Inside the Agentic Enterprise: Seven Conversations at Dreamforce 2025

Matthew Berman and co-host Nick Wentz open their October 15, 2025 live show at Dreamforce in San Francisco. They introduce a lineup spanning Formula 1, Salesforce, LIV Golf, Anthropic and Figma. Their first guest is Claire Sparks, F1’s head of technical initiatives.

Sparks oversees corporate technology projects and research with technology partners, aligning people, resources and investment. Asked about AI, she describes F1 as early in its adoption: the organization is exploring how technology can support a relatively lean workforce, streamline operations and communicate with fans.

Sparks says F1 itself has around 650–700 employees and cites a global fan base of 827 million. She credits entertainment such as Drive to Survive and the F1 movie with helping the sport reach new audiences. She says 43% of fans are female and 42% under 35, figures she presents as the organization’s view of its audience.

Sparks says her team considers return on investment alongside alignment and staffing. She frames AI as support for people who already know their jobs, not a substitute for understanding how F1’s operation works.

Only about 1% of the fan base typically gets to attend a race, Sparks says. That makes broadcasts, films and digital experiences important ways to help people feel connected. Asked how racing teams use AI for performance, she says the teams themselves are better placed to answer and turns instead to F1’s own operations.

F1 moves across 21 countries for 24 races in the season she describes. With only one or two people in some specialist roles, Sparks sees AI as a possible way to reduce the risk when key staff are unavailable and help the organization maintain its broadcasts.

The hosts ask how F1’s fan base has changed. Sparks says older European male fans were once the familiar demographic, while newer fans bring different interests and come from more places. She sees those differences as a reason to rethink how the sport communicates.

The logistics are extensive: crews arrive around ten days before a race, and F1 brings equipment beyond the site’s power, water and concrete base. Sparks describes hundreds of cameras and microphones, cables and telemetry feeding a traveling broadcast center; the international feed then goes to the media center in Kent and on to broadcasters.

The traveling broadcast center assembles the international feed at each track, Sparks says. Kent then distributes that feed to broadcast partners around the world. She uses the example to show how much infrastructure sits behind what viewers see on television.

Sparks cannot disclose all the projects under consideration, but says technology partners including Salesforce, AWS, Lenovo, Globant and Tata Communications bring different specialties. After COVID, F1 moved to a mix of remote and trackside broadcasting, substantially reducing the number of people at the circuit.

Sparks describes the pandemic as a catalyst for changing broadcast operations. She says the organization retained the hybrid approach afterward and had significantly fewer people at the track, illustrating how a disruption can prompt a lasting operational redesign.

The hosts ask whether a future of autonomous cars might change people’s relationship with racing. Sparks emphasizes the human driver’s judgment, reaction time and awareness, and says fans also want to know the people behind the machines: their routines, fitness, mindset and stories.

The hosts connect Sparks’s emphasis on drivers’ stories to AI-generated media. Sparks says people want a look behind the curtain: how a driver stays focused for two hours and handles the physical demands of racing. That human connection remains central to fan engagement in her view.

For Sparks, reaching fans means finding where they already spend time, from streaming platforms to in-person experiences such as F1 Arcade. She also highlights the physical preparation behind driving at speed and says technology could expand broadcasts, products and internal operations, without committing to a particular future feature.

Sparks is reluctant to select a single future technology opportunity. She sees potential across broadcast, fan engagement and internal processes, while noting that racing constructors make their own choices about technology for their teams.

Asked about her favorite way to watch, Sparks praises F1 TV’s technical discussion and race strategy. At a circuit the cars pass quickly, she says, while a broadcast can explain tire choices, garage decisions and how teammates work together.

F1 TV provides commentary that Sparks finds useful for understanding team strategy and tire decisions. She notes that a trackside spectator sees the spectacle directly, but a detailed broadcast can explain what happens in the garages and why.

Sparks remembers watching Monaco as a child and later attending the Australian Grand Prix while living near Melbourne’s Albert Park. She was fascinated by the event logistics: promoters build grandstands, bridges and other infrastructure well before F1’s ten-day setup. Other racing series and attractions make the event larger than the race itself.

Salesforce chief scientist Silvio Savarese joins next. Asked whether models are ahead of enterprise deployment, he says an agent is more than an LLM: it needs memory and retrieval, a reasoning engine, the ability to act through tools, and an interface. The infrastructure connecting those pieces determines whether customers can use it effectively.

