Salesforce Koa: The First CRM Reasoning Model and What It Means for Agentforce Buyers
Salesforce Koa is its first in-house CRM reasoning model, built on NVIDIA Nemotron. What it powers, what it costs, and 5 buyer questions to ask before piloting.
TLDR: Salesforce Koa, announced September 15 at Dreamforce 2026, is Salesforce’s first in-house reasoning model - a version of NVIDIA’s Nemotron 3 Super post-trained on synthetic CRM scenarios, built to power Agentforce agents. The headline claim (matching frontier models on CRM tasks with 3x fewer errors) comes from a Salesforce-administered benchmark nobody independent has verified, and Koa is still in pilot with GA expected winter 2026 in US regions only. If you’re already running Agentforce, Koa is worth a pilot conversation with your AE because in-house inference should eventually mean cheaper Flex Credit burn - but don’t sign anything that prices Koa as GA, and don’t assume it replaces Claude workloads already deployed in Slack, Headless 360, or Coworker.
Why This Matters Now
Salesforce just became a model vendor. That’s not a product note. It’s a procurement event.
On September 15, Marc Benioff brought Jensen Huang on the Dreamforce main keynote stage to announce Koa: a reasoning model Salesforce built itself, by post-training NVIDIA’s Nemotron 3 Super on what Salesforce describes as 27 years of CRM domain knowledge. The stated goal is blunt - cut reliance on costly third-party AI labs for the inference that powers Agentforce agents. TechCrunch called it “everything the AI labs should fear.” But you’re not an AI lab. You’re the person who has to decide whether this changes your Agentforce roadmap, your Flex Credit budget, or your negotiating position at renewal. That’s what this article is for, because nobody else at Dreamforce wrote that version.
Koa at a Glance
| Dimension | What We Know (verified Sept 20, 2026) |
|---|---|
| What it is | First Salesforce-built reasoning model, post-trained from NVIDIA Nemotron 3 Super |
| What it powers | Agentforce agents; selectable as a managed LLM in Data Cloud, as a model provider in Setup, or per-agent in Prompt Builder |
| Training data | Synthetic scenarios across 14+ industries; Salesforce says no customer data was used |
| Headline claim | Matches or beats leading models on CRM tasks with 3x fewer errors - on Salesforce’s own CRM Benchmark |
| Independent verification | None published. Zero. |
| Availability | Pilot only (1-800Accountant, Baxter Credit Union, Formula 1, UChicago Medicine, Xero). GA expected winter 2026, US regions |
| Pricing | Not announced. No SKU, no Flex Credit rate |
| Context window | 1 million tokens |
| Architecture | Hybrid Mixture-of-Experts (inherited from Nemotron 3 Super) — cheap per-inference by design |
| Trust boundary | Salesforce controls the weights; inference runs inside Salesforce infrastructure |
What Koa Actually Is (and Isn’t)
Strip away the keynote lighting and Koa is a specialization bet: Nemotron 3 Super plus supervised fine-tuning and reinforcement learning (GRPO via NVIDIA’s NeMo tooling) on synthetic scenarios mirroring Agentforce tool calls - lead gen, opp qualification, case resolution - across 14+ industries.
But the architecture matters less than the data choice. Salesforce trained on synthetic scenarios, not your records - a deliberate legal and privacy call, and the reason regulated pilots like Baxter Credit Union and UChicago Medicine can even consider it. Inference stays inside Salesforce’s trust boundary. No hop to an outside lab.
What Koa isn’t: a general-purpose frontier model, a replacement for Claude across the Salesforce surface area, or a generally available product. Hold all three of those lines when your AE brings it up.
Earned insight: I’ve watched this exact movie before with Einstein’s early “AI-powered” features - the vendor ships a model milestone, AEs start selling the roadmap as the product, and IT leaders get quoted timelines that engineering never committed to. Koa’s GA is “this winter,” pricing is unannounced, and the pilot list is five named customers. If your AE puts Koa in a Q4 proposal, ask for the pilot eligibility criteria in writing before you let it near a business case.
The 3x-Fewer-Errors Claim Deserves a Cold Read
Koa’s headline number - matching or exceeding leading models on complex CRM tasks with three times fewer errors - comes from Salesforce’s CRM Benchmark. Salesforce designed the test, Salesforce administered it, and Salesforce reported the result. No independent auditor has verified it.
