Salesforce is preparing for enterprise software to shift from fixed applications toward systems where AI agents carry out work. Its strategy puts Slack at the point where people and agents collaborate, while keeping business data and application rules underneath. For AI startups, that creates a practical question: does a product help agents do useful work inside an enterprise workflow, or does it merely add another screen?
Slack as a place for agents to work
Salesforce CEO Marc Benioff describes an enterprise AI system as four connected parts: underlying data, applications and analytics, an agent reasoning layer, and a user interface. Slack is central to the last part. A conversation can give employees a place to direct an agent, review its work and continue a task with colleagues.
▲ Slack as a shared agent workspace
Salesforce paid $25 billion to acquire Slack, which generated about $900 million in annual revenue at the time. Benioff says the plan was to make it a conversational hub for business agents, not just a messaging service. Salesforce has since introduced Slackforce for agent workflows, Slackbot integrated with Claude, Slack Code for developer scripting in channels and Slack CRM for managing enterprise activity.
Benioff points to Meta, Nike and General Motors moving teams to Slack for workflows that include agents. He also says OpenAI and Anthropic use Slack for internal engineering and development. These examples support Salesforce’s case that agents need a shared workplace, though they do not mean every company will organize work the same way. Benioff expects software agents soon to outnumber human employees in Slack workspaces; that remains a forecast.
Slackforce is presented as a way to run autonomous agents with zero data retention guarantees. That distinction matters when an agent handles proprietary records: a useful interface still needs clear rules for what data reaches an AI model and what the model retains.
Why Salesforce is backing more than one AI provider
Salesforce’s investment pattern reflects a related bet: enterprise customers may want access to several kinds of AI models rather than depend on one provider. Benioff says Microsoft’s exclusive alliance with OpenAI prevented Salesforce from investing directly in OpenAI. Salesforce Ventures instead backed Anthropic, Mistral AI, Sakana AI and Cohere.
Salesforce also invested in Hugging Face, which serves as an open-source hub for models and related development. Its importance to Salesforce is different from Slack’s: one is part of the model ecosystem, while the other is a place where employees can put agents to work. Salesforce’s acquisition of customer-agent startup Fin adds a product aimed at deploying autonomous customer-service agents.
For a startup, the pattern suggests a way to assess partnerships. A model provider, a repository, a customer-service agent and a workplace interface solve different problems. The strongest fit may be the layer where a startup can offer a specific capability without requiring customers to abandon their existing data and workflows. That is an interpretation of Salesforce’s choices, not a guarantee that any one layer will win.
Data and rules remain part of the product
Agents cannot work reliably with enterprise records unless they can reach the right information under the right permissions. Benioff argues that Salesforce’s metadata-driven architecture helps here. Metadata is the information that describes an application, including its fields, page layouts and sharing rules. Because those definitions are separate from one fixed screen, the application can be presented through different interfaces—an approach often called headless software.
Salesforce has long exposed its applications through technical interfaces and has added Model Context Protocol (MCP) integrations, which connect AI systems with external tools and data. Its Data Cloud is intended to make operational records available to reasoning systems rather than leave those records inaccessible inside separate applications. Salesforce has also deployed ClaudeForce for internal business operations.
The startup lesson is not simply to attach a chatbot to a database. A product needs to account for the definitions and access rules that give enterprise data its meaning. It also needs a clear policy for handling records sent to outside AI models.
Pricing may follow several measures at once
The agent shift also complicates the familiar software subscription. A human seat remains useful as a billing unit, but it may not capture the work performed by agents. Benioff describes six dimensions that Salesforce can combine in customer contracts:
| Pricing dimension | What it measures |
|---|---|
| Human seats | Named employees using software |
| Digital agents | Named agents available to do work |
| Software usage | Activity within the product |
| Resource consumption | Tokens or storage used |
| Completed transactions | Tasks finished through the system |
| Business value | An agreed percentage of value created |
Salesforce projects more than $50 billion in annual revenue for the coming year, and Benioff says a business of that size cannot depend on one AI pricing model. Customers instead need contracts that combine measures to suit their operations. Startups can take a narrower lesson: decide what customers can count, what they can control and which measure best reflects the work the product performs. Salesforce’s framework is a set of options, not a result other companies should expect to reproduce.
The interface may change again
Slack is a conversational surface today, but Benioff expects business interfaces to become more adaptive. Instead of opening a fixed set of icons and windows, an employee might describe a task and receive a workspace assembled around that request. The system would change the information and controls it presents as the work changes.
▲ An adaptive workspace interface
That is a prediction, not a description of a completed transition. It does, however, explain why Salesforce is connecting agents, enterprise records and Slack: the company wants its applications to remain useful even if the screen employees use looks different.
What startups should do next
Start with customer workflows rather than claims that AI will eliminate existing software. Identify the data an agent needs, the permissions it must respect and the place where a person will review or direct its work. Consider more than one model integration where dependence on a single provider would constrain customers. Then choose pricing measures that customers can understand alongside the costs of running the agent.
Benioff also argues for building responsibility into a company from the start. Salesforce’s 1-1-1 model commits 1% each of equity, product and employee time to nonprofit causes. Startups need not copy every part of Salesforce’s strategy to take its central point seriously: agents will be valuable when they fit real work, use enterprise data responsibly and give customers a clear way to judge what they accomplish.