No Clippy, but Sophie: AI according to Salesforce passes the (Microsoft) Copilot

At Dreamforce, Marc Benioff shares his vision of the future of AI for businesses. Copilots are finished, autonomous agents are the future. And let it just so happen that, according to him, no one is better positioned to deliver it, than Salesforce.

At Dreamforce, Salesforce CEO Marc Benioff presents a vision. According to him, the future of AI will be agents: autonomous digital AI bots that can correctly perform a multitude of tasks.

“We already helped you roll out your sales force,” Benioff says, “and then your marketing force, your analytics force, your data force and now your agent force.”

Agentforce is new, but then again, it’s not. Retroactively, Saleforce is bombarding previously launched agents such as the Einstein Service Agent as part of Agentforce, launched as a concept on the eve of Dreamforce. Conveniently, because it allows Benioff to announce Agentforce version 2.0 in one fell swoop, powered by the new Atlas Reasoning Engine.

Reasoning like the brain

Agentforce, as the name suggests, is a collection of AI-driven agents. Those agents are small applets that are integrated into your Salesforce environment and more specifically the data in the Data Cloud, but can also talk to integrated third-party environments such as AWS, Google and IBM.

At the heart of Agentforce 2.0 and its agents is the Atlas Reasoning Engine. This was developed by Salesforce itself and is supposed to mimic how people evaluate and answer a query. First a query is evaluated and possibly refined with additional questions, then Atlas retrieves the relevant data, after which the engine creates an execution plan. To do this, Atlas uses Retrieval Augmented Generation (RAG). “RAG was invented here at Salesforce, by the way,” Benioff flaunts.

No Copilot

Agents are not copilots, Benioff is clear about that. “Copilot is the next Clippy,” he believes. “Customers don’t see the added value. They want to skip the copilot step.” According to Benioff and Salesforce, copilots have missed their opportunity. They don’t bring the promised productivity improvements, it sounds. Agents need to fix that. So what about the Einstein Copilot? That one gets an upgrade and becomes Agentforce.

Copilot is the next Clippy.

Marc Benioff, CEO Salesforce

Agents stand out from AI applications today in several ways. Especially their autonomous functionality stands out. Like an intern, for example, they can complete certain tasks without human intervention. Think of the triage and follow-up of leads, or customer support.

Sophie vs. Clippy

On stage, we see a demo illustrating the functionality. Someone calls customer service at clothing store Saks, where an AI agent named Sophie answers. Sophie can talk (in English) and has access to various Saks systems, including customer data, product data and store inventory data.

When someone calls asking about an item of clothing that is too small, Sophie immediately knows which item it is about. The agent looks at previous orders from the customer, who is immediately identified through the phone number. From there, Sophie suggests shipping the item in a new size, or suggests that the customer pick it up at a location nearby.

read also

No Clippy, but Sophie: AI according to Salesforce passes the (Microsoft) Copilot

The entire conversation proceeds in natural language and Sophie immediately understands what is needed. In contrast to a classic chatbot, the conversation does not follow an imposed structure. The agent would solve 40 percent to even 90 percent more cases than a classic chatbot. Unlike a copilot, the AI does not assist a human support agent: Sophie does everything herself. The Atlas engine makes that possible.

How?

Agents live in the Salesforce platform, on top of Customer 360 applications. An agent consists of five components:

  • A role: The role describes what the agent does. Roles are similar to already existing Customer 360 profiles. Agents are therefore managed within the Omni Supervisor.
  • Data: What data can be accessed? Agents can work with structured and unstructured data, metadata and external data, all through the Salesforce Data Cloud. This is an essential part of agent functionality.
  • Actions: What capabilities do the agents have? What are the workflows? Agents can handle prompts, as well as flows (which you may have already created) and APIs through Mulesoft. You can set up an agent’s functionality through the low-code Agent Builder, without coding. You add most functions in natural language, according to the demo.
  • Channels: Where are the agents working and where are the interactions? The capabilities there follow those of Commerce Cloud, for example. Agents can be accessed in Slack, via Whatsapp, SMS, chat….

Security is guaranteed by the Einstein Trust Layer. That system keeps an eye on the agents so that they function within clearly defined boundaries and do not hallucinate. For example, if a Sophie would not find a clear solution to the problem of the too-small garment, but would come up with an answer as LLMs sometimes do, the Trust Layer stops it.

