# Salesforce Unveils Koa Reasoning Model and Air India AI Deal at Dreamforce 2026 > Source: > Published: 2026-09-16 13:07:05+00:00 **September 16, 2026, (Inside AI) —** Salesforce opened its annual Dreamforce conference in San Francisco on Tuesday with a clear message: the company that built its empire on customer relationship management software now wants to own the reasoning layer that powers it. The centerpiece is **Koa**, Salesforce's first AI reasoning model, built on top of **Nvidia**'s open-weight **Nemotron** family of large language models and post-trained specifically for sales, marketing, and customer support tasks. Koa represents a strategic pivot. Rather than letting enterprises route sensitive CRM data through third-party models like **Claude** or **ChatGPT**, Salesforce is offering a task-specific alternative that runs within its own security perimeter. The company says both Nvidia and Salesforce ensured no actual customer data was used to fine-tune Koa. Instead, the model was trained on synthetically generated data that mimics customer patterns, a decision that addresses growing enterprise anxiety about data leakage into foundation models. The timing matters. AI agents are becoming more autonomous, and that shift is forcing software-as-a-service giants to rethink how customers interact with their products and what they are willing to pay for. Reports point to revised contract terms with smaller footprints as companies experiment with agent-driven workflows. Salesforce is betting that owning the reasoning model, not just the interface, will keep customers anchored to its ecosystem. ## Token Efficiency Becomes The New Battlefield Salesforce claims Koa's inference architecture is designed to be token efficient, meaning it uses fewer tokens to complete work tasks that customers want an AI agent to handle. That efficiency could translate into lower costs compared to routing multi-step tasks to Claude or ChatGPT through the **Agentforce** AI gateway. In enterprise AI, where token consumption directly drives operational expenses, this is not a minor technical detail. It is a competitive weapon. **Read:** **OpenAI Launches GPT-6 Astra for Complex Business Work** The company also launched **AIforce**, a tool that lets customers carry enterprise data, workflows, permissions, and security into any AI interface without opening the Salesforce app or navigating its user interface. Users can generate custom live interfaces on the fly within **Slack**, Claude, and Coworker. Under the hood, AIforce runs on Salesforce's Headless Toolkit, an open architecture that exposes every element of the platform and gives developers MCPs, APIs, plug-ins, skills, and tools to build customized AI experiences. Salesforce says AIforce was designed with security, governance, and zero data retention as foundational principles. The company also upgraded **Claudeforce**, its existing offering with **Anthropic**. The tool now includes over **37** pre-built sales skills that let professionals automate pipeline updates, access revenue context, and take governed action. It is currently in open beta. Salesforce also announced deeper collaborations with **Google Cloud** and **Amazon Web Services**. Perhaps the most concrete deployment announced at Dreamforce involves **Air India**. The Tata-owned carrier will use Agentforce-powered AI agents to process refund requests in under **four hours** and resolve baggage and loyalty-related claims. It will also manage eligibility checks and reissue tickets with requested name changes in under **30 minutes**. The transformation is part of Air India's broader **Vihaan AI** project. Salesforce CEO **Marc Benioff**, speaking to CNBC, framed the company's approach around responsibility. **"We want AI to take care of these things and be held responsible, and take care of these actions and be ethical. The heart of ethics is responsibility."** (Marc Benioff, CEO, Salesforce) He added: **"We really look at a light on our brethren here in San Francisco and all over the world and say, 'We want you to be ethical in your actions and how you take care of your communities and the world,' because this technology is very powerful."** (Marc Benioff, CEO, Salesforce) Benioff also highlighted India's role in the company's AI-first rebuild. **"We've built incredible teams in India. I think this is a huge opportunity for India. That is, we're going to rebuild all of our enterprise systems to be AI-first, and I think these intelligent interfaces-they're going to do some of the work, and then they're going to need partners, trailblazers, and we have millions and millions of trailblazers in India. So I'm sure that they will be a key part of this incredible story."** (Marc Benioff, CEO, Salesforce) The announcements arrive amid heightened debate about AI safety. Anthropic CEO **Dario Amodei** recently [called for an intentional slowdown](https://insideai.news/news/ai-in-business/ai-slowdown-tech-stocks/11843/) of frontier AI progress to let safety measures catch up. Salesforce's decision to build Koa on an open-weight model and train it on synthetic data could be read as a direct response to that concern, though it also serves a commercial purpose: keeping customer data inside Salesforce's walls. Salesforce has not disclosed pricing for Koa or AIforce. The company also has not said when Koa will be generally available beyond the Dreamforce demonstrations. Air India's deployment timeline remains unclear, though the airline has positioned the move as central to its operational overhaul. For now, Dreamforce 2026 signals that the race to own enterprise AI is no longer just about building the best agent. It is about controlling the reasoning layer, the data perimeter, and the token economics that determine whether AI adoption actually saves money or quietly inflates cloud bills.