{"id":13027,"date":"2025-12-27T16:15:27","date_gmt":"2025-12-28T00:15:27","guid":{"rendered":"https:\/\/www.salesforce.com\/?p=13027"},"modified":"2026-02-03T17:02:36","modified_gmt":"2026-02-03T06:02:36","slug":"ai-learned-to-play-by-rules","status":"publish","type":"post","link":"https:\/\/www.salesforce.com\/au\/news\/stories\/ai-learned-to-play-by-rules\/","title":{"rendered":"In 2025, AI Grew Up \u2014 and Learned to Play by the Rules"},"content":{"rendered":"\n

As 2025 comes to a close, one thing is clear to us at Salesforce: enterprise AI<\/a> has entered a new phase. Over the past three years, large language models (LLMs) proved what was possible. They reasoned, summarised, and conversed in ways that reshaped expectations for software almost overnight. But as those capabilities became familiar, our conversations with customers began to change.<\/p>\n\n\n\n

Businesses stopped asking what AI could do next and started asking whether it was dependable and reliable enough to move beyond flashy demos and truly power their businesses. Customers wanted AI agents<\/a> they could trust inside their most critical workflows: agents that could operate within policies, respect regulations, and execute correctly every time. <\/p>\n\n\n\n

That last mile \u2014 between impressive demo and deployment at scale \u2014 is where most AI systems fail. Yet it\u2019s where enterprises can unlock the greatest business value. <\/p>\n\n\n\n

LLMs alone cannot close that last mile. Language models are exceptional at understanding intent and context but they are, by design, probabilistic. They generate likely outcomes, not guaranteed ones. In low-stakes scenarios, that\u2019s acceptable, even desirable. In core business operations, where policies, regulations, and customer trust are non-negotiable, it\u2019s a liability. In these environments, hope is not a strategy. You can\u2019t cross your fingers that your AI behaves correctly. You need to know it will, every single time. This requires more than intelligence; it requires the right architecture to ensure that AI plays by the rules.<\/p>\n\n\n\n

Fortunately, we\u2019ve spent the last several years focused on just this problem. This work wasn\u2019t always as attention-grabbing as the latest LLM model, and it wasn\u2019t always glamorous. But in 2025, it yielded real results. Today, more than 18,500 organizations have begun their transformations into Agentic Enterprises<\/a> with Salesforce. Agentforce 360 has processed over 3.2 trillion tokens, with usage surging 25% in October alone. And global leaders like Falabella<\/a>, Finnair<\/a>, Goodyear<\/a>, and Williams-Sonoma, Inc<\/a>. are deploying AI agents into real workflows to drive outcomes that matter: higher sales productivity, faster resolutions, and conversion rates that impact the bottom line.<\/p>\n\n\n\n

These companies aren\u2019t just turning LLMs loose on their business. They are marrying the flexibility of LLMs with the precision and reliability of deterministic frameworks. This hybrid approach allows AI agents to be flexible where flexibility adds value, and precise where predictability is required. Incorporating this determinism isn\u2019t a step back for AI, it\u2019s a step forward. It\u2019s the moment that intelligence becomes operational.<\/p>\n\n\n\n

Every business contains some functions that require judgement and interpretation, and others that require strict compliance with established rules. Most functions require both. Businesses cannot hallucinate a financial transaction. They cannot approximate a medical policy. When systems behave unpredictably, trust collapses\u2014and without trust, AI agents never move beyond pilots.<\/p>\n\n\n\n