Savarese contrasts the pursuit of general intelligence with enterprise requirements. A model may perform impressive creative or mathematical tasks yet fail a routine business task; consistency and accuracy on specific workflows matter for customers. Salesforce builds models for developer productivity, trust and embeddings as well as using other models.

Salesforce’s in-house work includes models used for a trust layer and embeddings for retrieval-augmented generation, Savarese says. The company also uses broader infrastructure to connect data and actions; he does not present an LLM alone as a complete enterprise solution.

He describes Enterprise World Model simulation environments as a way to stress-test agents. In his view, learning from interactions and feedback offers another path as existing training data and scaling approaches become constrained. Agents attempt tasks in simulated settings, and success or failure can inform prompts, context and reinforcement-learning fine-tuning.

In the simulated environment Savarese describes, agents try tasks and receive feedback about failure or success. He says the resulting learning can inform both the context provided to an agent and further model fine-tuning, including reinforcement learning.

On accountability, Savarese says agents need sufficiently high performance before deployment. He argues that exams and theorem-solving benchmarks do not necessarily measure whether an agent answers the right questions in financial services, healthcare, sales or customer service, so enterprises need evaluations tailored to those tasks.

The research-to-product path begins with foundational research and customer partnerships, he explains. Teams identify pain points, develop prototypes and bring successful proofs of concept to engineering and product teams. He cites the Atlas reasoning engine as a project incubated with customers; he estimates a pilot may take months, with a robust beta taking roughly six months.

Savarese says a successful research idea is incubated with customers before becoming a product. His timeline is an illustrative account of how a project can progress, not a promise that every research project will reach production.

Savarese describes agents as potential workforce multipliers. People can choose and orchestrate tools for tasks, while a further orchestration layer might coordinate agents when the work becomes too complex. He also anticipates agents communicating across organizations, a future he frames as a topic for further discussion.

He imagines a further layer of agent orchestration for complex work, and eventually communication between agents across organizations. He presents this as a possible direction rather than describing a completed system.

Wentz asks whether managing agents could simply replace one kind of drudgery with another. Savarese stresses that people should remain involved when agents are uncertain or face conflicts and should provide feedback. He sees the interaction as iterative augmentation: a product manager, for example, can develop and revise a prototype with AI rather than simply request a finished one.

As an example of augmentation, Savarese describes a product manager using AI to build a customer-meeting prototype. The person supplies ideas, reviews the result and asks for revisions until the prototype meets the desired standard.

Asked whether humans will remain at the start and end of workflows, Savarese describes a transition rather than a fixed endpoint. Jobs and orchestration roles will evolve as people learn the tools, he says, but their precise form is difficult to predict while the technology itself is changing.

An area that excites him is ambient intelligence: instead of responding only to explicit requests, an agent might follow a conversation and proactively surface useful information. A sales representative could receive customer context before a meeting and relevant details during it, potentially through future glasses or audio interfaces.

For a sales meeting, Savarese imagines an assistant preparing relevant documents in advance and surfacing context during the discussion. He sees future glasses and earbuds as possible ways to deliver this information without requiring the user to look at a conventional screen.

Savarese imagines personal agents helping with shopping, travel and other tasks, changing how people interact with digital services. He also discusses family conversations with his wife, Fei-Fei Li, including education and their research interests: her work on spatial and world models connects with his earlier work in 3D vision.

Asked about life with fellow AI researcher Fei-Fei Li, Savarese says family logistics and children’s education are part of their conversations alongside research. He describes their complementary interests in digital agents and spatial or visual intelligence.

Scott O’Neil, CEO of LIV Golf, describes the league as combining individual and team competition with a global schedule and event experiences such as concerts. In the format he describes, 13 teams give players an additional identity alongside their individual standings.

LIV combines individual rankings with teams that may share a geographic identity, O’Neil says. He also emphasizes concerts, art and fashion around the golf, comparing the overall event to the atmosphere at Dreamforce.

O’Neil says LIV Golf was the idea of chairman Yasir Al-Rumayyan and that Greg Norman led the league before him. He draws on prior work at the NBA and other sports organizations, describing LIV as an early-stage, well-funded business trying to grow golf around the world.

