Publish date:
SALESFORCE AI & AGENTFORCE
Salesforce Koa: A Practical Guide Beyond the Launch Headlines
Most coverage of Salesforce Koa repeats the launch headlines. This guide goes further: what Koa actually is, what Salesforce's own data does and doesn't prove, and how to test it in your Salesforce org before committing.
QUICK ANSWER
What You Need to Know About Salesforce Koa
- CRM reasoning model: Koa is Salesforce's first CRM reasoning model, built for Agentforce agents.
- NVIDIA foundation: It uses NVIDIA's open-weight Nemotron 3 Super, post-trained by Salesforce on synthetic CRM scenarios without customer data.
- Performance claims: Salesforce reports that Koa matches or exceeds leading models on its CRM benchmark with three times fewer errors. Its technical report also acknowledges that Koa remains below the strongest frontier models on some comparisons.
- Availability: Koa is currently in pilot, with US general availability expected in winter 2026.
- Adoption: Koa is a model choice inside Agentforce, allowing eligible teams to test it and switch back without rebuilding their agents.
What Is Salesforce Koa?
Salesforce Koa is Salesforce's first CRM reasoning model, designed to power Agentforce agents. Salesforce and NVIDIA announced it at Dreamforce on September 15, 2026.
Salesforce developed Koa by post-training NVIDIA's open-weight Nemotron 3 Super model on synthetic scenarios informed by nearly three decades of CRM deployments.
Unlike a general-purpose chatbot, Koa focuses on multi-step CRM workflows. It is designed to reason through tasks such as qualifying leads, routing service cases, updating opportunities, and selecting the appropriate tool at each step.
MODEL OVERVIEW
Salesforce Koa at a Glance
| What it is |
A CRM-specific reasoning model for Agentforce |
| Announced |
September 15, 2026, at Dreamforce, by Salesforce and NVIDIA |
| Base model |
NVIDIA Nemotron 3 Super (120 billion parameters, open weights) |
| Training data |
Synthetic CRM scenarios across 14+ industries; no customer data used for post-training |
| Hosting |
Operated by Salesforce within its own trust boundary |
| How you select it |
Data Cloud generative models catalog, org-wide Agentforce Setup, or per agent and sub-agent |
| Status |
Pilot customers now; US general availability expected in winter 2026 |
| Pricing |
No Koa-specific price published as of October 26, 2026 |
Sources: Salesforce Press Release, Salesforce Koa Product Page, Salesforce Technical Report.
Why Salesforce Koa Matters
01
It's a Model Choice, Not a New Product
Koa is the fourth model provider option in Agentforce Setup. Your existing actions, topics, and agents remain in place. Eligible teams can test Koa without rebuilding their Agentforce implementation.
02
It's Specialized Rather Than General
Most models used by CRM agents are designed for broad tasks. Koa focuses on choosing the right CRM action in the right sequence. Salesforce argues that reducing wrong-action failures can improve agent reliability.
03
It Changes the Data Conversation
Salesforce controls Koa's weights and operates it within its infrastructure. For organizations where security review has delayed AI adoption, this architecture may simplify parts of the approval process, subject to their own compliance requirements.
What Salesforce Koa Is Not
Not a Download
You can't self-host Koa or call it independently outside Agentforce.
Not a Replacement for Other Models
Koa is an additional model option alongside the models already available within Agentforce.
Not Generally Available Yet
Koa remains in pilot. Named pilot customers include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.
COMING UP NEXT
How Salesforce built Koa, how synthetic CRM training works, and why the training method matters when evaluating the model for your business.
SALESFORCE KOA TECHNOLOGY
How Was Salesforce Koa Built?
Salesforce built Koa in three steps. It started with NVIDIA's open-weight Nemotron 3 Super model, generated synthetic CRM training scenarios instead of using customer data, and then post-trained the model with supervised fine-tuning and reinforcement learning. Salesforce keeps the resulting model weights and runs Koa inside its own infrastructure.
