Salesforce AI Research pushes agentic enterprise with simulations

Salesforce AI Research has announced a suite of advancements aimed at helping enterprises transition into what it calls the “agentic enterprise”. Organisations that combine human expertise with AI-driven digital labour.

This quarter, the company introduced three major innovations: an enterprise simulation environment to test AI agents in realistic business conditions, a benchmarking tool to measure agent performance, and advanced data consolidation features in Data Cloud. Together, these developments are designed to give CIOs and IT leaders confidence in deploying AI agents at scale.

Simulating enterprise environments

Building on its earlier CRMArena project, Salesforce unveiled CRMArena-Pro, a simulation framework for training and testing AI agents in complex, multi-turn scenarios such as sales forecasting, case triage, and configure–price–quote processes.

The system uses synthetic data, safe API calls, and strict privacy safeguards to create a realistic test bed without exposing sensitive information. Much like flight simulators prepare pilots for turbulent skies, CRMArena-Pro allows enterprises to evaluate agents against edge cases such as service escalations or supply chain disruptions.

The result: AI agents that are more consistent, resilient, and ready for real-world deployment.

Measuring agent readiness

To address the question of which models perform best in enterprise settings, Salesforce launched the Agentic Benchmark for CRM. Unlike general-purpose evaluations, this benchmark assesses agents across five key enterprise metrics – accuracy, cost, speed, trust and safety, and sustainability.

Sustainability has emerged as a crucial addition, highlighting the environmental footprint of AI systems and the importance of aligning model size with task requirements. By applying these criteria, businesses can match the right models to the right agents and avoid wasteful deployments.

The company also released MCP-Eval and MCP-Universe, complementary benchmarks for evaluating large language models (LLMs). MCP-Eval uses synthetic tasks for scalable testing, while MCP-Universe introduces execution-based real-world challenges to stress-test agent performance.

Consolidating enterprise data

High-quality, unified data underpins reliable AI. Salesforce has enhanced its Data Cloud with Account Matching, an AI-driven capability that identifies and merges duplicate records across business units. By fine-tuning both small and large language models, the system reconciles millions of records in real time.

One early customer reported that Account Matching unified more than one million accounts within a month, achieving a 95 per cent success rate and cutting average handling time by 30 minutes. Sellers can now access authoritative records faster, reduce duplication, and accelerate sales cycles without manual intervention.

The bigger picture

Taken together, these innovations aim to solve two long-standing challenges in enterprise AI: building trust in AI-driven decisions and ensuring agents can operate in unpredictable environments. By offering tools to test, measure, and improve AI agents, while strengthening the data foundation. Salesforce is positioning itself at the centre of enterprise adoption of agentic AI.

Read More:

OpenText Partner Led Cybersecurity Strategy for Indian SMBs

PM Modi's Festive Bonanza - What next-gen GST reforms mean for Indian MSMEs

Backup as a Service in India rises as DPDPA reshapes data protection

Backup as a Service: Decoding India’s opportunity and reality