Business software is entering an era where companies must master model selection, system design, and performance evaluation to extract maximum value from AI—not just deploy the latest models.

What happened

Salesforce acquired Fin (formerly Intercom) for $3.6b, a company that repositioned itself through the AI upheaval using open-source models to maximize price/performance. This acquisition, combined with the US government's shutdown of Fable access and Satya Nadella's published thesis that the moat in AI must be human expertise and the system around the model (not the model itself), signals growing market consensus on what makes AI applications successful.

Why it matters

Building AI applications requires three new disciplines—picking the right models, designing feedback loops that let agentic systems improve, and ongoing evaluation of performance for each company's token budget. Unlike traditional software engineering, these tasks demand active, specialized labor; most companies cannot afford separate teams for each workflow. This complexity means the competitive edge belongs to vendors and companies that master these disciplines, not simply those with the fastest or most advanced models.

What to watch

The tension between model capability and operational cost. Different models have distinct strengths—Kimi K2.6 is fast and creative but less precise; Qwen 3.6 27b is small with strong performance but unreliable in toolchain calls; GLM 5.1 excels at coding but is slower—forcing companies to make tradeoffs based on their specific use case and budget constraints rather than defaulting to state-of-the-art.