THREE-STEP MODEL DEVELOPMENT
01 Base Model NVIDIA Nemotron 3 Super
02 Synthetic CRM Data Industry-specific simulated workflows
03 Post-Training Supervised fine-tuning and reinforcement learning
Koa's Training Pipeline at a Glance
| Stage |
What Happened |
Why It Matters to You |
| 1. Base model |
NVIDIA Nemotron 3 Super, a 120-billion-parameter open-weight model |
Salesforce didn't train a model from scratch, so Koa inherits a strong general foundation. |
| 2. Training data |
Synthetic scenarios modeled on nearly three decades of CRM deployments across 14+ industries |
No customer records were used, simplifying questions about training-data privacy and provenance. |
| 3. Post-training |
Supervised fine-tuning followed by reinforcement learning with GRPO |
Koa is trained to take correct actions in sequence, not merely generate plausible text. |
| 4. Tooling |
NVIDIA NeMo RL, NeMo Gym, and NeMo Automodel |
The training pipeline uses NVIDIA's open training stack. |
| 5. Hosting |
Salesforce controls the model weights and runs inference inside its own trust boundary |
Customer data doesn't need to cross to an external model provider for Koa inference. |
Sources: Salesforce Press Release and Salesforce Technical Report.
STEP 01
The Base Model: NVIDIA Nemotron 3 Super
Koa starts from NVIDIA Nemotron 3 Super, an open-weight model with approximately 120 billion parameters. Independent technical write-ups describe it as a hybrid architecture combining state-space layers with standard attention, supporting very long contexts of up to one million tokens in the base model.
Because the model weights are open, Salesforce could post-train the foundation model and retain control of the resulting CRM-specialized model.
IMPORTANT FOR YOUR ORG
The base model's maximum context window isn't necessarily the context window available to your Agentforce agents in production. Confirm actual limits during setup and testing.
STEP 02
Synthetic Training Data, Not Customer Data
Salesforce says no customer data was used to train Koa. Its training corpus consists of synthetic scenarios, each combining a persona, a set of CRM tasks, and the sequence of actions needed to complete them.
The scenarios span more than 14 industries, including manufacturing, financial services, healthcare, and travel.
WHY COMPLIANCE TEAMS SHOULD CARE
Salesforce's synthetic-data approach means customer records were not used in Koa's stated training process. This reduces concerns about customer-data exposure during training, although organizations should still assess runtime data handling, permissions, retention, and compliance requirements.
STEP 03
Training for Actions, Not Just Answers
Salesforce post-trained Koa using supervised fine-tuning followed by reinforcement learning with Group Relative Policy Optimization (GRPO).
In simple terms, the model repeats tasks, receives feedback on whether it followed the right steps, and learns from better-performing attempts.
This approach supports Salesforce's goal of improving action accuracy. In CRM workflows, an incorrect tool call, record update, or field change can create more serious business problems than an awkwardly written response.
TECHNICAL INSIGHT
The Design Choice Most Articles Miss
Salesforce's technical report describes a distinctive approach: the company reused workflow specifications that configure Agentforce agents as training environments.
For enterprise workflows, these specifications are written in Agent Script, Salesforce's declarative language for defining how an agent handles work. For public tool-use domains, the workflow structure was synthesized directly.
What this means in practice: Koa was trained against the same kinds of workflow structures that Agentforce agents use. That offers a plausible explanation for why Salesforce expects the model to follow Agentforce workflows reliably. However, this is an interpretation of the technical report, not a guarantee of results in your org.
Where Does Salesforce Koa Run?
Salesforce controls Koa's model weights and runs both post-training and inference within its own trust boundary. Koa isn't an external third-party model API that Salesforce calls for each request.
Salesforce-Controlled Weights
Salesforce maintains control of the post-trained Koa model.
Salesforce-Hosted Inference
Inference runs within Salesforce's infrastructure and trust boundary.
No Customer Self-Hosting
Customers cannot self-host Koa or independently fine-tune its weights.
IMPORTANT LIMITATIONS
What Koa's Training Does Not Tell You
1. Koa Wasn't Trained on Your Org
Your custom objects, bespoke Flows, and unusual field logic aren't necessarily represented in synthetic training data. Test Koa on your own workflows before relying on benchmark results.
2. Synthetic Training Isn't Proof of Production Reliability
Koa is currently in pilot, and its published results come from Salesforce's own evaluations. Real-world reliability still needs to be demonstrated in each customer's environment.