The Salesforce partnership includes what O’Neil calls a fan caddie: an assistant intended to answer questions about players, the competition and where to find things at an event. He says about 30% of attendees have not been to a golf event before, making an accessible guide valuable.

O’Neil explains the fan-caddie idea through a story about a human caddie at a course in Northern Ireland. A knowledgeable guide can explain the course and help a newcomer navigate choices; he wants an analogous source of information for spectators.

O’Neil distinguishes that fan tool from Agent Caddie for broadcasters. LIV uses a shotgun start, putting players on different holes at the same time, which he says creates a more compact round but complicates the broadcast narrative. The assistant is intended to surface what is happening and what may be relevant next.

The broadcast assistant is intended to help commentators track different players simultaneously and decide which developing stories to cover. O’Neil compares it to support from a producer and says users can ask follow-up questions in real time.

The partnership was developed during the season, and O’Neil says the league is building as it goes and expects the offseason to allow improvements. Asked where to draw a line on AI-enhanced performance, he discusses the rise of training, nutrition and injury analysis rather than specifying a boundary.

O’Neil says the team format gives players and their families more community around a traditionally individual sport. He also describes a change over his career toward more deliberate attention to sleep, nutrition and conditioning.

O’Neil argues golf’s precise variables—club angle, speed, spin, wind and trajectory—make it a promising area for analysis. He says players commonly use tools such as TrackMan and a phone to measure shots; he has not seen much use of augmented-reality glasses in golf.

On business strategy, O’Neil says the league has high structural costs and needs revenue growth. He lists tickets, premium experiences, sponsorship and television as principal sources, with merchandise, parking and concessions contributing as well.

He compares LIV’s attempt to alter golf’s format with changes in other sports intended to improve pace or flow. He also praises golf’s traditions of integrity, resilience and community, making clear these are his views on the sport.

For remote viewers, O’Neil describes an app feature called Any Shot, Anytime, which he says uses AI to make tournament shots available on demand. His larger aim is to tell stories about players and their families, because technology can help connect spectators with the people behind the sport.

He recalls Abraham Ancer losing a lead and team captain Sergio García comforting him, using the moment to illustrate competition, empathy and team golf. He also discusses the first season of the series Live to Win, says it lacked some major stars, and hopes a second season will tell a fuller story.

O’Neil says the first four episodes of Live to Win aired on FS1 and elsewhere, including YouTube. He wanted the initial series released rather than held for perfection, but says a later season should feature more of the league’s best-known players.

Salesforce chair and CEO Marc Benioff joins the hosts. He is enthusiastic about Figure’s humanoid robot demonstrations and imagines household tasks such as cleaning and laundry. These are aspirations and predictions, not capabilities established by the interview.

Asked about the OpenAI partnership, Benioff says Salesforce wants to bring AI into its products. He describes OpenAI’s use of Slack and Salesforce tools as an example of the agentic enterprise he wants to build: people and agents collaborating across business workflows.

Benioff says he first saw OpenAI’s use of Slack through Sam Altman and later heard OpenAI COO Brad Lightcap discuss its workflows. He treats that customer relationship as an example of how Salesforce hopes to integrate agents into its own products.

Benioff recalls founding Salesforce around the idea of making software accessible over the web. He says he now uses ChatGPT to prepare interview questions, while recognizing that language models have limits. His question is how to bring comparable assistance into work.

On his phone, Benioff describes using Agentforce inside Slack for a customer renewal and says Salesforce’s help site can route a user from an agent to a person when needed. His example is meant to illustrate an agent embedded in the flow of work with human escalation.

The hosts ask whether ChatGPT or Slack is suited to lighter work while deeper work remains in Salesforce’s interface. Benioff rejects a rigid division and says the relevant distinction is the use case: agents might assist throughout sales, service and other processes.

Benioff argues that agents will persist across different business processes rather than belong only to a category of “light work.” He says Salesforce had historically missed many potential callbacks and claims its sales agent contacted 50,000 customers in the week discussed; the show does not independently verify the figure.

Asked whether personal and workplace agents might eventually become one, Benioff says current systems can confuse contexts. He expects an awkward transition with some separation between personal and business use and between functions such as sales, marketing and service.

He offers a field-service example involving equipment at his home. A technician using Salesforce can consult records and, in the proposed agentic workflow, ask an agent for help before escalating to a person. Benioff says the interaction can then become part of the service record.

