By Srikanth Kamaraju, EVP & India Practice Leader, Bounteous
Every Dreamforce brings marquee announcements and customer stories. But from the perspective of someone working with Salesforce implementations, the more interesting story this year sat one level below the headlines: what it now takes to move AI agents from compelling demonstrations into production.
The strongest evidence came from three implementation archetypes – customer service, software engineering, and finance and operations automation. These are areas where agents are beginning to move beyond pilots and demonstrate measurable returns.
Consider three examples shared at Dreamforce. A global industrial automation group with nearly 120,000 suppliers reduced supplier onboarding from days to hours and deployed inbound pipeline agents in 45 days or less. A global staffing firm processing roughly five million applications each month reported 2.6 million agent interactions this year, while recruiters recovered 35–40% of their time and delivered 20,000 additional placements. Salesforce reported five million service conversations resolved by one agent and $500 million in pipeline generated by another in a single quarter.
What connects these examples is more significant than the numbers. None represents a fundamental model breakthrough. They are integration and workflow wins.
The interface is becoming the architecture
Salesforce framed its platform in four layers: data, applications and semantics, agents, and now interface. AIforce extends Salesforce into AI surfaces, while Claudeforce brings Salesforce natively into Claude. Agentforce Coworker applies the agent layer to the familiar Lightning interface, with Headless 360 providing the underlying foundation for experiences across Claude, Slack, Lightning, and other surfaces.
That shift matters. The interface is becoming less tightly coupled to the underlying system. Instead of rebuilding applications for every new experience, enterprises can increasingly recompose existing capabilities around the context in which work happens.
Model choice is changing, too. Koa, Salesforce’s CRM reasoning model for long-running agents, is built on NVIDIA’s Nemotron. More broadly, enterprises can combine foundation models from different providers on a common orchestration layer.
For executives, the implication is straightforward: model selection is becoming an architecture and economics decision, rather than a long-term bet on one provider.
The hard part starts after the demo
For practitioners, some of the less prominent announcements may prove more consequential.
Tools for specification generation, prompt grounding, validation, adaptive interfaces, multi-agent governance, security, privacy, and agent identity address work that engineering teams have often had to assemble themselves.
That is important because this is where many AI projects encounter reality. One operations agent discussed at Dreamforce stalled when it needed to write into SAP. The eventual solution combined a sandbox, learned business rules, and a library of trusted actions to turn agent learning into deterministic execution.
This is the unglamorous middle of enterprise AI: connecting agents reliably to existing systems, establishing permissions, encoding business rules, and governing what agents can do.The quality of that implementation increasingly separates a successful demonstration from a production system.
Where AI value gets real
Dreamforce also showed how agents are moving from generic assistants toward specific roles across sales, service, commerce, supply chain, HR, and IT.
But enterprises don’t need to automate everything at once. A more useful starting point is two or three journeys where the workflow is understood, the integration path is achievable, and the outcome can be measured independently. Verified results can then fund and inform the next phase.
The most interesting question after Dreamforce is therefore no longer simply whether AI models are capable enough. It is whether an enterprise can make them work reliably inside the systems, processes, permissions, and human decisions that already run the business.
That is where the next phase of differentiation will happen – not in the demo, but in the distance between capability and deployment.