3. Training Method Doesn't Determine the Performance Ceiling
According to Salesforce's technical report, Koa improves on its base model and outperforms a strong proprietary baseline, GPT-4.1, in the reported evaluation, but remains below the strongest frontier models.
COMING UP NEXT
What Salesforce's benchmarks actually show, how Koa compares with other models, and the context behind the headline claim of "3x fewer errors."
SALESFORCE KOA PERFORMANCE ANALYSIS
What Do Salesforce's Koa Benchmarks Actually Show?
Salesforce's technical report shows that Koa improves slightly on its Nemotron base model, beats GPT-4.1, and comes close to Claude Opus 4.8 on Salesforce's CRM Bench (0.86 vs. 0.87). It still trails the strongest frontier models across the overall benchmark comparisons.
The report states that Koa remains "below the strongest frontier models." The headline marketing figures come from Salesforce's launch materials and product page, rather than the benchmark table in its technical report.
KEY TAKEAWAY
Koa's strongest argument is not that it beats every frontier model. Instead, it offers a CRM-specialized, Salesforce-hosted reasoning model with competitive task performance and greater control over model operations.
Benchmark Results from Salesforce's Technical Report
| Model |
Tau2Bench
Weighted Avg. |
BFCL
Overall Accuracy |
CRM Bench
Weighted Avg. |
| Claude Opus 4.8 |
74.00 |
78.18% |
0.87 |
| OpenAI GPT-5.5 |
83.99 |
67.63% |
0.90 |
| OpenAI GPT-4.1 |
54.48 |
53.96% |
0.81 |
Nemotron-3-Super-120B
Koa's base model |
68.64 |
64.73% |
0.84 |
Salesforce Koa
CRM-SPECIALIZED |
69.41 |
66.63% |
0.86 |
All figures are vendor-reported. Source: Table 1, Salesforce Koa Technical Report.
What Each Benchmark Measures
Tau2Bench
Measures end-to-end, multi-turn customer service across airline, retail, and telecom scenarios, with four trials averaged.
BFCL
Berkeley Function Calling Leaderboard measures tool use, including multi-step calling, web search, memory, and stateful tools.
CRM Bench
Measures single-turn Salesforce and Agentforce workflows, including topic accuracy, function-call accuracy, and free-text accuracy.
How to Read These Numbers: 5 Important Findings
FINDING 01
Koa Clearly Beats GPT-4.1
Koa scores about 15 points higher on Tau2Bench (69.41 vs. 54.48) and approximately 13 points higher on BFCL (66.63% vs. 53.96%). Salesforce's claim that it surpasses a strong proprietary baseline refers to this older model.
FINDING 02
Koa Is Competitive with Opus 4.8 on CRM Bench, but Not Elsewhere
Koa scores 0.86 against Claude Opus 4.8's 0.87 on CRM Bench. On Tau2Bench, it trails by about 4.6 points, and on BFCL, it is approximately 11.5 percentage points behind.
FINDING 03
GPT-5.5 Leads on Two of the Three Benchmarks
GPT-5.5 scores 83.99 on Tau2Bench and 0.90 on CRM Bench. Koa trails by approximately 14.6 points and 0.04, respectively. On BFCL, Koa is within one percentage point of GPT-5.5 (66.63% vs. 67.63%).
FINDING 04
Koa's Improvement over Its Base Model Is Modest
Koa improves on Nemotron-3-Super-120B by less than one point on Tau2Bench (69.41 vs. 68.64) and by 0.02 on CRM Bench (0.86 vs. 0.84).
Its clearest improvement is on BFCL multi-turn tool use, which increased from 54.12% to 59.50%. The technical report identifies multi-turn tool use as a meaningful area for improvement.
FINDING 05
Function-Call Accuracy Is the Area to Test Most Carefully
On CRM Bench, Koa scores 0.77 for function-call accuracy, below Opus 4.8 (0.83), GPT-5.5 (0.82), and GPT-4.1 (0.85).
Koa performs better on free-text accuracy (0.85), ahead of Opus 4.8 (0.79) and GPT-4.1 (0.60). If your Agentforce agents depend on precise action selection, prioritize function-call accuracy in your pilot.