The problem is that the AI industry has a documented benchmark-reliability problem. A UC Berkeley team showed in April 2026 that eight standard AI agent benchmarks could be manipulated to near-perfect scores without solving a single real task. Apple’s GSM-Symbolic work found performance drops of up to 65% when high-scoring models faced minor variations of the same problems. The international AI safety community’s 2026 report put it plainly: “performance on pre-deployment tests does not reliably predict real-world utility.”
None of that proves Koa’s number is wrong. It proves a vendor-administered benchmark on a vendor-built evaluation set isn’t evidence you can put in a business case. The real evidence arrives when the five pilot customers - an accounting firm, a credit union, a racing outfit, a hospital system, and an accounting software vendor - report production results. Until then, treat “3x fewer errors” as a keynote slide, not a specification.
Warning: Salesforce’s benchmark optimizes for the CRM failure modes it chose to measure: wrong opportunity stage updates, misrouted cases, misdirected follow-ups. Your org’s failure modes - custom objects, industry-specific validation rules, the Apex triggers your predecessor wrote in 2019 - aren’t in that benchmark. Pilot Koa against your own workflows, not theirs.
Why Salesforce Built Its Own Model: Follow the Margin
The honest frame is economics, not innovation theater. Every Agentforce conversation routed to a third-party lab costs Salesforce per-inference money it can’t control. The August 2026 price war (GPT-5.6 Luna down 80%, Gemini agent costs down 65%) helped, but Salesforce still pays lab rates for every rented reasoning step. Koa moves inference in-house on a Mixture-of-Experts design that activates only a subset of parameters per request: big capacity, fraction-of-dense-model cost.
But there’s a second motive. Benioff’s keynote line - “the most valuable thing Salesforce has built isn’t our platform, it’s the accumulated knowledge of how enterprise business actually works” - is a moat argument. Twenty-seven years of deal structures and case lifecycles, distilled into weights nobody else has. A frontier model has seen more of the internet. Koa has seen more of a sales pipeline. Domain density versus parameter count: that’s the bet.
Earned insight: 24/7 Wall Street’s post-Dreamforce analysis nailed the variable that actually matters: whether Koa carries pricing power and whether pilot accounts convert to production before fiscal year-end. Salesforce stock is up 24% in the month into Dreamforce on AI momentum, but analysts lifted targets to $300 while CRM shares simultaneously retreated on SaaS pricing-power concerns. Translation: the market believes the story and doubts the monetization in the same breath. You should too - that’s your leverage at renewal.
Koa vs. Claude vs. Frontier Models: The Positioning Triangle
Here’s the question nobody at Dreamforce answered directly: Salesforce announced Claude as the default model for Slack AI, Slackbot, Salesforce in Claude, Headless 360, Agentforce Coworker, and Claude Code on August 26 - then announced its own house model three weeks later. So which model runs your work?
| Dimension | Koa (house model) | Claude (Anthropic partnership) | Third-party frontier (OpenAI, Gemini) |
|---|---|---|---|
| Best for | High-volume, routine Agentforce CRM actions | Conversational surfaces (Slack, Coworker), complex reasoning | Model-choice flexibility via AWS/GCP integrations |
| Availability | Pilot only; GA winter 2026 (US) | Default in five surfaces today | Available now through model router |
| Data boundary | Fully inside Salesforce infrastructure | Governed partnership boundary | Provider-dependent |
| Cost trajectory | Should be cheapest per action (Salesforce-owned inference) | Priced through the partnership | Subject to lab pricing; falling fast in the price war |
| Verification | Self-administered benchmark only | Established evals + production history | Established evals + production history |
| Lock-in risk | Highest - weights, hosting, and platform all Salesforce | Medium - portable model, Salesforce-exclusive surfaces | Lowest - multi-vendor routing |
The likely end state is the “Trusted Models” router vision: the right model per job, chosen on accuracy, performance, and cost, with Koa as the cheap house option for routine CRM volume. That’s a sensible architecture. What’s missing is the price list that makes the routing decision real. Until Salesforce publishes Koa’s Flex Credit rate, “cheapest per action” is a hypothesis, not a fact.
