Salesforce is rolling out three major AI research initiatives to accelerate enterprise AI deployment: CRMArena-Pro for realistic enterprise tasks, an Agentic Benchmark to evaluate AI agents in CRM, and Account Matching in Data Cloud to clean and deduplicate data.

Salesforce frames these efforts as part of pursuing Enterprise General Intelligence, aiming for capable, reliable agents across a range of tasks, with results to be showcased at Dreamforce in October.

The company stresses training and testing AI agents in simulated scenarios to close the gap between demos and real-world performance, noting MIT findings that 95% of AI pilots don’t reach production and Salesforce’s own 35% success rate in complex scenarios.

Security concerns are front and center after a breach where OAuth tokens were exploited to access Salesforce instances, leading to the removal of Salesloft Drift from AppExchange pending investigation.

CRMArena-Pro uses synthetic yet domain-validated data to test AI agents in realistic tasks—like customer service escalations, sales forecasting, and supply chain disruptions—running inside Salesforce production environments.

The Agentic Benchmark measures AI agents on five enterprise metrics—accuracy, cost, speed, trust and safety, and environmental sustainability—with an emphasis on sizing models to task complexity to reduce environmental impact.

Account Matching uses fine-tuned language models to consolidate duplicate records across systems, delivering high match rates (a client hit 95%) and cutting time spent on manual cross-checks.