Salesforce AI Research identifies trends shaping agentic AI

Salesforce AI Research today showcased the three major agentic AI trends it predicts will shape AI in the enterprise through 2027.

It also announced AI Foundry, a new initiative shaped by those trends and geared toward helping enterprises move from model-level AI to system-level AI that focuses on how AI components work together across teams, workflows, and organizational boundaries.

“As we push the envelope of agentic AI, we need to ensure that not only are we making progress in terms of capabilities, but also that we’re moving along the axis of consistency, accuracy, and trust,” said Silvio Savarese, EVP and chief scientist of Salesforce Research, in a press conference Wednesday. “This is what we call enterprise general intelligence, which is our true north star for 2026 and 2027.”

Savarese identified three major trends that will shape how Salesforce approaches AI for the enterprise in the coming year: simulation environments, agent-to-agent ecosystems, and ambient intelligence.

Simulation environments enable agents to learn from experience

Late last year, Salesforce unveiled eVerse, a simulation environment to train voice and text agents for the enterprise. It leverages synthetic data generation, stress testing, and reinforcement learning to optimize agents. Savarese said simulation environments like eVerse will be the key to continuous learning for AI agents.

“Even if models and agent systems are increasing in size and capabilities, even if these models have been fed by billions of tokens, their performance is fundamentally saturating,” he said.

Salesforce Research calls this the saturation law, or the scaling law. Agents frequently struggle with tasks that require multiple steps, long horizontal reasoning, and when they encounter corner cases not well represented in training. The answer, Savarese said, is learning from experience, which means continuously absorbing positive feedback from positive behavior while penalizing mistakes.

“In the context of enterprise AI, we can do this by true simulation environments,” he said. “Simulation environments are critical for creating thousands of realistic business scenarios that can be populated with synthetic data, which mimics real customer data or real business logic.”

Within those environments, Salesforce can measure how agents handle complex business cases and reward positive outcomes and penalize mistakes.

Agent-to-agent ecosystems cross organizational boundaries

The second trend is agent-to-agent ecosystems, where AI agents interoperate not only within the same organizations but across organizations.

“We soon will have personal agents, and local agents will be interacting with business agents,” Savarese said. “We’ve already seen the deployment of protocols of communication such as A2A or MCP, and we’re going to need to go a step further by not only looking at how two agents will interact at the protocol level, but also at the semantic level.”

That means establishing the rules of conduct for agents to interact with each other in a safe and productive manner, and ensuring these negotiations between agents are kept within legal and safe boundaries. As part of AI Foundry, Salesforce is investing in an enterprise multi-agent semantic layer with standardized protocols, guardrails, decision logging, and coordinated escalation. The research team is also working with Salesforce’s legal counsel and its Office of Ethical Use of Technology to define the legal frameworks for autonomous agent negotiation.

Ambient intelligence surfaces insights just in time

The third trend is ambient intelligence: context-aware, proactive, and timely agents that can disappear into the background, anticipate needs, and surface insights just as they’re needed.

Itai Asseo, Salesforce AI Research’s head of incubation and brand strategy, points to the company’s newly redesigned Slackbot as an example of ambient intelligence.

“It leverages powerful LLMs, but it does that in the context of the business, and in the context of enterprise knowledge and all the different systems we have access to,” he said.

In addition, Salesforce AI Research has been working on a new ambient intelligence project called Proactive in-Meeting Support Agent (PISA), a sales assistant with access to CRM data that can sit in on sales meetings and surface insights as needed by the sales team.

Also in development are capabilities that embed ambient intelligence directly into enterprise workflows, with a focus on human-AI interaction patterns. The goal is to help humans by sifting through overwhelming volumes of data to surface exactly the information needed in real-time.

AI Foundry initiative brings the trends together

The research team’s AI Foundry initiative brings together AI research, strategic customers, and academic partners to develop, test, and validate new AI capabilities and move them from foundational research to product innovation.

Plus, AI Foundry seeks to help Salesforce transform its product roadmap through the simulation environment, agent-to-agent ecosystem, and ambient intelligence trends.