Marc Benioff on AI Agents, Human Work and Salesforce’s Strategy

Matthew Berman welcomes Salesforce co-founder and CEO Marc Benioff following a Salesforce event at the St. Regis. On-screen branding identifies the interview as a partnership with Salesforce. Berman asks how agents might change work and communication, starting with Slack as an interface.

Berman asks how Slack can be a primary interface while Slackbot also reaches people wherever they work. Benioff credits Salesforce chief futurist Peter Schwartz with urging the company to acquire Slack years earlier, saying Schwartz anticipated the need for an open, conversational interface to AI models.

Benioff says Slack’s ecosystem has become richer than he expected and describes Salesforce applications as increasingly Slack-first. Berman asks whether traditional Salesforce interfaces might fade behind agents and the data layer. Benioff says Lightning still has a role, but sees Slack as a strong conversational workspace for Salesforce, Tableau and outside products.

Asked whether Slackbot should work beyond Slack, Benioff says it should be a composable capability available in Salesforce applications and other collaboration tools, including Microsoft and Google environments. He cites Sales Cloud and Service Cloud as examples of where the agent should accompany users.

Berman asks how AI will change team structure: will companies use smaller teams or pursue larger ambitions? Benioff says both can happen alongside a proliferation of agents coordinated by humans or AI. He imagines agents helping with sales, customer service, qualification, marketing and customer conversations.

Benioff argues that language-based work is especially suited to language models, including service conversations and coding. He contrasts current tools with his own early experience in assembly-language programming. These are his expectations rather than measured results across all tasks.

Berman says agents are producing more work, but he feels like a bottleneck because he must still prompt and verify them. Benioff agrees that humans are needed in the loop: models can become confused and inaccurate, making verification essential at this stage.

Benioff says Salesforce’s support agent handles initial conversations on its help site but roughly half the calls reach a point where a customer needs a person. He describes handing the conversation and context to a human through an omnichannel supervisor, enabling the person to synthesize the problem. The figure is his company-specific account.

Berman asks whether better tools could let people verify agent work at greater scale. Benioff expects such tools and more accurate models, and speculates that future systems incorporating more kinds of sensory information may improve decisions. He treats those capabilities as a future direction, not an existing guarantee.

Asked about generalists and specialists, Benioff turns to software engineering. He says Salesforce has about 15,000 engineers who can use coding tools and supervise agents, but maintains that engineers remain necessary because models cannot yet work autonomously at the required level.

Benioff estimates Salesforce’s engineering organization is more than 30% more productive, though not twice as productive. He points to hiring by AI companies as evidence that even organizations with advanced models still need people. The interview gives no measurement method for the productivity figure.

Berman asks why other firms are cutting staff if AI increases human productivity. Benioff says Salesforce itself has rebalanced its workforce and reports more than 83,000 employees at the time. He cautions against treating every reduction as having the same cause.

Some companies, Benioff argues, are reducing costs; others face data-center commitments or are shifting the mix of workers in response to AI. He calls it too easy for executives to make AI a scapegoat and urges CEOs to be specific about the reasons for changes.

Asked what a new graduate should study, Benioff recalls an MIT computer science student who wondered whether to change majors. He says he tried to reassure and recruit her, and describes Salesforce’s interest in interns and graduates from demanding universities. His account reflects Salesforce’s stated hiring needs, not a guarantee for individual job seekers.

Berman asks which functions besides engineering could be transformed. Benioff highlights sales, saying face-to-face communication remains vital. He cites approximately 15,000 salespeople and 15,000 engineers at Salesforce, along with a large Salesforce and Slack customer base, as his description of the company at the time.

Benioff argues that AI tools can help engineering leaders take on some product, design and marketing work, while marketing leaders can begin work that previously required engineering. He wants organizational structures to adjust around customer success and Agentforce adoption, while emphasizing the practical work of implementing products with customers.

Berman sees an opportunity for generalists: persistence can let someone start a project outside their formal role. Benioff agrees that someone with an idea can now get further into building it, although an engineer or other specialist may still be needed to finish.

Berman mentions a preference for demos over presentations. Benioff adds a qualification: seeing a demo can make people think a product is complete, but rapid prototyping only goes so far. Building a dependable product requires filling in the remaining work.

Berman asks why AI inspires such intense interest in San Francisco and whether Benioff feels it too. Benioff connects the city’s current enthusiasm to its history of social movements, entrepreneurship and technological change, describing innovation as part of the place’s culture.

Benioff says his own early AI excitement came around 2012 or 2013 as he watched models emerge from Stanford. He links that period to Salesforce’s Einstein work and acquisitions. He also claims Salesforce played a role in prompt engineering’s origins; that historical claim is his and is not established by the interview.

Benioff says Salesforce held roughly 1% of Anthropic and had been interested in AI for years. He believes society remains at an early stage of AI adoption, while noting that some countries use more AI-enabled surveillance and facial recognition. He distinguishes possible beneficial uses from uses he considers harmful.

Berman asks about the Anthropic partnership. Benioff says Salesforce wanted to invest in OpenAI and alleges Microsoft blocked the investment; the interview does not establish the other parties’ account. He says Salesforce then invested in several model companies, including Cohere, Mistral and Anthropic, and reports about $330 million invested in Anthropic by that point.

Benioff says Anthropic and OpenAI both participate in Salesforce’s model ecosystem. He describes a stack beginning with language models and continuing through an integrated and harmonized data layer, which Salesforce calls Data 360. In his view, data quality and integration are central to useful enterprise AI.

Above that data layer, Benioff places applications such as Slack, sales, service, marketing and Tableau, followed by Agentforce and agents that work with customers and employees. He also imagines local and employee agents and a supervisory or control layer spanning the architecture.

Berman asks whether Benioff has installed OpenClaw. Benioff says he tried it on a separate iMac but does not consider it enterprise-grade. He describes an internal Salesforce research project called Albert intended to explore a trusted, secure, reliable local agent that could work in Slack and other applications.

Benioff says Salesforce wants customer-facing, employee-facing and local agent capabilities and is willing to incorporate technology developed elsewhere. He cites the acquisition of Qualified and its Piper agent as part of that approach.

Berman asks which AI future worries him most, naming automation, concentrated power and surveillance. Benioff compares AI’s safety challenge with his longstanding criticism of social media, especially risks to children and misuse of platforms.

Benioff refers to reporting about chatbots and harm to children, and says model providers need to take safety and trust seriously. He argues that growth and safety must advance together if the industry is to build durable products. These are his concerns and policy judgments, not a clinical assessment.

Berman asks how to address public concern about AI. Benioff advocates stronger action, possibly including regulation, and draws an analogy to restrictions on young people’s social-media use in some jurisdictions. He says safeguards can come from industry or government and should balance safety with avoiding excessive regulation.