Major breakthrough

Benioff himself is already very impressed with the interplay of capabilities within the Salesforce platform. “Agentforce is the biggest breakthrough we’ve ever had in technology, and I think it’s the biggest breakthrough I’ve ever seen in AI,” he lets slip with some sense of superlatives.

Salesforce has been working on agents and the Atlas engine for some time, but Benioff chose only a few weeks ago to make Agentforce the theme of the conference. “The keynote was all over the place as usual,” he told the assembled press. “And it worked well. But in my conversations with customers, I discovered that they have both high expectations and many frustrations with AI.” That insight must have come sometime before the release of quarterly results at the end of August, because even then he made remarkably negative comments about Copilot.

CEO Marc Benioff hung the entire Dreamforce conference on Agentforce last minute.

Benioff then decided to work out a new keynote, which was not final until the eve of Dreamforce. “We want to break the hypnosis,” he said. “The AI implementations with fundamental LLMs are do-it-yourself science projects. Companies don’t need do-it-yourself.”

AI is more than an LLM

“Enterprise AI is much more than an LLM,” adds Clara Shih, CEO of Salesforce AI. “There was a lot of confusion about that for 18 months. An LLM needs to be surrounded by entire technology stack in a company to enable productivity.”

With Agentforce, Salesforce states that it has that stack on offer. “Only Salesforce offers what you need,” Benioff continued. “We have an unfair advantage with the Salesforce platform. Customers already have their data in the Data Cloud, they’ve already built flows and integrated MuleSoft, the whole platform is already there.”

Benioff doesn’t shy away from big words. Although Salesforce is not really inventing anything new at Dreamforce, Salesforce is proposing an actionable combination of technologies that is easy to implement in a platform that many companies already have. Salesforce itself is launching several agents with relevant functionality within its industry-specific cloud offerings, and especially encourages customers to develop their own agents.

Objectives

Agents can help companies scale capacity during peak periods without the need for additional people, we hear. Agents will take over daily boring tasks and give people more space to do more inspiring work. We even see that agents will prevent burnouts because they give employees extra oxygen.

Whether it’s all going to work out that way? Benioff expects so. He backs up his supperlatives with concrete goals: “By the first of February, we want to have thousands of customers with Agentforce. Next year at the next Dreamforce conference, we are aiming for one billion consumers who have interacted with an agent.

By February 1, we want to have thousands of customers already with Agentforce.

Marc Benioff, CEO Salesforce

Agentforce does have potential, especially for companies already deep in the ecosystem. Salesforce itself, meanwhile, is eliminating legacy code in its platform as quickly as possible. Everything is being converted to the Salesforce Core, which means applications speak the same language. That must be done for legacy applications and acquisitions, such as Slack and Tableau. Only then can AI agents integrate cleanly with the entire platform.

An organization with a lot of data in the Data Cloud and Salesforce Foundations, can get started quickly. A conversation with an agent will cost two dollars an hour (or less). In many cases, that’s easy to justify. After all, which is more efficient: having a paid workforce send a hundred leads a personalized follow-up email, or automatically entrusting that task to an agent?

Big words?

Still, Benioff is quite brutal in his communications. Copilot, so to speak, has gone all the way and all other implementations of AI are relegated to do-it-yourself projects that don’t deliver value. Agentforce, he says, is the solution. In practice, generative AI has not been around for long. Everyone wants an immediate return on their investment, and Agentforce has the potential to deliver on that quickly for customers who are sufficiently in the ecosystem.

read also

No Clippy, but Sophie: AI according to Salesforce passes the (Microsoft) Copilot

Along the other hand, you can argue that generative AI is very new, and evolving very quickly. No one knows what the best implementation is, including Benioff. “We are making a bit of a gamble here,” he admits. AI models are continuing to evolve, frameworks are maturing, and AI is gaining enterprise frameworks. For example, within Cortex AI, Snowflake also envisions an AI implementation on a secure data cloud equipped with retrieval augmented generation.

Benioff and Salesforce tell the story with more focus and have concrete examples of early users to back it up. In this, Agentforce is unique. Salesforce also indicates that thanks to the Einstein Trust Layer, the AI technology is reliable enough to work not as a human’s helper, but as an independent colleague.