MARKETING CLAIMS VS. PUBLISHED EVIDENCE
Where Do the "3x Fewer Errors" Figures Come From?
| Claim |
Source |
In the Technical Report? |
| Matches or exceeds leading model performance on CRM actions with 3x fewer errors |
Salesforce press release, based on its CRM Bench |
Not found in the report text |
| 11% more precise at calling the right action |
Salesforce product page, compared with today's default general intelligence models |
Not found in the report text |
| 2.1x more reliable context recall |
Salesforce product page |
Not found in the report text |
| 15% better at remembering long conversations |
Salesforce product page |
Not found in the report text |
Sources: Salesforce Press Release, Salesforce Koa Product Page, Salesforce Technical Report.
What Should Businesses Make of These Claims?
These figures may be based on internal evaluations that Salesforce has not fully published. As of this writing, the reviewed technical report did not identify the methodology or comparison models behind these specific marketing claims. Treat them as reasons to run a controlled pilot, not as guaranteed performance or budget assumptions.
Limits of the Published Evidence
- Vendor-reported results: All published results come from Salesforce. No independent replication was available at the time of writing.
- Limited task samples: Tau2Bench's weighting uses 50 airline, 114 retail, and 114 telecom tasks.
- Different test conditions: The report serves Salesforce's checkpoints through vLLM in BF16 but doesn't specify all versions or settings used for proprietary models.
- Starting-point limitations: The report notes that Koa's base model was already reinforcement-learning post-trained, so findings may differ with another starting model.
- Not your Salesforce org: Benchmarks use public or synthetic tasks rather than your custom objects, Flows, integrations, and production data.
BUSINESS DECISION
What Does This Mean for Your Salesforce AI Strategy?
Koa's published evidence supports a specific, measured conclusion: it is a Salesforce-hosted, governed model that performs competitively on CRM-focused tasks and outperforms some older proprietary models.
It does not establish that Koa outperforms the strongest frontier models overall.
For enterprises evaluating Agentforce, the strongest potential case for Koa is model control, data handling, and routing flexibility, rather than raw benchmark leadership. Test its action accuracy, reliability, and operational fit against your own workflows before deciding.
COMING UP NEXT
How Salesforce Koa's security model works, what Salesforce's trust boundary means for enterprise data, and the questions your security and legal teams should ask before adoption.
SALESFORCE KOA SECURITY & GOVERNANCE
Is Salesforce Koa Secure? How the Trust Boundary Works
Salesforce says Koa is secure because it runs entirely inside Salesforce's own infrastructure. Salesforce controls the model weights, no customer data was used in training, and no customer data crosses Salesforce's trust boundary during training or inference.
Salesforce also describes two additional controls: Koa is optimized to follow your instructions and guardrails, and a dedicated serving harness adds trust and safety protections around the model.
That's a stronger data-path story than sending prompts to an external model provider. It isn't a complete security assessment, though. Here's what Salesforce claims, what each claim covers, and what your organization still needs to verify.
QUICK SECURITY ANSWER
Is Salesforce Koa Safe for Enterprise Use?
Koa's Salesforce-hosted architecture may simplify data governance by keeping model inference within Salesforce's trust boundary. However, security approval still depends on your organization's access controls, grounding design, compliance obligations, regional requirements, logging policies, and testing results.
Salesforce Koa's Three Security Layers
| Layer |
What Salesforce Says |
What It Means for You |
What It Doesn't Cover |
| 1. Location |
Koa lives entirely inside Salesforce's trust boundary, with customer data remaining under customer ownership and control. |
Prompts don't need to go to an outside model provider. |
Your organization's permission model and data quality. |
| 2. Behavior |
Koa is optimized to follow instructions and guardrails and is hosted at temperature 0. |
More consistent behavior and easier testing. |
A guarantee against incorrect actions. |
| 3. Serving Harness |
A dedicated serving layer adds trust and safety controls around the model. |
An additional layer of protection if the model behaves unexpectedly. |
Unpublished details of the harness's specific checks. |
Source: Salesforce Koa Product Page and FAQ.
What "No Customer Data in Training" Does and Doesn't Mean
Salesforce states that no customer data is used to train Koa and that its training corpus consists entirely of synthetic scenarios. At runtime, Salesforce says Koa only sees what you choose to share through grounding data, records, and instructions.
TRAINING
The Provenance Question
Salesforce's claim means Koa's training data did not include your customer records. This removes a major question: whether customer information shaped the model's weights, subject to verifying the vendor's statements.