Koa Strengths:
- Purpose-built for CRM tool-calling workflows, not retrofitted from general chat
- 1M-token context window reasons across full customer histories and long case chains
- No customer data in training; inference stays inside Salesforce’s trust boundary
- MoE architecture means structurally cheaper inference than dense frontier models
- Selectable at three levels: Data Cloud catalog, org-wide Setup default, per-agent Prompt Builder
Koa Weaknesses:
- Pilot-only with five named customers; no production track record
- Performance claims rest entirely on an unverified self-administered benchmark
- No pricing, no SKU, no GA date more specific than “this winter”
- US regions only at GA - data residency questions unanswered for EMEA/APAC
- Nemotron open-weight lineage raises license-term questions for the largest buyers
Pricing Reality: The TCO Math Nobody Can Finish Yet
Let’s be honest about what we can and can’t compute. The known inputs, verified September 2026:
- Flex Credits cost $500 per 100,000 credits. A standard Agentforce action burns 20 credits ($0.10); voice actions burn 30 ($0.15). The legacy $2-per-conversation model still exists as an alternative.
- Agentforce packaging runs four routes: a $5 user license, $125 and $150 add-ons, Agentforce 1 Editions from $550, and pure consumption.
- SiliconANGLE’s September 19 channel check (via Qualitate) found most organizations interviewed still in Agentforce pilots or proofs of concept - meaning most buyers haven’t yet felt production-scale Flex Credit burn.
The unknown input is Koa’s rate. If Salesforce prices Koa inference below third-party routing - which the MoE economics say it can - then high-volume, routine agent actions (lead qualification, case triage, follow-up scheduling) get materially cheaper per action, and the TCO case for scaling Agentforce past pilot improves. If Salesforce prices Koa at parity and pockets the margin, your bill doesn’t move and the “reduces reliance on costly labs” line describes Salesforce’s P&L, not yours.
Tip: Before your next Agentforce renewal conversation, pull your last 90 days of Flex Credit consumption by action type and split routine CRM actions (triage, qualification, routing) from complex ones (multi-step resolution, cross-object reasoning). That split is your Koa leverage: it’s the volume you’d move to the house model the day it has a published rate. Walk in with the number.
Pricing verification date: September 20, 2026. Koa-specific pricing: not announced as of this writing.
Who Should / Shouldn’t Care Right Now
Good fit - start the pilot conversation now:
- You’re already running Agentforce in production with meaningful Flex Credit burn on routine CRM actions. You’re the buyer Koa was built for.
- You’re in a regulated industry (financial services, healthcare) where the no-customer-data-training plus in-boundary inference story clears a compliance review that third-party routing wouldn’t.
- You’re an architect standardizing the “Trusted Models” router pattern and want the house option evaluated alongside Claude and frontier models before GA locks in defaults.
Not a good fit - wait for GA and a price list:
- You haven’t deployed Agentforce yet. Koa doesn’t change the pilot-first sequencing - Qualitate’s channel checks show most orgs are still in PoC, and a model announcement doesn’t shortcut that.
- Your AI workloads live primarily in Slack, Coworker, or Headless 360 surfaces where Claude is the default. Koa doesn’t touch those yet.
- You’re outside the US. GA is US regions only, and data residency answers for EMEA/APAC haven’t been published.
- You need audited model evaluations for procurement sign-off. There are none. Your security review will stall on this - plan for it.
The Buyer Checklist: 5 Questions for Your AE
- What are the pilot eligibility criteria, and are we on the list? Five named pilots means gated access. Get the criteria in writing.
- What will Koa’s Flex Credit rate be relative to third-party model routing? If they can’t answer, the cost story is aspirational. Don’t model savings you can’t source.
- How does Koa interact with Claude workloads we’ve already deployed? Claude is the default in five surfaces. You need the routing rules, not a vibe.
- What’s the data residency roadmap beyond US regions at GA? Winter 2026, US-only leaves EMEA/APAC buyers with a compliance gap.
- What are the Nemotron open-weight license terms as they apply to us? Open weights flow downstream into obligations. Your legal team will ask - ask first.