In Benioff’s field-service scenario, the agent supplements the technician’s existing record of the customer, equipment and prior service. A human colleague remains available for escalation, and Benioff expects the interaction to enrich the record for future service.

The conversation turns to San Francisco, where Benioff says he is a fourth-generation resident. He celebrates the city’s history and describes AI as another local wave of innovation. He says Salesforce kept Dreamforce in the city and discusses event security and the city’s recovery from the pandemic period.

Benioff characterizes San Francisco as a place of recurring waves of innovation and says Dreamforce continued to bring people to the city. His account is personal and promotional, reflecting his connection to the city and his role as conference organizer.

Benioff cites company and personal philanthropy in local hospitals, schools and parks, along with volunteer work and Salesforce’s 1-1-1 giving model. These are his own figures and characterizations. He recommends the Presidio Tunnel Tops park as an example of a place for families.

Benioff says Salesforce and his family have supported local schools, children’s hospitals and parks. He also describes the company’s philanthropy and volunteer programs, presenting these as part of a broader obligation to give back.

On the future of work, Benioff says Salesforce had fewer service staff than a year earlier and more salespeople, describing humans and agents as working together. He claims comparable customer-satisfaction scores for 1.6 million human and 1.6 million agent conversations over nine months; those figures are his account, not an independent evaluation.

In the staffing example, Benioff argues some sales work still benefits from face-to-face communication. He says coding tools can make people more productive while keeping them involved. His reported service metrics do not by themselves establish causation for staffing changes.

Benioff argues that language models remain limited and that people will remain involved in work for some time. He is interested in world models and Fei-Fei Li’s research, and rejects claims that current systems have already achieved AGI. His definitions and timelines are opinions about a contested and evolving field.

Benioff thinks another generation of models may help address limits he sees in systems built around language. He mentions world models as one research direction and notes that modern AI builds on decades of earlier work.

He returns to Figure founder Brett Adcock’s ambitious robot predictions and contrasts them with the absence of robots walking through the conference that day. Benioff urges realism about present capabilities and links technological abundance to his interest in supporting public services and local communities.

Reflecting on ambitious claims about autonomous robots, Benioff urges the audience to distinguish current evidence from predictions. He returns to philanthropy and public institutions as part of his own vision for technological progress.

Paul Smith, Anthropic’s chief commercial officer, says he joined to help enterprises adopt AI while commercializing the research team’s work. After decades in enterprise software, he wanted to work at a frontier model provider and apply his experience with organizational technology adoption.

Smith says the pace of research and product releases drew him to a frontier model provider. His stated role is to connect that research with enterprise customers while maintaining Anthropic’s stated commitments to safety and trust.

Smith discusses the Salesforce partnership: Claude is used in Slack and, he says, Claude Code is used by 16,000 Salesforce developers. He describes a preferred-model arrangement within Salesforce’s trusted boundary for regulated customers and says the arrangement is designed around data controls and a private-cloud environment.

Smith describes a closer partnership in which the two companies coordinate product roadmaps rather than treating the model provider as an interchangeable supplier. He expects enterprises to use a variety of models and providers for different needs.

Enterprise customers may encounter the same model in several ways, Smith says: through startup products built on APIs, a direct Claude Code deployment to developers, Salesforce’s agents, or specialized providers serving a particular domain. He views those channels as additive rather than a simple choice between direct and indirect sales.

He anticipates large companies using Claude Code directly while also consuming Claude through products such as Agentforce and specialized applications. In his view, selling through partners and directly to enterprises can coexist.

Asked whether businesses built on frontier models can be profitable despite inference costs, Smith says Anthropic sees evidence of real enterprise value. He describes multiyear customer-care deals and banking use cases such as know-your-customer workflows, while emphasizing that revenue must ultimately come from customers with viable businesses.

Smith says Anthropic’s enterprise focus is intentional. He points to collaboration with cloud partners and the release of Claude Haiku 4.5 that morning, and argues coding tools can be an entry point to wider business use cases. His examples include fraud detection, customer due diligence and documentation in drug development; the benefits are presented as company or customer ambitions.

Smith gives examples of coding adoption leading to broader agent projects, from financial services to pharmaceutical documentation. These examples illustrate his proposed commercial playbook; their stated benefits are not independently measured in the conversation.

