Salesforce just built an AI model that isn't smarter than GPT or Claude, and it's beating them anyway. Koa, Salesforce's new CRM-only reasoning model, delivers 3x fewer errors than a general frontier model on routing and updating CRM tasks. The same week, a startup called TypeSafe launched a model that can't hallucinate because it only ever picks from a fixed list of answers. And OpenAI itself quietly conceded the "one chatbot for everything" era is over by launching a version of ChatGPT built specifically for financial services. If you're running a business on a single general-purpose AI subscription, the smartest companies in the industry are telling you that's the wrong architecture.
Salesforce's Koa: 3x Fewer Errors by Narrowing the Focus
Salesforce built Koa by post-training Nvidia's Nemotron 3 Super on synthetic data modeled on 27 years of its own CRM deployments, rather than licensing a frontier model from OpenAI, Anthropic, or Google. On real-world tasks like updating an opportunity, routing a case, or scheduling a follow-up, Salesforce says Koa matches or exceeds leading general models with 3x fewer errors, 11% more precision at calling the right action, and 2.1x greater reliability. Koa is already running inside Salesforce and moving into customer pilots this October with Formula 1, Baxter Credit Union, 1-800Accountant, Engine, and Xero, with general availability expected this winter.
Koa beats general models because it was trained on nothing but the exact task it needs to do, not because it's a smarter model overall. That's the pattern worth copying: a narrowly trained tool beats a generalist whenever the task itself is narrow and repetitive.
TypeSafe's Jev: The AI That Can't Hallucinate
TypeSafe, the startup from former OpenAI researcher Diogo Almeida, launched Jev the same week. Jev is what TypeSafe calls a "System One" model: instead of generating open-ended text, it only ever returns a value from a schema you define, whether that's a yes/no choice, a score, or a pick from a fixed list. Because it's mathematically impossible for Jev to return something outside that schema, TypeSafe says it can't hallucinate.
That claim comes with a real nuance worth knowing before you build on it: Jev can't produce an invalid answer, but it can still pick the wrong one inside the schema you gave it. "Hallucination-free" here means something narrower than "always correct." But for tasks like scoring a lead, routing a support ticket, or screening another AI's output for a jailbreak, that narrower guarantee is exactly what you need, and it's a lot cheaper and faster than a full chat model.
The Digital Door Paradox: When General Models Win
Not every case favors the narrow model, and the counterexample matters. A Nature Medicine study led out of NYU Langone pitted two specialized clinical AI tools, OpenEvidence and UpToDate Expert AI, against three general frontier models: GPT-5.2, Gemini 3.1 Pro, and Claude Opus 4.6. Using real, de-identified questions doctors were actually asking on hospital floors, the general models won across all three evaluation stages.
The reason had nothing to do with medical knowledge. The specialized tools were accessed through restrictive, consumer-facing web forms with no control over the model's behavior. The general models were tested through a clean, direct API where researchers could set the temperature to zero, forcing strictly factual output and eliminating creative hallucination. The interface, not the training data, decided the outcome. This study is contested, though: OpenEvidence, one of the tools that lost, has publicly asked Nature Medicine to retract it. Treat the finding as directionally useful, not settled science.
The practical rule this leaves you with: a model's training doesn't matter if the access layer around it is broken. Before you assume a "specialized" tool will outperform a general one, check whether it's actually giving you clean, parameter-controlled access, or just a locked-down form.
Build a Specialized, Orchestrated Team, Not One Chatbot
Put together, these three stories point at the same shift: stop routing every task through one general-purpose chatbot subscription. Use a mid-tier general model for the 80% of daily work that's genuinely open-ended. Reserve your best model for complex planning. And for the narrow, repetitive, high-volume stuff, tagging, routing, scoring, screening, build or buy a small tool that only does that one job, connected to everything else through an API, not a locked web form.
FAQ
Is Salesforce's Koa model available to the public yet?
Not yet. Koa is running internally at Salesforce and moving into customer pilots (Formula 1, Baxter Credit Union, 1-800Accountant, Engine, and Xero) starting this October, with general availability expected in U.S. regions this winter.
Does TypeSafe's Jev model really never hallucinate?
It can't produce an answer outside the fixed schema you define for it, which makes an invalid or fabricated response mathematically impossible. It can still choose the wrong option within that schema, so "can't hallucinate" doesn't mean "can't be wrong."
Why did general AI models beat specialized clinical AI tools in the Nature Medicine study?
Not because they knew more medicine. The general models (GPT-5.2, Gemini 3.1 Pro, Claude Opus 4.6) were tested through a clean API with the temperature set to zero and web search enabled, while the specialized tools were only accessible through restrictive consumer web forms with no parameter control.
Do I need to build a custom AI model for my business?
Usually no. The pattern across these stories favors wrapping existing general models with tight API controls (zero temperature, structured outputs, verified data) over training something from scratch, unless you have Salesforce-scale proprietary data and a genuinely narrow, high-volume task to justify it.
Companion episode:
Want more breakdowns like this? Check out joebuildsai.com for daily AI system-building content.