nvidia underscores a critical business consideration. As AI adoption scales, the economic viability of these technologies will become increasingly paramount.
The development and launch of Koa at Dreamforce 2024 mark a significant milestone in Salesforce’s AI journey. It represents a strategic pivot towards greater control, specialization, and efficiency in enterprise AI, setting a new benchmark for how businesses can harness the power of artificial intelligence to drive growth and customer success. The collaboration with Nvidia on the Nemotron architecture has proven instrumental in achieving this ambitious goal, paving the way for a future where AI is not just powerful, but also deeply integrated, secure, and cost-effective for the enterprise.
About the Author
