BENGALURU: Salesforce is taking a fresh look at its artificial intelligence approach. The company wants to change how large language models work within business software. Many organizations face a common problem. They cannot move generative AI from test projects to dependable, fully operational systems.
The Gap Between Benchmarks and Reality
Srini Tallapragada serves as president and chief engineering officer at Salesforce. He recently shared his observations with TOI. The past two years have shown something important. There is a growing difference between how LLMs perform on standard tests and how they actually behave in real business environments.
"Large language models represent foundational technology," Tallapragada stated. "They will remain relevant for many years to come. However, enterprises are learning an important lesson. Strong benchmark performance does not automatically lead to consistent business results."
The Pilot Project Problem
According to Tallapragada, most large companies spent 2024 and early 2025 running AI demonstrations and pilot programs. They discovered a frustrating reality. Very few of these systems could advance into full production use.
The main challenge lies in what experts call the "last mile." AI systems must operate predictably across unusual situations. They need to maintain performance over extended periods. They must also function under strict regulatory scrutiny.
Probabilistic Systems Meet Business Needs
Large language models are probabilistic by design. They excel at understanding intent, language, and context. Yet they do not always follow fixed instructions with absolute certainty.
"These models may comply ninety-seven percent of the time," Tallapragada explained. "But enterprises require workflows that function one hundred percent of the time." This requirement becomes especially critical in specific areas. Financial services demand perfect accuracy. Customer refunds require precise handling. Policy enforcement needs consistent application.
Combining Approaches for Better Results
Salesforce is addressing this challenge through a combined approach. The company is pairing generative AI with deterministic systems. These systems enforce non-negotiable rules and standard operating procedures.
In practical terms, this means using LLMs where flexibility, reasoning, and empathy are necessary. Meanwhile, the company relies on rule-based logic for compliance-heavy steps. Audit-sensitive processes also benefit from this traditional approach.
"People initially tried to use the same tool for everything," Tallapragada noted. "But sometimes a simple 'if-then' rule provides the right answer. The real challenge involves making these different approaches work together seamlessly."
Moving Beyond Industry Benchmarks
Tallapragada also offered a caution about industry benchmarks. He warned against over-reliance on these measurements. Many tests remain theoretical in nature. Some can even be manipulated to show better results.
"A perfect score does not guarantee reliable real-world performance," he emphasized. "Businesses need systems that work consistently in actual operating conditions."
Continued Commitment to LLMs
Despite this more disciplined approach, Salesforce is not reducing its use of large language models. The company works with multiple models, both large and small. Overall usage continues to increase across the organization.
Salesforce optimizes for three key factors: performance, cost, and sustainability. The company wants to balance technological capability with practical business considerations.
The Turning Point for Enterprise AI
Looking ahead, Tallapragada identified 2026 as a potential turning point. Enterprise AI adoption could reach a new stage of maturity during this period.
"The focus is shifting from excitement to outcomes," he observed. "Our job involves turning powerful models into systems that deliver real business value. They must perform consistently. They must operate at scale."
Salesforce CEO Marc Benioff has previously outlined the company's AI philosophy. The strategy aims to augment human decision-making rather than replace it. AI agents handle routine tasks efficiently. Humans retain judgment-driven roles that require deeper understanding.
This balanced approach reflects Salesforce's evolving perspective on artificial intelligence. The company recognizes both the potential and the limitations of current technology. By combining different systems thoughtfully, Salesforce hopes to help businesses achieve more reliable AI implementation.
