Last week, Salesforce stated that AI models are becoming less relevant, and that it’s increasingly about the data, the context, and the right actions (workflows) to achieve a final result. They’ve been working toward that for over a year. Yet we’re seeing developments at Salesforce that counter that philosophy. A partnership with Anthropic to make Claude the default model, and now the introduction of its own AI model. Isn’t that contradictory?
Together with Nvidia, Salesforce has introduced its own AI model, a so-called CRM reasoning model named Koa. This model is based on Nvidia’s open Nemotron model, to which Salesforce has added a large amount of synthetic data and CRM processes based on its 27 years of CRM experience.
Claude and Claudeforce
Last month, Salesforce announced a major partnership with Anthropic’s Claude. This led to the launch of Claudeforce, which allows you to connect directly to Salesforce from within Claude and access data. However, Claude is also used by the Atlas Reasoning Engine, Agentforce Vibes, Slackbot, Slack AI, and more. This makes Anthropic a very important partner for Salesforce. Claude is more powerful and more effective than the models previously used by Salesforce.
Trusted Enterprise AI Harness makes the AI model less important
Then, last week, Salesforce launched the Salesforce Trusted Enterprise AI Harness. As we see it, this deterministic layer (Harness) ensures actions, context, security, and the AI models come together correctly to execute actions on the Salesforce platform.
One of the most important statements made during the introduction of the AI Harness came from Rohan Kumar, President, Chief Platform and Engineering Officer: “The Agentic Enterprise won’t be defined by which model a company chooses. Models will continue to change, and intelligence will increasingly be available everywhere. What will differentiate an enterprise is the trusted, proprietary context it brings to that intelligence…”
So, on the one hand, the AI model is less important; on the other, Salesforce is entering a major partnership with Anthropic to use Claude and introducing its own AI model, built in collaboration with Nvidia.
Koa is Salesforce’s proprietary CRM AI model
If AI models are less important, why have a proprietary AI model at all? The press release states that many AI models fall short when it comes to performing CRM actions. According to Salesforce, “Koa already matches or exceeds leading model performance on CRM actions with 3x fewer errors.” However, if we dive into the numbers and examine their research paper, published one day before the press release, that isn’t entirely accurate. Koa “surpasses a strong proprietary baseline (GPT-4.1) while remaining below the strongest frontier models.” On the CRM Bench, Koa (0.86) scores better than GPT-4.1 (0.81), is roughly on par with Opus 4.8 (0.87), but trails GPT-5.5 (0.90). Notably, the tests did not include Opus 5.
On Tau2Bench, which measures end-to-end customer service conversations, Koa still lags significantly behind the competition. Koa scores 69.41 points here, compared to GPT-5.5’s 83.99 and Claude Opus 4.8’s 74.00 points. So, to a certain extent, the model does indeed matter for the final result, so why choose Koa? The answer is simple: it’s cheaper.
AI Harness is about model selection, and Koa is about cost savings
Organizations that invest heavily in AI already know this: AI can make your organization much more efficient and effective. However, AI can also cost a lot of money quickly when used at scale. Salesforce has millions of users on its platforms, resulting in an enormous amount of token usage. That comes with a massive price tag, and that’s where Koa and tokenomics come into play. Tokenomics is simply about choosing the right model to use the fewest tokens and incur the lowest costs.
If you need to analyze a highly complex and disjointed customer service conversation using AI, Claude is currently the best choice for Salesforce. That said, many simple conversations and questions still require AI, but Claude is pure overkill. If Salesforce can handle those using Koa instead of Claude, it will save a tremendous amount of money.
Koa was developed in collaboration with Nvidia and is based on Nemotron-3-Super-120B-A12B, a model with 120 billion parameters, 12 billion of which are active per token. It’s also an open-weight model with an open license. Running this model is much cheaper than running a frontier model, and Salesforce doesn’t have to pay Nvidia for the license or for using Koa. Nvidia is simply giving this model away for free. Nvidia is doing this because Salesforce must invest in AI hardware not only to train Koa but also to run it in production. Salesforce can do this by setting up its own data centers with Nvidia chips or by renting AI capacity from one of the many AI data centers. Nvidia will simply sell more chips as a result.
Tokenomics
Salesforce presents Koa as an innovative CRM model. In our view, it’s simply tokenomics, and there’s nothing wrong with that. AI incurs enormous costs, and if the AI Harness works well and selects the right model for the right action, that’s not only cheaper but also more sustainable.
If we compare it to normal business processes, it also makes sense: you wouldn’t have your smartest employee or most senior engineer, who, if all goes well, is significantly more expensive than a junior or less experienced employee, answer the simple questions within your organization either.
If the AI model makes the difference, how important are data, context, and actions?
Looking at the acquisition of Fin, it seems data, context, and actions aren’t all that relevant after all. Fin is capable of autonomously resolving an average of 76 percent of customer service tickets. Salesforce already had Casey for this purpose, an AI agent that can also autonomously resolve customer service tickets. Reportedly, Casey can’t match Fin’s 76 percent rate for autonomously resolving tickets.
Fin doesn’t rely on data, context, or a frontier model, but rather on its own specially developed Fin Apex models to resolve its tickets. Casey, on the other hand, relies more on data and context from the Salesforce platform. Salesforce must have been hugely impressed, given the $3.6 billion acquisition. This might lead one to conclude that data and context aren’t that important as long as you use an advanced model. But when you look at the full picture, we can draw a parallel here with Koa: a proprietary model is significantly cheaper than a frontier model.
This brings us back to tokenomics: the more optimized the dataset, context, and actions, the fewer tokens are needed. Salesforce can now combine this technology with an optimized model in the form of Fin Apex, giving it the best of both worlds. This combination will also increase the speed, which is critical for live conversations.
The future of Fin and Casey is therefore straightforward: Fin excels at getting the most out of its own AI models, while Casey excels at optimizing the data and context. By combining these two worlds, Salesforce will soon have exceptionally effective customer service agents at its disposal.
The future of Salesforce’s AI agents
Returning to Rohan Kumar’s statement about how important data, context, and the right actions are, and that they are more important than the AI model itself, that is the entire vision and philosophy upon which Salesforce has built its strategy.
There’s nothing wrong with that vision and strategy; we’re seeing this across the industry, with companies working hard to optimize AI input. Tokenomics is also a widely discussed topic when it comes to selecting the right AI model for the right action. Many major vendors are doing exactly the same thing as Salesforce in this area.
It is telling, however, that Salesforce is shelling out billions for a company that owes its lead to its own domain-specific model rather than to data and context optimization. At first glance, this seems to run counter to the Salesforce vision. At the same time, it’s exactly the same conclusion Salesforce reached with Koa. For a well-defined domain, it pays to train your own model rather than rely on frontier models. In the coming year, the company will have to prove that it can truly combine these two worlds.
Today, with AIforce, Salesforce does finally connect the dots, but only on one floor. Claudeforce, Slackforce and Agentforce Coworker are presented as a single interface layer on top of Data 360, Customer 360 and Agentforce. The model layer is absent from that picture. Koa doesn’t appear in it, and neither does the reason it exists.
Where Salesforce has missed the mark a bit is in its communication over the past few weeks. Salesforce released all announcements in isolation, without considering the bigger picture until today. As a result, the message about AI models seemed completely contradictory to the “Harness” narrative and the previously communicated strategy regarding context. However, there’s a much deeper story behind this that still hasn’t been discussed, as it’s a bit more technical. Hopefully, we’ve now provided the necessary clarification here.