Salesforce CEO Benioff with details on the Agentforce AI platform

Contents

A large part of office work consists of repetitive tasks, and there is a lot of optimization potential in this, according to CRM provider Salesforce. Large language models (LLMs) already perform simple writing tasks, such as summarizing a text. In future, LLMs will support all e-commerce areas more actively than before, for example to process product inquiries or warranty claims from customers. Salesforce recently presented the AI platform Agentforce. At the Dreamforce in-house conference in San Francisco, Salesforce CEO Marc Benioff gave an insight into what this will look like.

Benioff compared his new AI to self-driving cars and took a swipe at the competition: instead of just assisting as a co-pilot in the passenger seat, Agentforce uses customer data to make decisions independently within the limits set for it. The first wave of artificial intelligence made predictions, the second assisted with all kinds of tasks and now the time has come for the third generation.

To stay in the picture: The AI gets behind the wheel to place an order, for example, without an employee signing off on the action. However, this also poses a major risk, as generative AI is known to have a tendency to hallucinate, i.e. to spit out incorrect, irrational answers.

To prevent Agentforce from sending 10,000 T-shirts to one customer, it is set certain limits, such as ordering a maximum of five items. This minimizes the risk of liability in the event of damage. According to Salesforce, the AI Agentforce, which emerged from Einstein Copilot, is hardly supposed to hallucinate. Benioff spoke of "low hallucination" in the keynote speech. But it is clearly not fully trusted.

Salesforce offers a range of ready-made agents that are designed to perform standard tasks. Salesforce works together with Nvidia, Google and IBM to achieve this. "You don't have to build your own LLM," promises Benioff. Your own AI helpers are created using natural language input.

AI helps in sales and marketing

The individual AI agents are defined by a handful of characteristics: their role, for example to process order inquiries, the data and APIs they are allowed to access, their scope of action and its limits, as well as the department in which they work.

They are available for customer service, sales and marketing, among others. Some contact end customers directly, others contact employees within the company.

Agents in practice

In customer service, chargeable service agents are intended to replace dry call bots, where in the end you would rather press 3 to speak to an employee. For example, a service agent can help if a customer wants to exchange a purchased item, check the order number and status and change the order. To do this, it accesses the customer's previous purchase history.

The commerce agent assists end customers with their online purchases, answers questions about the product range and suggests suitable items, such as the right mounting bracket for a solar module. Unlike human customer service staff, the AI agents promise to be patiently available around the clock.

The "Campaign Optimizer" uses artificial intelligence to create personalized marketing campaigns. The "Sales Coach" plays through a role play with a salesperson. It takes on the role of the customer based on the CRM data. This is intended to help build a stable customer relationship. In turn, new, user-defined agents are created with platform agents.

The intelligence in the background

The so-called "Atlas Reasoning Engine", the brain of these agents, works behind the scenes. The artificial intelligence converts structured and unstructured company data into work steps via the agents.

In the "Agent Builder", users can create their own agents with little or no programming effort. It is based on the existing "Prompt Builder". In the Agent Builder, the Atlas engine plays out the input and output of a user chat in the form of code for control purposes. Unlike conventional chatbots, the AI agents do not require a dialog tree.

This allows workflows to be defined for general inquiries, price negotiations or the processing of warranty claims. To achieve this, you define a title and an avatar, for example the image of the Salesforce mascot Einstein, add libraries that the agent can access and describe its task. All the necessary building blocks are entered in natural language. Once they are defined, the AI finds out for itself what needs to be done, according to the manufacturer.

Prices and availability

Agentforce for customer service and sales is due to be released at the end of October 2024. Salesforce does not plan to release some components of the Atlas Reasoning Engine until February 2025. Prices start at 2 US dollars per conversation with an agent.

Transparency note: The author was invited to Dreamforce in San Francisco by Salesforce. Salesforce covered the travel costs. There were no specifications regarding the nature and scope of our reporting.

(akr)