Salesforce Koa and AIforce: What Sales Teams Need to Know
Salesforce announced Koa — its first CRM-specific reasoning model — at Dreamforce on September 15, and the part that matters is happening now: customer pilots are running inside Agentforce this October, with US general availability targeted for this winter. Alongside it, Salesforce unveiled AIforce, a new interface layer that pushes your CRM's data, permissions, and workflows out to the tools your reps already live in — Claude, Slack, and the Lightning search bar.
Most Dreamforce coverage treated these as two more logos in a crowded keynote. For sales teams, they're actually one bet: Salesforce is wagering that a smaller model trained exclusively on CRM work will beat frontier models at CRM work, and that the interface to your CRM should follow your reps around instead of making them log in.
Here's what shipped, what it costs, and whether a purpose-built CRM model changes anything for your pipeline.
What Koa Actually Is
Koa is a reasoning model built by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset modeled on 27 years of Salesforce CRM deployments — scenarios spanning 14+ industries including manufacturing, financial services, healthcare, and travel. Salesforce says no customer data was used in training; the corpus is synthetic workflows (updating opportunities, routing cases, scheduling follow-ups) generated from institutional knowledge, refined with supervised fine-tuning and GRPO reinforcement learning on NVIDIA's NeMo stack.
The headline claim: on Salesforce's internal CRM benchmark, Koa matches or exceeds leading frontier models on CRM actions with three times fewer errors.
Two details worth flagging before you repeat that number in a buying committee:
- It's an internal benchmark. No third-party validation has been published, and no latency or throughput figures were disclosed. "3x fewer errors" is Salesforce grading Salesforce's homework — plausible (specialized models routinely beat generalists on narrow tasks), but unverified.
- The trust boundary is the real product. Salesforce owns the model weights and runs post-training and inference entirely inside its own infrastructure. No customer data crosses to OpenAI, Anthropic, or Google during inference. For regulated industries, that single property may matter more than any benchmark.
Pilot customers already running Koa include 1-800Accountant, Baxter Credit Union, Formula 1, UChicago Medicine, and Xero. GA is expected this winter in US regions only.
Why a CRM-Specific Model Is a Big Deal (and Why It Might Not Be)
The case for: most Agentforce failures aren't intelligence failures, they're action failures — the agent picks the wrong tool, updates the wrong field, or loops on a multistep workflow. A model post-trained on exactly those action patterns should fail less, and since errors in a CRM agent mean corrupted pipeline data, error rate is the metric that matters. It mirrors what we're seeing across the stack: Gong, Apollo, and ZoomInfo all shipped agent builders in the same two-week window, and everyone is converging on the idea that generic chat is over and domain-tuned execution is the game.
The case against: Koa only reasons over what's in Salesforce. Your CRM is a record of what already happened — meetings logged, stages updated, deals closed. The signal that creates pipeline (who visited your pricing page today, which target account just hired a new VP of Sales, who's comparing you against a competitor right now) lives outside the CRM boundary that makes Koa trustworthy in the first place. A model that's 3x more reliable at updating opportunities doesn't help if the opportunity never gets created because the intent signal died in a tool Koa can't see.
That's the structural trade of every CRM-native agent: the tighter the trust boundary, the smaller the world the agent can act on.
AIforce: Salesforce Everywhere Your Reps Already Are
AIforce is the second half of the announcement, and for day-to-day sellers it's the more tangible one. Salesforce describes it as a live interface layer that takes the data, workflows, business logic, and permissions inside your org and pushes them out to external tools — no migration, existing permissions enforced, zero data retention promised from model providers.
It ships as three surfaces:
| Surface | What it does | Status |
|---|---|---|
| Claudeforce | Salesforce inside Claude — 37 prebuilt sales skills via a Salesforce MCP server | Beta |
| Slackforce | Salesforce inside Slack, including dashboards generated from natural-language requests | Beta |
| Agentforce Coworker | An AI teammate living in the Lightning search bar | GA — 100,000 users activated in the first 35 days |
We covered Claudeforce in depth when the beta opened — the short version is that pipeline reviews, account research, and opportunity hygiene from a Claude chat window are genuinely useful, with real limits around write actions.
Under the hood, all three surfaces run on the same MCP-based architecture (Salesforce calls it Headless 360), which means agents in Claude, ChatGPT, and Cursor can invoke Salesforce capabilities directly. Add the new long-horizon agent runtime — agents that pursue multi-week objectives like reviving a stalled deal, expected GA in November — and the direction is unambiguous: Salesforce wants to be the system agents act through, not the tab reps click into.
This is the same play HubSpot made with Agent Hub in its Fall 2026 Spotlight, and the same logic behind OpenAI's always-on Dots agents. Every vendor has concluded the CRM UI is where rep productivity goes to die.
What It Costs
Salesforce disclosed no Koa-specific pricing, which almost certainly means it's consumed through existing Agentforce billing. That model, as of October 2026:
- Flex Credits: $500 per 100,000 credits. A standard agent action burns 20 credits — $0.10 per action. Voice actions burn 30 credits ($0.15).
- Per-conversation: a flat $2.00 per resolved conversation, regardless of complexity. Break-even against Flex Credits is 20 actions per interaction — below that, credits are cheaper.
Run the math on a modest sales motion: 5,000 agent conversations a month at three actions each is $1,500/month at list on Flex Credits — before Sales Cloud seats, before Data 360, before the premium SKUs that AIforce surfaces will likely require once they leave beta. If you're evaluating the full stack, start with our Salesforce Sales Cloud pricing breakdown, because the agent line item sits on top of everything in it.
The pattern to watch: specialized models like Koa are cheaper to run than frontier models, but Salesforce bills you per action, not per token. Efficiency gains in inference don't automatically show up on your invoice.
Should Your Sales Team Care?
If you're a Salesforce shop: yes, concretely. Agentforce Coworker is GA and effectively free to try if you have Lightning. Put it in front of your ops team this quarter. For Koa, ask your AE two questions when GA lands: whether Koa-powered actions are priced the same as frontier-model actions, and whether you can see per-model error rates on your org's workflows instead of the internal benchmark. If the 3x error claim holds on your data, that's real money in prevented pipeline hygiene cleanup.
If you're not on Salesforce: Koa changes nothing for you directly, but the strategic signal matters. Every major platform now has a domain-tuned agent story — Salesforce has Koa, Google has Gemini 4 Argon with its hallucination-rate pitch, Anthropic's Opus models anchor half the GTM tools shipping this quarter. The buying question has shifted from "which model is smartest" to "which system sees the signals that matter for my pipeline."
Either way: the gap in every CRM-native agent story is the same one. Koa reasons brilliantly over records that exist. It can't see the anonymous visitor on your pricing page, the target account researching your category, or the champion who just changed jobs — the signals that happen before a CRM record exists. That's the layer MarketBetter operates in: identifying in-market accounts and orchestrating outreach from intent signals, then handing clean, qualified context to whatever CRM and agent stack you run. A reasoning model that's 3x more accurate at step twelve doesn't fix a pipeline that loses deals at step one.
If you want to see what signal-to-outreach orchestration looks like in front of your own traffic, book a demo.
The Bottom Line
Koa is the most credible "small specialized model beats big general model" bet a CRM vendor has made — real training infrastructure, a defensible trust-boundary story, named pilot customers, and a winter GA date. AIforce is the more immediately useful half for sellers, because it meets reps where they already work. Treat the 3x benchmark claim as a hypothesis to test on your own org, budget for per-action pricing that doesn't care how efficient the model is, and remember that the best CRM reasoning in the world only operates on the pipeline you've already managed to create.