The big bets are on as Salesforce pitches the need for enterprise transition from model to system level AI

Salesforce AI Research has announced AI Foundry, an initiative focused on developing transformative AI innovation for the enterprise and built around the basic premise that there needs to be a shift from model-level to system-level AI.

According to the Salesforce AI Research team, the hardest challenges in enterprise AI live at the system level where components work together to deliver accuracy, consistency, and reliability at scale, a case in point being autonomous agents that need to communicate across organizational boundaries without compromising security.

The old consumer rules don’t apply at enterprise level, explains Itai Asseo, VP of Salesforce AI Research:

The models that we all know and love and are all using today  in many senses, have matured and saturated and reached some of the scaling laws. While many of those models are very, very powerful on a consumer level, when we are thinking of them for enterprise grade and the enterprise scale, some of these models don't really address the biggest problems that we see every day from our enterprise business customers.

[At] the model layer, where for the past few years we've really been focused on, there's been a lot of focus on how do you get the right responses and prompting from from these models. The way that we've seen that evolve is that we're moving from the so-called model layer to what we call the system level AI.

System level AI is going one step above, because there are many, many different things that you need to create, especially for enterprise and for businesses, to make this AI actually work in confidence and predictably and consistently. Things like the agentic memory or orchestration, and some of the things that you've also seen from from Salesforce, like observability and other things, These are the things that actually make AI work for businesses.

Bets

AI Foundry is pitched as bringing together AI research, strategic customers, and academic partners to develop, test, and validate new AI capabilities designed to move from foundational research to product innovation faster than ever before. Asseo expands:

The AI Foundry is an initiative from AI Research to really focus on these, what we call 'big bets'. These are the big bets that we're focused on for for this year in terms of taking foundational research, the amazing work that our researchers and engineers at AI Research are really trailblazing, inventing new ways of creating these system level AI models in different systems so that we can bring that more quickly to our product.

AI Foundry is focused on three strategic areas where Salesforce AI Research sees the greatest opportunity for enterprise impact - simulation environment, agents that learn from experience; ambient intelligence, agents that disappear into the environment; and agent-to-agent ecosystems, agents that interact on behalf of companies across organizational boundaries. Asseo goes on:

To validate these technologies and make sure that these are addressing some of the biggest challenges, we work very closely with our most strategic customers,. That allows us to accelerate the way that we bring that innovation to our product by partnering very closely with our engineering teams and our product teams.

Communicating the importance of this transition from model level to system level to customers is another challenge ahead. Asseo pitches Salesforce’s Slackbot as a useful of exemplar, with the bot using LLMs (Large Language Models) to actually understand and carry out a certain reasoning. But he’s quick to emphasize:

It's not working in a vacuum of just prompting an LLM. It's grounding its answers. It's being able to orchestrate certain tasks. So, I think Slackbot is a fantastic example for system level AI. Of course, there are many others out there as well. Something that was big in the news recently was OpenClaw that was, you know, that took the world by storm. You know, OpenClaw did not actually point to any one specific innovation. It was actually a combination of a lot of different things that were put together in a specific way to create a system level AI that created new types of possibilities.

I'm sure there's many other examples that you can see, whether it's within Salesforce or in the broader ecosystem, but this is really the big trend that we're seeing, that movement from model-centric and just having a model almost as a product, to having a model as one piece, an important piece, of a larger architecture that is really changing the behavior of how people and businesses are really working together with AI.

Real world

How people really work is, of course, another vital factor that needs to be taken into consideration. For example, quite how much involvement/assistance does a salesperson want or need before that actually gets in the way of a successful conversation with a prospect? Asseo says:

This is not really any more about the technology itself. This is about human behavior, and this is about how we're able to perceive certain things. Especially today, we have this information overload of so many different things that we're doing, and you're saying, 'OK, so now I'm supposed to have a conversation and you're going to add additional information while I'm just trying to have a connection?!?’.

One of the key features that we have, and we're still testing it and validating it is, how much content should you really have, and how should it be displayed?  There's one option of having just very, very small bullet points, just a few words to jog your memory. Or if you're in listening mode, you can have a more verbose type of answer. The other thing is, this ambient type of notification that you can glance at it, and if you want to read more, you can click on that and kind of expand that. So the human behavior and the user experience is actually one of the key things that we're working on beyond the technology.

And this is model-to-system AI journey isn’t something that’s going to happen overnight, cautions Asseo, with use cases emerging over time:

We're focusing on Service right now, Marketing obviously has some really interesting use cases, but right now, that's not our focus. We're trying not to spread too thin in terms of trying to boil the ocean.

My take

LLM realism in action.