Bottom Line
Koa is the most structurally interesting thing Salesforce announced at Dreamforce 2026, because it’s the only announcement that changes the unit economics of running agents rather than just the surface area. A house model with MoE inference costs, trained on CRM-shaped work, selectable per-agent in Prompt Builder - that’s a real architecture, not a keynote rendering. But architecture isn’t a product: it’s in pilot with five customers, its headline benchmark is self-graded, and it has no price. The practitioner move is to open the pilot conversation now (access is gated and early positioning matters), refuse to let Koa into any Q4 business case as GA, and walk into renewal with your routine-vs-complex Flex Credit split in hand. Within 30 days, pull that 90-day consumption breakdown and send your AE the five checklist questions in writing.
FAQ
What is Salesforce Koa?
Koa is Salesforce’s first in-house CRM reasoning model, announced September 15, 2026 at the Dreamforce main keynote. It was built by post-training NVIDIA’s Nemotron 3 Super base model on synthetic CRM scenarios spanning 14+ industries - lead generation, opportunity qualification, case resolution - using supervised fine-tuning plus reinforcement learning. Salesforce controls the weights, runs inference inside its own infrastructure, and says no customer data was used in training. It powers Agentforce agents and can be selected as a managed LLM in the Data Cloud model catalog, as an org-wide model provider in Setup, or per-agent in Prompt Builder.
Is Salesforce Koa generally available?
No. As of September 20, 2026, Koa is in pilot with five named customers: 1-800Accountant, Baxter Credit Union, Formula 1, UChicago Medicine, and Xero. Salesforce expects general availability in winter 2026, limited to US regions. There is no GA SKU, no published pricing, and no self-serve enablement path - pilot access is gated, so talk to your account team about eligibility rather than looking for a Setup toggle.
How much will Koa cost in Flex Credits?
Salesforce hasn’t announced Koa pricing. For context, standard Agentforce actions currently cost 20 Flex Credits ($0.10) and voice actions cost 30 ($0.15), with credits sold at $500 per 100,000. Because Koa runs on Salesforce-owned infrastructure using a Mixture-of-Experts architecture with structurally cheaper per-request inference, it should eventually price below third-party model routing - but “should” isn’t a rate card. Don’t model Koa savings into any budget until a per-action rate is published.
Does Koa replace Claude in Salesforce products?
No - and this is the positioning question to get straight. On August 26, 2026, Salesforce named Claude the default model for Slack AI, Slackbot, Salesforce in Claude, Headless 360, Agentforce Coworker, and Claude Code. Koa, announced three weeks later, targets Agentforce agent workloads. The long-term vision is a “Trusted Models” router that picks the right model per job, with Koa as the house option for high-volume routine CRM actions. If you’ve already deployed Claude-powered surfaces, nothing about Koa displaces them today.
Was customer data used to train Koa?
According to Salesforce, no. The training corpus was built from synthetic scenarios modeled on 27 years of Salesforce CRM domain knowledge - persona-plus-task simulations with mapped tool calls - rather than real customer records. That’s why regulated organizations like a credit union and a hospital system are in the pilot cohort. Inference also runs inside Salesforce’s trust boundary with Salesforce-controlled weights, so there’s no per-request hop to an outside model provider.
How does Koa compare to frontier models like GPT and Claude?
Salesforce claims Koa matches or exceeds leading models on complex CRM tasks with 3x fewer errors - but that figure comes from Salesforce’s own CRM Benchmark, with no independent verification published. Architecturally, the argument for Koa is domain density over parameter count: a smaller model post-trained on CRM-shaped reasoning versus a generalist that has seen more of everything. The argument against is track record: frontier models have established evals and production history, Koa has a keynote slide. Judge it on your own workflows in pilot, not on the benchmark.
Should my Agentforce agents use Koa or a third-party model?
If you’re in the pilot and running high-volume routine actions (triage, qualification, routing), Koa is the natural candidate - that’s the workload its training and its cost structure target. Keep complex multi-step reasoning and conversational surfaces on Claude or frontier models until Koa has production evidence there. Architecturally, select the model per agent in Prompt Builder rather than setting an org-wide default you can’t yet justify. Revisit the split once Salesforce publishes Koa’s Flex Credit rate.
Does Koa work outside the United States?
Not at launch. General availability expected winter 2026 is US regions only, and Salesforce hasn’t published a data residency roadmap for EMEA or APAC. If you’re subject to data sovereignty requirements, Koa’s in-boundary inference story only helps once there’s a boundary in your region. Non-US buyers should treat Koa as a 2027 planning item and get the residency timeline in writing from their account team.
Related Articles
Discussion