At the TDX 2025 developer conference in San Francisco, Salesforce executives presented its Agentforce agentic AI technology as a “whole system” approach, where large language models (LLMs) are less significant than a “trinity” of data, applications and agents. Relatedly, they consistently disparage “DIY” artificial intelligence (AI) programmes.
Paula Goldman, the supplier’s chief ethical and humane use officer, said: “I think a lot of the public discourse about AI has been about [large language] models. But if you think about Agentforce, it’s a whole system. There’s a foundation model, and then there’s a series of smaller models that go into our Atlas system, and there are workflows that are automated that people can draw on. We’ve got used to talking about AI as models over the past few years, but I think we need to be talking about systems.”
David Schmaier, president and chief product officer at Salesforce, said the supplier’s entire technology stack, including Slack and Tableau, comes into play with Agentforce. He also pointed to its Data Cloud platform as central to its AI offer.
“You couldn’t have a computer without a microprocessor; you need storage and RAM and a display and an operating system around it. That’s what we’ve done. We have our data cloud, which harmonises hundreds of thousands of systems. It gives you the data, the metadata and the semantics. That’s why we can outperform an LLM by itself. LLMs have hallucinations, they have bias, toxicity. An LLM is necessary but insufficient. We add to the LLM. Our view is the data powers the AI and then the AI powers the customer experience of the future,” he said.
An LLM is necessary but insufficient. We add to the LLM. Our view is the data powers the AI and then the AI powers the customer experience of the future David Schmaier, Salesforce
“We call it the ‘holy trinity’. We have the Data Cloud, then we have our Sales Cloud, Service Cloud and Marketing Cloud apps – which is how we got the name Salesforce – as well as Slack, Mulesoft and Tableau. And now we have Agentforce on top of all that. That’s how we can turn on 10,600 customers over three days with agents. It’s because we are using the same platform as we have for 25 years. So, with a healthcare company, for example, that has workflows it has bult in its Salesforce deployment, it can make all those available for [virtual] agents,” Schmaier added.
He believes too many organisations are doing DIY AI. “Most people are just trying to take whatever apps they have, whether it’s Salesforce or SAP or Workday, and just buying ChatGPT and trying to plug it in. No other competitor has what we have, in terms of agents. We think we have a real lead in this agentic field. We’ve sold to 5,200 customers since launching at Dreamforce [in September 2024]. Now, we have 200,000 customers, and most don’t use Agentforce today,” he said.
Rahul Auradkar, executive vice-president and general manager of Unified Data Services and Einstein at Salesforce, made a similar argument about what the provider calls DIY AI.
“What we are doing with agents is an entire system. We’re not shipping a model, an app or a copilot. We’re shipping an AI system on a deeply unified platform. What that system allows our enterprise customers, who don’t want to do the DIY, to do is surface customer-centric analytics and workflows, and listen to the customers to feed back to the system so the agents get better. Copilots are a narrow sliver of what AI can be,” he said.
“The difference between a DIY AI and an enterprise using [our] system is that the enterprise can focus on things that they are good at, which is plenty of things. They have their data. The have their transactions. They have their engagement data. They have their AI policies, their workflows, their automations. We bring all that together within a deeply unified platform and drive value for our customers,” added Auradkar.
DIY AI programmes strongly in evidence among users
And yet, analyst research from Informa TechTarget’s Enterprise Strategy Group (ESG) offers a contrast with Salesforce’s disparagement of DIY AI – a complicating contrast rather than a confutation, but a contrast nevertheless.
Towards the end of 2024, ESG surveyed 832 professionals at organisations across the globe involved in the strategy, decision-making, selection, deployment and management of generative AI (GenAI) initiatives and projects at their organisations and familiar with their organisation’s use of third parties to support GenAI initiatives.
The resulting report, The state of the generative AI market: Widespread transformation continues – authored by Mark Beccue, principal analyst, Mike Leone, practice director and principal analyst, and Emily Marsh, associate research director – does find support for an agentic AI philosophy: “Respondents most often said that they see AI agents, virtual assistants, and intelligent chatbots powered by AI as valuable productivity tools, though they also often said they view them with cautious optimism (41%). Over two-thirds of organisations are planning for or considering AI agents, which represents a significant opportunity for AI vendors to target these requirements with capabilities and services.”
They also note, however: “The AI agent market is extremely nascent and loaded with challenges, including managing single-task agents, interoperability problems, the potential emergence of multitask agents and security.”
But the authors also remark, similarly to Salesforce’s Auradkar, that: “A wide majority (84%) of respondents agreed it is important to incorporate their own enterprise data into models that support generative AI. GenAI models themselves are not a competitive differentiator. Rather, effectively identifying, organising and vetting internal data for use with GenAI models is the key to creating unique and highly actionable insights.”
The research also found user organisations to be embracing a variety of LLMs – open source and proprietary. The largest percentage of respondent organisations (43%) are both proprietary and open source models.
Alongside this enthusiasm for using large language models, the study found that organisations are placing “their bets on internal resources, planning to reskill or upskill employees (58%) and provide education and awareness training to employees (43%)”. This suggests a growing cadre of employees who will want to do DIY AI.
The authors comment: “Employee enthusiasm for these technologies is likely at a high point as GenAI excitement pervades many facets of society, so this internal investment will likely be a win-win situation whereby personnel receive welcome development opportunities and the business gains valuable GenAI expertise.”
At Dreamforce in September 2024, Marc Benioff, co-founder, chairman and CEO of Salesforce, was in combative mood in respect of Agentforce, positioning it as a wholescale alternative to generative AI copilot usage, associated with Microsoft and Google, but with other vendors too.
“There’s a lot of narratives out there from vendors, and a lot of it is not true,” he said at the time. “You need to sit with those customers [at the Dreamforce event], look at the code and break the hypnosis coming from all the vendors. There’s plenty of real customers here who are really deploying real AI. But there are billions being invested in copilots, delivering how much productivity increase? Is there a better way to do it? And so, that’s our gambit.”
The game is still being played. The middle game lies ahead.
