Every AI agent needs reasoning abilities. It’s what gives them the ability to analyse, adapt, and make meaningful decisions. But until Salesforce’s Agentforce came along, no AI agents, copilots, or assistants had a true, enterprise-class reasoning engine for transforming raw data and metadata into intelligent actions that mimic or even surpass human thinking.
“Reasoning engines combine models, data, business logic, events, and workflows into unified cognitive architectures,” said Phil Mui, Ph.D., SVP of Technology, Head of Products and Engineering, Salesforce AI Research and Agentforce. “Reasoning engines are compound systems. The good ones are fast inference time System 2 reasoners. They try to understand the nuances of user queries, contexts and provide accurate, faithful, and sometimes actionable responses.”
Reasoning engines combine models, data, business logic, events, and workflows into unified cognitive architectures.
Phil Mui, Ph.D., SVP of Technology, Head of Products and Engineering for Salesforce AI Research
This is a striking shift from the AI assistants we’ve come to know, with their scripted prompts and unpredictable replies. Inspired by noted psychologist Daniel Kahneman’s bestseller, Thinking, Fast and Slow, reasoning engines integrate the dual-process framework of human thought — the idea that we can sometimes be quick to react in a moment because a situation calls for it. Or we can take a moment, think about things more deeply, and come up with more intricate solutions to problems or needs.
In AI terms, many do-it-yourself (DIY) assistants, which are essentially thin software veneers over large language models (LLMs), go down that first rapid, or “System 1” path because they’re built for basic tasks like answering frequently asked questions or taking product orders. But organisations thinking about employing agents to offload or augment human work need higher levels of reasoning. This is where “System 2” inference comes into play with its ability to understand and think through a situation, deliver relevant insights, and suggest actions.
From quick fixes to thoughtful solutions
That’s part of what makes Agentforce — the first digital labor platform for enterprises — unique. By using System 2 inference-time reasoning, its Atlas Reasoning Engine, the “brain” behind Agentforce, can retrieve the most relevant data then reason and act, greatly improving processes across customer service, sales, and operations. When faced with a request, Atlas enables Agentforce to refine the query — expanding it with additional context — then performs advanced retrieval augmented generation (RAG) that pulls in relevant data and metadata while assessing the quality of its own response. This ability to answer a question, then reflect on the answer before seeking to answer again, enables more accurate responses and precise actions compared to other System 2 assistants that don’t have deep data and metadata context in Salesforce.
That step-by-step deliberate thinking also helps minimise one of the most pervasive and worrying pitfalls of AI hallucinations. Indeed, 78% of workers would stop using an AI agent if it produced incorrect responses, according to a new Salesforce study. While System 1’s fast, instinctive response times can handle simple queries, it often falters in complex or ambiguous situations, leaving room for LLMs to fill in the blanks of its knowledge by delivering results that look real but are inaccurate, creating significant risks for businesses.
With hallucination rates for commercial AI chatbots ranging from 3% to as high as 27%, the value of deliberate, context-rich System 2 analysis is clear. Grounding decisions in deeper reasoning reduces errors while providing more complete and nuanced responses to complex questions. AI agents using System 2 reasoning, fueled by solutions like Salesforce Data Cloud that unify and harmonise business and customer data, further minimise hallucination risks while delivering more contextual and actionable insights — including inline citations that show the sources that Agentforce consulted to craft its response.
“We all know LLMs can hallucinate when there are requests without restraint,” said Claire Cheng, Ph.D., VP of Machine Learning and Engineering for Salesforce AI. “But by accessing the business domain and customer data, agents can reason much better and understand the context of a request to deliver more personalised and relevant responses.”