The hosts ask about platform risk when model companies release products that compete with developers building on them. Smith says Anthropic wants to act as an enterprise provider without that conflict and argues that regulated customers require sustained attention to safety, explainability and regulatory requirements.

He argues that Anthropic’s emphasis on enterprise customers shapes investment in trust, explainability and regulated-sector needs. The hosts frame that focus as a possible way to reduce conflict with businesses building on the company’s models.

For international expansion, Smith expects local commercial and applied AI teams in markets including the UK, France, Germany, India and Japan. He says large customers want specialists on site, but Anthropic plans to scale through partners such as Salesforce and consulting firms instead of building a sales force that grows linearly with revenue.

Local teams matter because customers often ask for applied AI specialists to work with them on site, Smith says. He nevertheless expects consultancies and platform partners to extend Anthropic’s reach beyond what its own staff could cover.

Asked whether regulation could slow startups, Smith says he wants both innovation and safe, trusted AI. He points to constitutional AI and interpretability research as investments that he believes can support adoption in banking and healthcare. These are his claims about the approach; they do not guarantee any model’s outputs or regulatory compliance.

Smith presents constitutional AI and mechanistic interpretability as safety efforts that may make deployments more useful to regulated customers. He argues safety and innovation can reinforce each other, while offering no detailed regulatory framework in this interview.

Figma CTO Kris Rasmussen explains that Figma is a design and product-development platform. Figma Make, he says, uses AI to turn Figma designs and prompts into functional prototypes or web experiences. The product relies on Anthropic’s Claude models while connecting generated work to the broader Figma design environment.

Rasmussen distinguishes a prompt-only workflow from one that lets teams explore and refine designs. A team can generate an initial design, work on it collaboratively with visual tools, then return that context to Figma Make. He sees AI as accelerating an iterative relationship between design and engineering.

The Figma workflow Rasmussen describes can start with a prompt, continue with visual exploration and team alignment, and then return that context to a model. He sees the surrounding design ecosystem as a differentiator from isolated prompt-to-code tools.

He expects more blending of roles: designers can use code as a tool, while engineers and product managers can express ideas visually. He does not claim everyone becomes an expert designer; rather, an individual can make ideas concrete and give a design team better material to develop.

Asked about design careers, Rasmussen says easier coding shifts attention toward choosing the right experience, sharing prototypes and aligning teams. Producing more software does not automatically produce good software. He argues that intention, taste and collaboration become important when teams can create alternatives quickly.

Rasmussen says easier prototyping makes it more important to share alternatives and decide which ones deserve to reach users. Without that selection, rapid generation can create confusing or low-quality experiences.

Rasmussen says design involves more than appearance: designers think about the whole system, users’ mental models and the emotional response to an experience. He believes taste can be learned, even if some people begin with an advantage, and AI may help non-designers communicate their ideas more clearly.

He says design is a learnable skill and that AI can help people who lack formal training visualize an idea. That does not make every user a top designer; it can improve communication with the specialists responsible for the final experience.

The hosts connect that argument to empathy and systems thinking. Rasmussen says those skills were already valuable, but AI can free time to attend to them. He still needs to care about details such as database performance, even if a coding model reduces time spent implementing a fix.

One initiative Rasmussen highlights is Figma’s MCP server. It can deliver design context to coding agents so they can translate a design while considering existing code. In the example he describes, developers provide a Figma URL to GitHub Copilot, which requests design context through the server and works with the existing codebase.

In his MCP example, a development team gives a coding tool the URL of a Figma design. The tool retrieves context through the server and reconciles it with an existing codebase, an example of combining a visual specification with implementation context.

Asked about a future in which a neural network generates the entire software experience, Rasmussen says it is theoretically possible but faces performance and design challenges. In such an extreme, context and taste would matter even more, while people would still need to agree on predictable experiences that preserve an organization’s identity.

Rasmussen cautions that model-generated interfaces would still need consistency, shared expectations and an organization’s identity. He treats fully generated software as a theoretical possibility, not an inevitable or imminent product.

Rasmussen contrasts extreme personalization with practical needs for consistency and user muscle memory. He sees a future design system as a way to curate boundaries within which a model may have freedom. The precise form of such a system remains an open design question.

