It describes bringing an organization's data, AI agents, and work surfaces into one governed system rather than running them separately. In practice this means agents reason over the same current data, hand off to each other within a defined process, and act inside the same permissions and business rules. It matters because most enterprises now run several AI tools at once, each with a partial picture of the customer. Fragmented context produces confident answers built on incomplete information, which is harder to detect than an obvious failure.

Salesforce and Anthropic have partnered to bring Claude into the Salesforce platform, including for regulated industries, and to make Salesforce data and actions available inside Claude. The division of work is practical: Claude handles reasoning and natural language, while Salesforce provides the data, business rules, and governed workflows that turn a recommendation into an action a business can stand behind. Most enterprise use cases need both, because language ability without governed data produces plausible answers rather than correct ones.

Not inherently, and most enterprises will. Problems appear when agents reason over different copies of the same data, when no one owns the handoff between them, and when each operates under its own permissions model. These are integration and governance problems rather than model problems, and they are solved at the orchestration layer, by defining which agent owns which step, what data each can reach, and what happens when a step fails.

Failures cluster at the handoffs rather than inside individual agents. The most common causes are undefined behavior when a step fails, agents working from inconsistent data, business rules that exist informally and cannot be encoded, and approval paths that assume every participant has a system login. Most of these surface only under real volume and real exceptions, which is why a pilot can pass while the rollout stalls.

Yes. Approval requests can be sent by email, text message, or a callout to an external system, with every response recorded against the record in Salesforce. This matters because approvers in finance, legal, and partner organizations frequently sit outside the CRM. When those approvals are handled informally through email or messaging apps, the result is delay, lost context, and an incomplete audit trail.

By defining eligibility, discount, margin, and escalation thresholds as explicit rules the agent calls, rather than judgments the model makes. The agent determines what to ask and in what sequence. The system determines what is permitted. Every step produces a record showing which rule was applied and why, so a decision can be reviewed after the fact. Agents that reason about pricing without calling governed rules are the most common source of commercial risk in agentic sales processes.

Readiness depends less on technology than on whether the business logic behind a process is written down. Processes with clear rules, structured data, and defined approval paths can move now. Processes that depend on judgment, incomplete data, or undocumented exceptions need that work done first. A practical test is to ask whether a new employee could run the process from documentation alone. If not, an agent cannot either, and the first phase of the project is capturing the rules rather than building the agent.