RUNTIME
The Access-Control Question
The training claim doesn't limit which records Koa can access during inference. If an agent is grounded on sensitive records, Koa may reason over them. Your permissions and grounding design remain essential.
Temperature 0: Consistency, Not a Guarantee
Salesforce hosts Koa at temperature 0 to encourage consistent and repeatable responses. This helps governance because similar inputs are more likely to produce similar action plans, making testing, audits, and incident reviews easier.
PRACTICAL TESTING TIP
Treat temperature 0 as more consistent, not guaranteed identical. Test which action the agent calls, which arguments it passes, and which records it changes. Run those tests before and after any model change instead of relying on exact response wording.
The Data Path Changes, So Rerun Your Reviews
According to independent coverage from SFDC Developers, Koa runs within Salesforce's infrastructure. This changes the model data path compared with using models hosted by external providers on platforms such as AWS or Google.
If a data protection impact assessment, vendor review, or security assessment influenced your current model choice, review it again before adopting Koa. The new architecture may improve the outcome, but you still need to complete the review.
TECHNICAL LIMIT TO VERIFY
Context Size with Data Masking
SFDC Developers reports that although Nemotron 3 Super supports up to one million tokens of context, the Einstein Trust Layer limits supported models to 65,536 tokens when data masking is enabled.
This limit was identified in one independent source and was not confirmed in the Salesforce pages reviewed for this draft. If your organization relies on data masking, verify the current context limit in official Salesforce documentation before designing large grounding payloads.
ENTERPRISE SECURITY CHECKLIST
6 Questions to Ask Salesforce Before Approving Koa
01 / COMPLIANCE
Which Certifications Apply?
Which compliance attestations cover Koa specifically? The reviewed sources don't identify Koa-specific certifications.
02 / DATA RESIDENCY
Where Does Inference Run?
With US-first availability and no confirmed EU or APAC date, where will inference run for your org?
03 / LOGGING
What Is Stored and for How Long?
How are prompts, outputs, and action logs stored, accessed, retained, and deleted?
04 / THIRD-PARTY ACCESS
Does NVIDIA Access Runtime Data?
Koa is based on NVIDIA's open-weight model. Confirm whether NVIDIA has any access to runtime data, despite Salesforce-hosted inference.
05 / ACCOUNTABILITY
Who Is Accountable for Wrong Actions?
Which agreement governs incidents when an Agentforce agent makes an incorrect decision or updates the wrong record?
06 / CHANGE CONTROL
How Are Model Updates Managed?
How are model versions tracked, what advance notice is provided, and how can your team test changes?
What the Trust Boundary Does Not Fix
The reasoning model is only one step in the Agentforce workflow. According to SFDC Developers, topic classification, grounding retrieval, and Trust Layer checks remain unchanged when switching to Koa.
Wrong Topic Selection
If your agent selects the wrong topic, Koa may reason about the wrong task.
Incorrect Grounding Data
If retrieval returns the wrong account or record, Koa may act on incorrect information.
Overly Broad Permissions
If permissions expose too much information, Koa may have access to more data than intended.
RECOMMENDED CONTROL
Keep human approval checkpoints for high-impact actions involving money, customers, or compliance. Design clear error handling, rollback, and escalation paths before deploying agents into production.
SECURITY TAKEAWAY
The Bottom Line on Salesforce Koa Security
Koa's hosting model is one of its strongest enterprise arguments. For organizations where security review has delayed AI adoption, a model hosted within Salesforce and trained without customer data may simplify part of the discussion.
Before approval, verify compliance coverage, data residency, retention, permissions, and model update procedures. Then test Koa's behavior in a controlled Agentforce pilot rather than relying on architecture claims alone.
COMING UP NEXT
When Salesforce Koa becomes available, which regions will get access first, and what Salesforce has disclosed about pricing and rollout plans.
SALESFORCE KOA RELEASE & PRICING
When Is Salesforce Koa Available, and How Much Does It Cost?
Salesforce Koa is currently available only to select pilot customers. Salesforce expects general availability (GA) in winter 2026 in US regions, with an open beta planned shortly after GA.
As of October 26, 2026, Salesforce has not published a Koa-specific price, credit consumption rate, or confirmed international availability date.
