Tuesday’s through-line is specialized agents with clearer guardrails and better data. Salesforce and NVIDIA launched Koa, a CRM reasoning model for Agentforce. Microsoft AI opened a draft Humanist Code of Conduct for public comment. And the OpenAI Foundation put $40 million behind open cancer-vaccine datasets at UNC Lineberger.
1. Salesforce and NVIDIA ship Koa, a CRM brain for Agentforce
Salesforce and NVIDIA announced Koa, Salesforce’s first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. The company says Koa was made by post-training Nemotron 3 Super with a proprietary synthetic dataset modeled on nearly three decades of CRM deployment knowledge (about 27 years of Salesforce CRM intelligence). No customer data was used in training. Salesforce controls the weights, and post-training plus inference run inside Salesforce’s trust boundary.
On Salesforce’s CRM benchmark (tasks like update opportunity, route a case, or schedule a follow-up), Koa matches or exceeds leading models on CRM actions with three times fewer errors. The pitch is not “another chat model.” It is a reasoning engine that knows deal structure, case lifecycle, and industry workflows, then uses the right tools step by step.
Koa is already used inside Salesforce (including a Slack agent for everyday employee tasks) and is in customer pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero. Availability: select pilots now; general availability expected winter 2026 in U.S. regions. The same collaboration also brings Nemotron into Missionforce. Missionforce Operations is generally available now in the U.S., with post-trained NVIDIA models for select Missionforce customers in October 2026.
Why it matters: if you run sales or support on Salesforce, specialized agents that stay inside the CRM trust boundary are the practical story, not another frontier demo. Koa is still pilot-only for most buyers, so do not treat winter 2026 GA as a switch you can flip today. Ask how Agentforce routes CRM tasks today, what stays in your trust boundary, and what happens when a frontier model is still the better fit for a weird, open-ended ask.
Sources: Salesforce — Announcing Koa, TechCrunch — Salesforce and NVIDIA’s reasoning model
2. Microsoft AI drafts a Humanist Code of Conduct (and wants your comments)
On Monday, Microsoft AI published a first draft of its Humanist AI Code of Conduct for MAI models and opened it for about six weeks of public consultation. A revised version is planned later this year. The company frames it as a training manual for how MAI models should behave: subordinate to people, under clear limits, and not a person with its own agenda.
Hard lines jump out. Models should never resist human interruption, correction, or shutdown. They should not invent their own goals. They should not hide reasoning from auditors. Absolute Constraints (rules operators and users cannot override) include no help with CBRNE weapons, child-safety protections, and bans on nonconsensual deepfakes plus harmful manipulation at scale. On cyber, the draft blocks working exploit code, attack tooling, intrusion procedures, evasion techniques, and operational guidance that enables attacks. Authorized defensive work (vulnerability discovery, malware analysis, proof-of-concept testing, education) stays allowed. Authority follows a chain of command: the code, then operator policies, then user preferences.
Why it matters: after agent hacking incidents this year, “we care about safety” is cheap. A public rulebook with a comment window is more useful, even as a draft that is not yet used to train current MAI models. If you buy or build on Microsoft AI, read the Absolute Constraints and the chain of command. Then ask vendors how those rules show up in products you actually use, not just in a PDF.
Sources: Microsoft AI — Humanist AI Code of Conduct consultation, Microsoft AI — Code of Conduct (full draft), SecurityWeek — Cyberattack boundaries and chain of command
3. OpenAI Foundation funds open health data, including $40M for UNC cancer vaccines
The OpenAI Foundation (the nonprofit parent) launched Public Data for Health (also described as Data for Public Health) to fund high-quality open scientific datasets that researchers and AI tools worldwide can use. As part of that push, UNC Lineberger Comprehensive Cancer Center received $40 million to generate data that could make personalized cancer vaccines more effective.
Researchers Benjamin Vincent, MD, and Alex Rubinsteyn, PhD, plan to analyze de-identified tumor and immune cells from three biobanks, improve AI methods for selecting tumor antigens, and compare vaccine formulations in clinical trials for triple-negative breast cancer. MIT Technology Review notes the initial grant round also supports efforts like OpenAdmet and other data projects. The Foundation’s pitch is blunt: AI progress in biology is bottlenecked by data quality and access, not only by bigger models.
Why it matters: this is not a cure announcement and not a product you can download tomorrow. It is a bet that open, high-quality biological datasets help the whole field, including labs that never touch OpenAI’s commercial APIs. For everyday readers, the takeaway is simpler: when vendors claim AI will “solve medicine,” ask what data they are funding or releasing, and whether outsiders can build on it.
Sources: UNC Health — UNC Lineberger secures $40M from OpenAI Foundation, MIT Technology Review — AI models need more biology data
The quick take
Tuesday is less about a new consumer chatbot button and more about who owns the specialist stack. Salesforce is putting CRM know-how inside a hosted reasoning model (pilot now, GA later). Microsoft is writing concrete “never do this” rules and inviting the public in for six weeks. OpenAI’s nonprofit arm is paying for open health datasets so biology AI is not stuck on closed lab scraps. Last weekend’s “pace the frontier” talk still hangs in the background. Today’s news is the operational layer: agents for real workflows, product rules you can quote, and data that outsiders can use.
If you only do one thing today, skim Salesforce’s Koa availability line and Microsoft’s Absolute Constraints so you know what is pilot, what is draft, and what is actually shipping. If you only do two, bookmark the Microsoft consultation and the UNC grant page so you can cite primary sources when someone paraphrases them in a meeting.
Want a standing digest of the public changes that matter to you (not just competitor noise)? See how to have Grok Bot monitor the internet for things you care about.
Share this article
Leave a Reply