Salesforce intros action models to fuel AI agentic workflows

xGen-Sales and xLAM

XGen-Sales is a large action model, a model that produces actions instead of texts or images like large language models (LLMs) do. Users who finetune the model can automate sales tasks such as generating customer insights, summarizing calls and tracking the sales pipeline, Salesforce said.

Salesforce also introduced different sizes in the xLAM line: xLAM-1B, xLAM-7B, xLAM8x7B and xLAM-8x22B.

With a range of small to large sizes, even the smaller xLAM models are still powerful, said Shelby Heinecke, senior AI research manager at Salesforce.

“Smaller models are super important,” Heinecke said in an interview with TechTarget Editorial. “Small models are super-fast and [have] comparable performance to models many times their size.”

The xLAM-1B, which Salesforce describes as a “tiny model,” is suitable for on-device applications, Salesforce said. The model can create AI agents or assistants that run on smartphones or other devices with limited computing resources.

The small xLAM-7B model can be used to create AI assistants that perform planning and reasoning tasks. The medium xLAM-8x7B is a mixture-of-experts model that is ideal for AI agents within industrial applications. The large xLAM-8x22B is also a mixture-of-experts model for organizations looking to create AI agents with intensive computational resources.

Non-commercial open source versions of the xLAM models are available now on Hugging Face for community review and benchmark testing.

AI agents

The introduction of the xGen-Sales and xLAM models coincides with recent activity in the AI market around AI agents and a surge in development of large action models.

The xGen-Sales and xLAM models are meant to power agentic (or agent-driven) workflows, Salesforce said.

They provide the power needed to create an AI system that can not only draft an email but also take the action of sending the email, for example.

"Large action models will be a very important underpinning model component of these agents because action models emulate human behavior," said Gartner analyst Arun Chandrasekaran.

Looking to get out front

With the introduction of its new LAM, Salesforce is attempting to move into a leadership position in the AI market despite not being known as an AI vendor as are tech giants Google, Meta and Microsoft, said Dion Hinchcliffe, an analyst with Futurum Group.

"It's a good move to demonstrate leadership, to be first and brave," Hinchcliffe said. "If they don't get it quite right, like more proven AI companies do, then it'll look like a misstep."

However, Salesforce’s strategy differentiated from the big cloud vendors with their substantial AI portfolios is that Salesforce will buttress its sales applications with LAMs like xGen-Sales, said David Nicholson, a Futurum analyst.

“Salesforce knows Salesforce better than anyone, and so it can look at all of the workflows that humans are involved with from a Salesforce perspective,” Nicholson said. "Salesforce is creating something for itself and its users, just like an individual enterprise might engage Google to create something for itself, the end user large company and its users and customers and ecosystems.”

On the other hand, while Google also creates models for its own applications, most of them are outward facing for customers’ use.

Moreover, since Salesforce is focused on business users primarily, its use of LAM technology such as xGen-Sales can help customers who find prompt engineering challenging, Chandrasekaran said.

These users will benefit from the AI agents fueled by LAMs to help them perform business tasks, he said.

Challenges with LAM

While AI agents and large action models are relatively new additions to AI technology, they also add a new level of complexity compared to already hard-to-develop LLMs.

"Large action models are a lot more complex to build ... and a lot more expensive to build as well," Chandrasekaran said. A few startups have tried to build them in the past and realized it would cost a lot more money than anticipated.

LAMs are also complicated to use because the models need to interface with other services such as enterprise data services, he added.

The consequences with LAMs could also be more serious than those of LLMs, Hinchcliffe said.

"It's really scary that you're going to have AI doing things for you," he said. Since the AI technology will be able to directly call any existing IT system, it can do different tasks such as file taxes or use a bank account. "The bad guys can use these to do all sorts of things that we may not want them to do," he added.

Another challenge for Salesforce will be getting people to use its models, Nicholson said.

"The real test for this is, frankly, a year from now whether everyone that uses Salesforce will be using these tools to enhance productivity," he said.

Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.