{
  "id": 12522653,
  "title": "The Case For Evals In A Multi-Model World",
  "url": "https://urgent.news/2026/10/07/the-case-for-evals-in-a-multi-model-world",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-07T01:30:07.000Z",
  "source": {
    "name": "Inc42",
    "slug": "inc42",
    "url": "https://inc42.com/features/the-case-for-evals-in-a-multi-model-world/"
  },
  "original_language": "en",
  "account": "Evaluations, or evals, are becoming essential in the development of AI models and systems, according to leaders from major tech companies like Swiggy, Rapido, Razorpay, and Shadowfax. The traditional approach of selecting a single model for use is shifting, as these companies now prioritize creating evaluations and monitoring into the model lifecycle. Evaluation helps determine if a change to a model is safe and beneficial for the product. Instead of testing models in isolation, companies assess changes against their specific tasks and the customer experience. For instance, Swiggy's contact-center agent was tied to one model; when the provider faced capacity issues, migrating to a new model took nearly a month. To streamline this process, Swiggy now uses evaluations and experimentation to make model changes easier. The key is to measure what matters for each product and workflow, such as quality, latency, cost, and customer impact. Monitoring is also crucial, as it must show what is happening in production, identify meaningful regressions, and connect these to decisions about keeping, adjusting, or rolling back a model. Continuous evaluation and monitoring, along with automation, can help companies make model changes quickly and minimize the risk to users.",
  "summary": "Another day and another new model grabs the limelight. Product teams scramble to see how it can be used for…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}