Pilot Current Access
Winter 2026 Expected US GA
US First Initial Regions
Unannounced Koa Pricing
RELEASE STATUS TRACKER
Salesforce Koa Availability Tracker
| Question |
Status |
| Available now? |
Select pilot customers only |
| Named pilots |
1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, Xero |
| General availability |
Expected winter 2026, US regions only |
| Open beta |
Planned shortly after GA |
| EU, UK, APAC |
No date published |
| Koa pricing |
Not published |
| Credit consumption per action |
Not published |
| Public pilot sign-up |
No public mechanism documented |
Sources: Salesforce Press Release, Koa Product Page, Aptivus Solutions and SFDC Developers.
Information current as of October 26, 2026. Recheck official Salesforce announcements for changes.
Which Regions Get Salesforce Koa First?
Salesforce's initial general availability plan covers US regions. Independent coverage also notes that Salesforce has not announced a release date for other regions.
One analysis suggests that organizations requiring EU or UK data residency should not treat Koa as a confirmed 2026 deployment option. If your Salesforce org is outside the US, focus on preparation: build a testing baseline, define a pilot workflow, and establish your evaluation criteria without assuming a go-live date.
IMPORTANT TIMELINE NOTE
"Winter 2026" is Salesforce's stated release window, not a confirmed calendar date. Ask your Salesforce account team how that timing applies to your org and region.
How Much Will Salesforce Koa Cost?
Salesforce has not publicly disclosed Koa-specific pricing or whether it will consume more or fewer credits per action than existing Agentforce model options.
One analyst expects Koa to use existing Agentforce billing, but that remains an inference rather than a confirmed Salesforce policy. If a vendor or implementation partner provides a Koa-specific price, verify it directly with Salesforce before using it in your budget.
EXISTING AGENTFORCE PRICING REFERENCE
What We Do Know: How Agentforce Is Billed Today
Salesforce's published Flex Credits rates provide a starting point for budget scenarios. These are general Agentforce rates, not confirmed Koa-specific charges.
| Item |
Published Rate in Supplied Draft |
| Flex Credits |
$500 per 100,000 credits ($0.005 per credit) |
| Standard Agentforce Action |
20 credits = $0.10 |
| Voice Action |
30 credits = $0.15 |
Reference: Salesforce Agentforce Pricing and Salesforce Help. Salesforce also offers other billing models, including per-conversation and per-user options. Check current terms before budgeting.
Illustrative Monthly Agentforce Cost at Today's Standard Rate
The examples below assume an average of 3 actions per interaction at $0.10 per action. Replace that assumption with your measured usage when planning a real deployment.
| Monthly Interactions |
Actions |
Illustrative Monthly Cost |
| 1,000 |
3,000 |
$300 |
| 5,000 |
15,000 |
$1,500 |
| 20,000 |
60,000 |
$6,000 |
Important: These Are Not Salesforce Koa Prices
The figures are arithmetic examples using the general Agentforce rate supplied above. They exclude other possible charges and do not predict Koa's eventual billing structure.
Why Koa's Efficiency May Not Lower Your Bill
Specialized models can be less expensive to operate than large frontier models. However, Agentforce's action-based billing means lower inference costs don't necessarily translate into lower customer charges.
One third-party analysis also suggests that actions involving roughly more than 10,000 tokens may count as multiple billable actions. This threshold should be verified against Salesforce's current billing documentation rather than assumed.
BETTER COST METRIC
Evaluate Koa using cost per successfully resolved task, not model efficiency alone. Include action consumption, escalations, rework, and the percentage of tasks completed correctly.
What to Do While You Wait for Salesforce Koa
STEP 01
Ask your Salesforce account team about pilot eligibility, regional timelines, and expected pricing.
STEP 02
Choose one high-volume, well-understood workflow with low risk if an agent makes a mistake.
STEP 03
Measure your current model's action accuracy, escalation rate, resolution time, and cost per resolved task.
STEP 04
Build a regression test set from real transcripts so you can compare Koa with your existing model under consistent conditions.
STEP 05
Create an illustrative budget scenario using current Agentforce rates, then update it once Koa-specific pricing becomes available.
