Salesforce introduces Koa Decisions for faster enterprise AI choices
Salesforce has introduced Koa Decisions, a decision model that reads a request and a user-defined list of categories before selecting an answer in a single step. The company says the model can make these choices in milliseconds, reducing the need to use larger AI models for every task.
- Koa Decisions is the next version of Salesforce’s HyperClassifier model.
- It selects from user-defined categories in a single step.
- Salesforce says it supports images and audio.
- The company measured roughly 26 milliseconds for Agentforce routing requests.
- A Trailhead demo is available to people who are not Salesforce customers.
A model for targeted decisions
Koa Decisions is the next version of Salesforce’s HyperClassifier model, which the company uses inside Agentforce to route requests to the appropriate part of an agent. A question such as “Where’s my package?” can be directed to the part of the agent that handles shipping and delivery, while other components handle FAQs, promotional offers, or escalations.
Salesforce says Koa Decisions expands the use of decision models beyond routing. It can return an option from a list, make a yes-or-no decision, or assign a score. The model also supports images and audio, and Salesforce plans to extend it into flows, Apex, and prompt templates for Salesforce builders, as well as provide an API for developers.
The model joins the Koa family, which also includes a separate reasoning model built with NVIDIA. Salesforce says its data science team created Koa Decisions by applying novel training techniques to Google’s Gemma, an open-weight model.
Why Salesforce says decision models are faster
Salesforce describes large language models as systems that generate text, while decision models are designed to make focused choices that other software can use immediately. For example, a customer saying “I can’t log in to my account” could be assigned to accounts, engineering, or sales. A decision model returns the selected category and a confidence score, rather than producing a longer response.
According to Salesforce, the difference comes partly from how the models produce answers. An LLM generates text one token at a time, while a decision model reads the full request and produces its answer in one step. The company says this can require less computing for short answers.
In measurements of routing requests in Agentforce, Salesforce says its decision model took roughly 26 milliseconds, compared with about 1.4 seconds for a general-purpose LLM. The figures are the company’s measurements and relate specifically to routing requests.
Testing and business use cases
Salesforce says people can test Koa Decisions without being Salesforce customers by signing up for a free Trailhead account and running its demo. Users can add categories and change the sample request to see how the model updates its selection, including without retraining after a new category is added.
The company contrasts this approach with traditional software classification, which often required teams to define an outcome, collect and label data, train a predictive machine learning model, and check its accuracy. Salesforce says some enterprise customers previously took months to move from an idea to production, while a decision model can sometimes reach production in hours after a category is added.
Salesforce presents Koa Decisions as a way to reserve LLMs for writing and reasoning while the decision model handles quick operational choices. In one example, it can flag a sales deal as being at risk and select the team that should respond, while an LLM summarizes what changed. Salesforce says pricing and packaging are subject to change, availability may vary by region, and customer agreements govern availability.
Salesforce: Less Reasoning. More Doing. Meet Koa Decisions.