At Figma, teams use their own design tools and AI coding products including Cursor and Claude Code. Rasmussen says the company is tool-agnostic and experiments with post-training open-weight models. He expects future interfaces to use richer visual and other modalities, not just linear chat.

Figma uses its own canvas and FigJam for briefs and mockups, Rasmussen says, giving designers and engineers a shared space to consider alternatives. The company also explores open models and uses multiple coding tools rather than requiring one preferred model.

Rasmussen says about one-third of Figma’s user base consists of designers, with the rest in other roles. Product managers, marketers and executives also collaborate on product development; he expects AI to help more of them contribute, while preserving a place for professional design.

Nancy Xu, vice president at Salesforce AI, explains that her work on Agentforce spans the agent development lifecycle: building tools, helping customers reorganize workflows, and observing how agents behave in production. She argues that a capable model alone does not make an enterprise agent successful.

Xu lists data connections, tools and a trust layer as essential scaffolding. She describes Agent Script, announced at Dreamforce, as a way to make agent behavior more predictable and controllable. For example, answering a retirement-transfer question is insufficient if the agent does not follow the relevant business process.

Xu says a conversational model can sound helpful while failing to follow the process needed for a business transaction. Agent Script is intended to add control around that behavior; the interview does not establish that scripting alone eliminates errors.

The hosts contrast creative tasks with situations where an agent must choose the correct product or follow a repeatable procedure. Xu cites a customer, 1-800Accountant, and says it achieved a 90% deflection rate on tax queries. That is a company-presented customer metric, not a measure of universal accuracy.

Xu sees evaluation and observability as major opportunities. Customers need to see which production interactions work and which do not, feed those signals back to builders, and coordinate multiple agents. She frames this as a continuous development lifecycle rather than a one-time deployment.

The hosts note that evaluation is often treated as an afterthought. Xu says customers need to measure agents against their own business objectives, while Salesforce tests model changes so upgrades do not silently reduce performance.

Salesforce also benchmarks model updates internally, Xu says, including versions from OpenAI, Anthropic and Google. The purpose is to help customers understand whether changing a model improves or degrades the agent on their use case, rather than assuming a newer model is automatically better.

Xu cites recruiting company Adecco as another partner and says around half of off-hours candidate cases are successfully addressed through its agent. She previously worked at Moonhub on recruiting and other vertical agents. The reported customer metric is her account of the deployment.

Asked about the labor market, Xu acknowledges that AI is changing work quickly, potentially faster than earlier technological shifts, but says she is optimistic. She imagines more workers managing agents, delegating execution and focusing on the strategic reason for a task.

Xu’s agent-manager vision shifts the worker’s attention toward deciding which outcomes matter and delegating steps. She acknowledges that the pace of change creates a serious labor-market question even as she expresses optimism about new roles.

Her longer-term vision is an agent that can pursue an objective over time, such as improving a customer-satisfaction score over six months, with simulation and evaluations helping guide its work. This is a prospective scenario, not a demonstrated result. The hosts ask whether management will become tedious; Xu emphasizes strategic collaboration instead.

On the model landscape, Xu expects a mix of general and specialized models rather than a single winner. Customers already choose among providers, she says, and models trained for healthcare, law or finance may serve particular domains. She wants Agentforce to support multiple models and agents that communicate across platforms.

Xu also discusses communication between agents through A2A, asking whether a coordinated system should be understood as one agent with subagents or several agents working together. For Salesforce, she says, interoperability and orchestration are key design goals.

Accountability depends on the circumstances, Xu says. She distinguishes a possible model-training copyright issue from a platform’s handling of user data, but does not give a legal determination. Her practical recommendation is to build security and trust in from the beginning and involve business leaders alongside developers.

As illustrations of responsibility, Xu points to different layers of an AI system: the foundation model and a platform that sends user information to it. She says responsibility is situation-dependent and emphasizes business leadership’s involvement in trusted deployment.

Adoption can be both top-down and bottom-up. Xu says complex service-center agents often require a coordinated team and leadership commitment, while narrower tools such as document summaries may spread through individual users. The interview closes with thanks, and the hosts wrap up the Dreamforce broadcast.

For simple task-oriented tools, an individual can discover an agent and recommend it to colleagues. For a deeply integrated service-center agent, Xu says implementation requires coordination, leadership and ongoing participation from the people who will use it.