The news
Salesforce (CRM) unveiled Koa, a CRM reasoning model built on Nvidia (NVDA) Nemotron, on September 15, six days after legal AI firm Harvey raised $550 million touting its own model. Domain-specific AI models, tuned for one industry or task, are becoming a vendor selling point.
Koa is post-trained on synthetic data modeled on enterprise knowledge from Salesforce's CRM deployments, using Nvidia's Nemotron 3 Super as its base. It is in pilot with select customers, with general availability planned for winter 2026 in U.S. regions. Salesforce said that on its own CRM benchmark, Koa matched or beat leading models while making three times fewer errors. The company did not disclose pricing.
TechCrunch reported that Koa is meant to use fewer tokens, the units AI models are billed by, than frontier models such as Claude or ChatGPT, and that it can be routed automatically through Salesforce's AI gateway. Because it was trained on simulated customer-service and sales scenarios rather than customer records, it contains no customer data, TechCrunch reported. Salesforce AI executive Jayesh Govindarajan told TechCrunch the company had always depended on frontier labs for reasoning until now.
Harvey made a similar pitch on September 9. The legal AI company raised $550 million at a $15.5 billion valuation in a round led by Diffusion and Lightspeed Venture Partners, and highlighted Tenet, which Harvey describes as an open-weight model post-trained for long-horizon legal work, alongside a legal agent benchmark. Harvey says 80% of Am Law 100 firms and five Fortune 10 in-house legal teams use its software.
AI coding company Cognition is training its own model based on open-source alternatives, TechCrunch reported on September 8 in a story on Cognition's $2 billion raise at a $48 billion valuation. In TechCrunch's assessment, relying less on costly OpenAI and Anthropic models will help Cognition cut costs. Citing The Information, TechCrunch also said Cognition leases an Nvidia server cluster costing hundreds of millions of dollars a year, which could push its total cash burn to $800 million in 2026, a reminder of how much compute these businesses consume.
Vendors are not abandoning the labs. At Dreamforce, Salesforce expanded its Google Cloud deal, with Gemini Enterprise now generally available inside its Reasoning Engine, Salesforce Ben reported. The outlet also listed Bedrock access through Amazon Web Services and Claudeforce, which puts Salesforce's CRM inside Anthropic's Claude.
The numbers
- Harvey funding / valuation
- $550 million / $15.5 billion
- Cognition funding / valuation
- $2 billion / $48 billion
- Cognition annualized revenue (reported)
- $900 million
- Koa error rate vs leading models (Salesforce benchmark)
- Three times fewer errors
- Am Law 100 firms using Harvey
- 80%
Why CEOs should care
CIOs should ask every vendor a simple question: which model handles which task? Many AI features will be routed between a vendor's own model and a frontier model based on cost and complexity. Buyers need to see that routing policy, have a say in it for sensitive workflows, and get notice when the underlying model changes, since a model swap can change outputs. Treat vendor benchmark claims as a starting point and test on your own records before scaling.
CFOs should follow the token savings. If a vendor's own model is cheaper to run, the vendor's margin improves. Ask whether those savings show up in lower usage rates or larger credit allotments, or whether the vendor keeps them. Watch for premium tiers that charge extra for frontier models while defaulting to the vendor's cheaper one.
CISOs and legal teams should examine provenance. Salesforce says Koa was trained on synthetic data, not customer records; ask other vendors for the same clarity about training data, retention and whether your data will be used to improve their models. Check regional availability too, since Koa launches in U.S. regions first, which matters for companies with data residency rules.
The bigger picture
Open-weight base models such as Nemotron have made it practical for application companies to build their own reasoning models without training from scratch. For vendors, owning the model offers lower costs, more control and a story about differentiation as AI labs push into applications themselves. For buyers, it adds a new layer of vendor lock-in: workflows tuned to one vendor's model may not move cleanly to another.
What’s next
Koa's general availability in winter 2026 is the next test, along with any independent benchmarks comparing it with frontier models on real CRM tasks. Watch whether Harvey and Cognition publish evaluations of their own models, and whether other large software vendors announce similar post-trained models at their autumn conferences.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
