Salesforce AI Foundry: Why System Reliability Now Beats Model Power in Enterprise AI
AI / Machine Learning | 5 min read
Salesforce AI Research has launched AI Foundry — a strategic initiative designed to accelerate the shift from model-level to system-level enterprise AI. Rather than competing on benchmark scores or raw model capability, AI Foundry focuses on the infrastructure, protocols, and validation frameworks required to make AI agents work reliably in production across complex, real-world business environments. The initiative unites Salesforce AI researchers, strategic enterprise customers, and academic partners in rapid co-development cycles — with the explicit goal of moving foundational research into production-grade product innovation faster than the traditional development cycle allows.
"The problems that matter most for businesses don't live at the model level anymore. They live at the system level, where components work together to deliver accuracy, consistency, and reliability at scale. AI Foundry is the engine we've built to make that a reality."
— Silvio Savarese, EVP and Chief Scientist, Salesforce Research
"Many of the old rulebooks simply don't apply anymore. AI Foundry connects foundational research to real business problems by collaborating closely with our strategic customers in rapid iteration cycles."
— Itai Asseo, VP of Salesforce AI Research
The Shift: From Model Wars to System Reliability
For over a decade, AI progress meant bigger, faster, more capable models — and competitive advantage was measured in benchmark scores. Salesforce AI Research contributed to many of those advances, from predictive AI for customer behaviour forecasting to generative AI for developer productivity and agentic AI that acts on behalf of users. But as large language models mature and begin to commoditise, the central challenge for enterprise adoption has shifted. As Savarese framed it: the industry must now progress along two axes simultaneously — expanding AI capabilities while also driving forward on consistency, accuracy, and trust. A model that performs brilliantly in isolation but fails unpredictably in production is not yet an enterprise-grade system. AI Foundry is Salesforce's response to that gap. Its north star concept — Enterprise General Intelligence (EGI) — describes AI systems designed for reliable, consistent performance across the full complexity of business scenarios, not just narrow use cases.
Three Strategic Bets: Where AI Foundry Is Focused
- • Simulation Environments (eVerse) — enterprise AI agents trained on static data fall apart at edge cases and multi-step operations. AI Foundry's simulation platform, eVerse, exposes agents to thousands of realistic business scenarios, using feedback loops to reward correct behaviour and penalise errors before any agent touches production. eVerse has already been used to stress-test Agentforce Voice and to pilot UCSF Health's contact centre billing agents.
- • Ambient Intelligence — context-aware, proactive AI that is embedded directly into enterprise workflows and disappears into the background. Ambient intelligence anticipates needs before they surface and delivers just-in-time insights — always available without creating information overload. Salesforce points to its newly redesigned Slackbot as an early example of ambient intelligence in production.
- • Agent-to-Agent Ecosystems — as AI agents increasingly interact with each other across organisational boundaries, governance becomes critical. AI Foundry is building an enterprise multi-agent semantic layer with standardised protocols (including "agent cards"), guardrails, decision logging, and coordinated escalation. Salesforce's legal counsel and its Office of Ethical Use of Technology are involved in defining the legal frameworks governing autonomous agent negotiation.
Why Enterprise AI Needs the Grid, Not Just the Power Plant
The framing used by Salesforce AI Research is instructive: the large language model is the power plant. Orchestration, governance, and observability are the grid. A power plant without transmission infrastructure is a science experiment. The lesson from 2025's enterprise AI deployments — where agents skipped required steps, produced audit gaps, and failed in ways that looked like success until a customer escalation surfaced the problem days later — is that autonomy alone does not make a system. The more ambitious the AI deployment, the more essential engineering rigour, deterministic rules, and human oversight become. AI Foundry is Salesforce's public commitment to building that grid — compressing the research-to-production cycle, validating in environments that mirror real operations, and establishing measurement standards centred on reliability, latency, auditability, and consistency rather than benchmark leaderboard positions.
Key Takeaways
- • Salesforce AI Research has launched AI Foundry — an initiative to accelerate the shift from model-level to system-level enterprise AI, uniting researchers, strategic customers, and academic partners to move foundational research into production-grade products faster than the traditional development cycle allows.
- • The central thesis: as LLMs commoditise, competitive advantage shifts from raw model power to how intelligently and safely AI components are orchestrated — the "grid" of infrastructure, governance, and observability that makes agents reliable in production.
- • AI Foundry is organised around three strategic bets: Simulation Environments (eVerse — stress-testing agents against thousands of realistic scenarios before production); Ambient Intelligence (context-aware, proactive AI embedded into workflows); and Agent-to-Agent Ecosystems (standardised protocols, guardrails, decision logging, and legal frameworks for cross-organisation agent interactions).
- • Salesforce's north star concept is Enterprise General Intelligence (EGI) — AI systems designed for reliable, consistent, and trustworthy performance across the full complexity of real-world business scenarios, not just narrow use cases or benchmark environments.
- • eVerse has already been validated in real deployments — stress-testing Agentforce Voice and piloting UCSF Health's contact centre billing agents. AI Foundry's success metrics are operational: reliability, latency, auditability, and consistency, not leaderboard scores.