LLMs are important, but they're not up to muster just yet in terms of being viable human replacements, a lesson that some CEOs have yet to learn, it seems.
I think we've got this right - humans and agents working together, this idea that we have a role, and the agents have a role.
Alongside the big pitch around the new role of Slack and Slackbot as Salesforce dis-intermediators, CEO Marc Benioff took time out yesterday to make the case for the criticality of the human/AI balancing act that organizations need to get their strategic thinking working towards. Human intelligence has to co-exist alongside its artificial counterpart, is the thrust of the thesis. It’s not an either/or scenario, he posits:
When you look at the power inside of a company for the agentic layer to become an Agentic Enterprise, what we say [is] that all of a sudden, now I have my agents who can help me service my customers, sell to my customers, qualify my customers, market to my customers, have conversations with the customers. Agents are built on Large Language Models. The key word is language. It's good at speaking language. So things that are basically impacted by language, like a Service conversation, are going to be dramatically better using these tools.
The human is the bottleneck here in many ways, he adds, and that’s a good thing:
These Large Language Models are still wildly inaccurate times. All of us have had the experience where all of a sudden we're using a Large Language Model and it gets totally confused, and it's taken a right turn, and it doesn't know what we're talking about. So it's very critical that human beings, at least at this stage of AI, stay in the loop. It's maybe not forever, but right now, I think it is extremely important because the accuracy levels are just not there.
Just far enough and no further...for now
It’s critical to know just how far agents can go and when humans need to intervene, he adds:
A good example is if you go to help.salesforce.com right now, you know that you're working with Agentforce to resolve your customer service issue. But for about half the calls, it hits a moment where that customer is like, 'I need to speak to somebody', and the agent goes, 'You're right,' and boom, through our omni-channel supervisor, it goes to a human who then gets a screen in front of them. That screen - it could be in Slack, it could be in Lightning [Salesforce UI] - is looking at all the conversations, all of the information. Humans are really good at that moment, looking at a screen going, 'Actually, the problem [appears] to be this'. We're just very good at synthesis. That human being won't be 100% accurate, but, for the Large Language Model, is still needed a very high percentage of the time
Things will change, he points out:
New models that will be more accurate over time, as we have new AI breakthroughs. Remember, we are in a continuum of models. For decades, we've had more and more models, and now we're at the Large Language Model stage, We'll move to world models, but we're going to move to other kinds of models that we can't even describe yet that will just, over time, radically increase accuracy.
Our Chief Scientist Silvio Savarese has the vision that we're moving into multi-sensory models. We haven't really seen them yet, but the reason why that's important is the model then just has a lot more data to be able to make a decision, not just language, but all these other kinds of sensory data that we have. We have our eyes and ears and our brain and our memories and all these things, and we're able to kind of make these decisions. Eventually the models will get to that point.
Death of the expert?
But does the increased role of agents already imply the 'death of the expert' and the rise of the generalist within organizations? Up to a point, concedes Benioff, arguing that this is especially true in software engineering:
Obviously, we have 15,000 software engineers at Salesforce, each one of them an 'A-player'. The key thing about these 15,000 engineers is that they're all out there, all over the world. All of them can now be hugely augmented with these coding models. It could be Anthropic, it could be OpenAI, it could be Cursor AI, it could be others, but when they start to use these models, they're now working not only with the AI, but with agents to help them code.
But, he adds, the human engineers are needed:
The models still cannot operate autonomously. We're not at that level of AI yet, so it's really critical. Our engineering organization is probably more than 30% more productive, but I wouldn't call it 100% more productive. That's why, even in the top AI companies, if you go to their job boards, you'll see they are hiring a lot of engineers. They're hiring a lot of everybody, because they have to build companies of humans. Even though these top AI companies have these unbelievable models, they need a lot of humans. That's probably the canary in the coal mine [by which] we know that the models are not at that level yet.
So, hiring humans is not a thing of the past whatever some of the more ‘glass half empty’ commentators would insist upon. But, of course, we’ve seen a lot of companies, tech and non-tech, laying off huge tranches of their workforce, with AI thrown into the mix in some form as justification. Just this week, Benioff’s alma mater, Oracle, was the latest to announce sweeping headcount reductions.
Salesforce itself was the subject of negative mainstream media headlines last year stemming from the idea that it had laid off 4,000 of its support personnel, a mis-interpretation of what actually happened which saw most of those people re-deployed into new areas of the company rather than shown the door. In fact, says Benioff:
Salesforce right now has hit a new record number of employees, more than 83,000, but the balance of those employees and where they are placed is different than where it was five years ago. We went through an uncomfortable period ourselves over the last five years of re-balancing our workforce. That's just a difficult thing for any company. In some cases, these companies are cutting because their costs are just too high; in other cases, these companies are cutting because they've made financial commitments, specifically to data centers that they have to pay for; and in other cases, these companies are cutting because they need to re-balance their workforce to reflect the changes in Artificial Intelligence.
These are not all the same reason, he points out, although that’s not a message that’s necessarily being widely understood, he admits:
You cannot bucket all these companies together. If you do, you're making a fundamental mistake. Even though I've spoken about this, I think somewhat aggressively, I don't think most people still really understand what is going on. It's too easy to basically take AI and make it the scapegoat. I think for some CEOs, it's the lazy way out. That's up to them. I think it's better just to say, ‘Here's what's really going on’, and trying to be specific as you can. You're going to take bullets no matter what, because that's your role as CEO, and then you have to kind of get go forward and put everything back together.
Spoken with the voice of experience there, but the wider point is well-made - organizations cannot do without human intelligence and that has to be sought out, brought on board, nurtured etc etc. Benioff goes on:
We're hiring interns. We're hiring the Freshman class. We want to send our recruiters out to the Top 25, high academic threshold universities, like MIT, and bring top computer scientists and others to Salesforce, because we badly need that talent. That idea that those people are so critical for a company like ours, that is really part and parcel of what I think the discussion is, because for a lot of them, they've been told, 'Well, it might be a difficult for you getting a job for the summer, or getting a job when you graduate'.
He concludes:
The reality is that companies like ours badly need these people. We need to kind of put two-and-two together to really show that this is still going to be a critical part of our workforce going forward.
April 1, 2026
