Marc Benioff on Slack Agents, AI Jobs and Enterprise Safety

Matthew Berman interviews Salesforce co-founder and CEO Marc Benioff about AI agents, work and the company’s direction. The conversation follows an event where Salesforce discussed Slackbot and its role in business workflows.

Berman asks how Slack can serve as the main agent interface while an assistant also follows users into other tools. Benioff credits Salesforce futurist Peter Schwartz with arguing years earlier that a conversational, open interface would matter as AI and agents matured.

Benioff says Slack’s ecosystem has grown beyond what Salesforce expected and that the company wants its applications to become increasingly Slack-first. He sees conversational access as useful not only for Salesforce customers but also for Tableau and third-party applications.

Asked whether the traditional Salesforce interface will fade, Benioff says Lightning still has a role alongside Slack. He describes Slackbot as a composable assistant that should be available in Salesforce apps and potentially in other collaboration environments, including Microsoft Teams and Google Workspace. These are his product ambitions, not proof that every integration already exists.

On company structure, Benioff expects a range of team sizes alongside more agents coordinated by people or AI. His central view is that humans and agents will work together in service, sales, marketing and customer conversations, rather than that one fixed organizational structure will suit every company.

Berman says reviewing and prompting agents is making him a bottleneck. Benioff agrees humans can slow the process, but argues that present language models remain inaccurate enough that keeping people in the loop is essential. He expects more accurate models and better verification tools over time; that is a forecast.

Benioff gives Salesforce’s customer-service system as an example: he says roughly half of calls reach a point where the customer wants a person, who can review the conversation and use judgment. The figure is his description of Salesforce’s operation and is not independently verified in the interview.

Benioff predicts future models will use more kinds of sensory information and improve accuracy. He invokes a progression from current language models toward world and multisensory models, while acknowledging that some future forms are not yet clear.

Berman asks whether AI is reviving the generalist. Benioff says engineers can supervise coding agents and use them across a wider range of work, but he does not believe the models operate reliably without humans. He says Salesforce has about 15,000 engineers and estimates its engineering organization is over 30% more productive; the method behind that estimate is not given.

Berman raises companies that are cutting staff despite optimism about AI. Benioff says the causes can differ: high costs, commitments to data centers or a rebalancing of skills. He argues that using AI as a blanket explanation can become an easy scapegoat, and CEOs should explain their decisions specifically.

Benioff says Salesforce has more than 83,000 employees after a difficult period of workforce rebalancing, with a different distribution of roles than five years earlier. His headcount and characterization refer to the time of the interview and should not be treated as a current staffing figure.

Asked for advice to recent graduates, Benioff recalls an MIT computer science student who wondered whether to change majors. He says he tried to recruit her and that Salesforce was hiring interns and early-career talent. He remains concerned about the uncertainty students face but argues that technical talent remains important.

Beyond engineering, Benioff points to sales: he says Salesforce has about 15,000 salespeople and that face-to-face communication and explaining a product’s value still matter. These staffing numbers are his account of the company at the time.

Benioff argues that AI can soften boundaries between engineering, product, design and marketing. An executive may be able to prototype outside their traditional specialty, but the company still has to work directly with customers to make products succeed; adoption cannot simply be delegated to a language model.

Berman welcomes the idea that persistent people can begin building outside their job descriptions. Benioff agrees that a non-coder with an idea can get further than before, while noting that an engineer may still be needed to finish. He cautions that a convincing demo is not the same as a complete product.

Berman asks why AI occupies so much attention in San Francisco. Benioff connects the enthusiasm to the city’s long culture of innovation and transformation, while saying other places have different local experiences. The account is his interpretation of the city’s culture.

Benioff says his own interest in AI intensified around 2012–2013 and describes Salesforce’s earlier work on Einstein. He also makes a claim about Salesforce research and the origins of prompt engineering; the interview does not substantiate a claim of exclusive invention.

He says Salesforce owns approximately 1% of Anthropic and has invested in AI for years. He also warns that some places have adopted AI-enabled surveillance more extensively, illustrating both positive and negative uses; he regards society as still early in the transition.

Berman asks about the Anthropic partnership. Benioff claims Salesforce wanted to invest in OpenAI but Microsoft blocked a corporate investment, and says Salesforce then invested in several model companies. This is Benioff’s account; the interview does not independently establish Microsoft’s role or the underlying terms.

Benioff says Salesforce had invested about $330 million in Anthropic and that both Anthropic and OpenAI models are part of the Salesforce ecosystem. The amount and ownership estimate are his statements at the time, and are not a valuation or a promise of future returns.

He describes Salesforce’s stack as models, a connected and harmonized data layer, applications, and an agent layer. Products such as Data 360, MuleSoft, Slack, sales and service applications are meant to work together; he argues that weak data integration limits the quality of the resulting AI.

On top of these layers, Benioff envisions customer-facing, employee-facing and local agents, with a supervisory layer helping coordinate them. He frames this as Salesforce’s architecture and strategy, not a universally required design for all agent systems.

Berman asks whether Benioff tried OpenClaw. Benioff says he installed it on a separate iMac and liked it, but considers it insufficient for enterprise requirements. He describes a Salesforce research project called Albert to explore a secure, reliable local agent that could work with Slack and company applications.

Benioff says Salesforce wants different agent products and architectures to work together, mentioning its acquisition of Qualified and its Piper agent. He presents Albert as a research project and an intended direction, not as a finished product with verified enterprise guarantees.

Asked what AI future worries him most, Benioff points to safety failures and draws an analogy with social media. He refers to reports of harmful chatbot interactions involving children and argues that infrastructure providers should make trust and safety a core part of product development.

Benioff says improving public trust requires safeguards and may also require regulation. He discusses age restrictions and other guardrails as possible approaches, while saying he does not want excessive regulation. His remarks are a policy view, not a complete description of laws in any particular country.