Should You Test Koa at GA or Wait for the Open Beta?
| Your Situation |
Suggested Approach |
| US-hosted org, already using Agentforce, with security review as the main bottleneck |
Ask about pilot access now and prepare to test at GA. |
| US-hosted org that wants more evidence first |
Wait for the open beta and evaluate early customer findings. |
| Organization outside the US |
Prepare your baseline now, but don't assume a 2026 go-live date. |
| Organization not yet using Agentforce |
Start by evaluating Agentforce itself, because Koa isn't a standalone product. |
FREQUENTLY ASKED QUESTIONS
Salesforce Koa Availability and Pricing FAQs
Is Salesforce Koa generally available?
Not yet. Koa is available to select pilot customers, with US general availability expected in winter 2026.
Can I get Salesforce Koa outside the US?
Salesforce hasn't published an availability date for regions outside the United States.
How much does Salesforce Koa cost?
Salesforce hasn't published Koa-specific pricing or a credit consumption rate. Existing Agentforce pricing can only be used for illustrative budgeting.
Do I need a separate Salesforce Koa license?
Salesforce hasn't confirmed whether a separate license will be required. Koa is described as a model option inside Agentforce, so verify licensing with your Salesforce account team.
COMING UP NEXT
How to enable Salesforce Koa inside Agentforce once your org has access, where to select the model, and what to configure before running your first pilot.
SALESFORCE KOA SETUP GUIDE
How to Enable Salesforce Koa in Agentforce
Salesforce Koa isn't generally available yet, so you can only enable it if your organization has access through Salesforce's pilot program.
Once you have access, Salesforce says you can select Koa in three places: org-wide in Agentforce Setup, at the agent or subagent level, or as a managed model in the Data Cloud generative models catalog.
The safest approach is to start with a single subagent while keeping the rest of your organization on its existing model.
BEFORE YOU START
Salesforce hasn't publicly documented Koa's exact model API name or released Koa-specific setup instructions. The guidance below combines Salesforce's Koa product information with its general Agentforce model-configuration documentation. Menu labels, model identifiers, and availability may change at general availability.
The Three Ways to Select Salesforce Koa
| Method |
Scope |
Best For |
Risk |
| Org-Wide Model Provider Agentforce Setup |
Sets the org default. Koa becomes a fourth provider option, with customer opt-in. |
Final rollout after a successful pilot. |
Highest: affects agents using the org default and reduces your control group. |
| Agent or Subagent Level Agentforce Builder / Agent Script |
One agent or subagent. |
Your first controlled pilot. |
Lower: other agents can retain their existing model. |
| Managed LLM Data Cloud Generative Models Catalogue |
Prompts, AI applications, and agents using the catalogue. |
Prompt-template and Data Cloud use cases. |
Depends on which applications and prompts reference the model. |
References: Salesforce Koa Product Page and Salesforce Agent Script documentation.
RECOMMENDED IMPLEMENTATION PATH
How to Pilot Salesforce Koa on One Subagent
STEP 01
Confirm Your Koa Access
Koa is currently pilot-only, and Salesforce hasn't documented a public sign-up process. Contact your Salesforce account team to confirm eligibility, regional availability, and the model's exact API identifier.
STEP 02
Choose the Right Subagent
Select a subagent handling a high-volume, well-understood workflow with limited risk if an action fails. Case routing and lead qualification are useful examples.
STEP 03
Record a Performance Baseline
Before changing models, capture representative transcripts from the selected subagent. One independent analysis suggests reviewing 20 to 40 transcripts per high-traffic topic.
Record the actions called, arguments passed, escalation frequency, repeated loops, and task outcomes.
STEP 04
Configure Koa for That Subagent Only
Use the model selector in Agentforce Builder or configure the model through Agent Script. Salesforce's general documentation illustrates the use of a model:// identifier.
AGENT SCRIPT EXAMPLE
model_config:
model: "model://"
Replace with the identifier Salesforce provides. The example is a configuration template, not a working Koa model identifier. Don't guess the value.
STEP 05
Test Before and After Deployment
Salesforce recommends thoroughly testing agents with their selected models before deployment. Some supported models may not be suitable for every agent or tool.
- Missing or incomplete data
- Conflicting customer records
- Permission restrictions
- Failed actions and tool errors
- Unexpected escalation scenarios
Repeat critical tests after publishing, as Salesforce also recommends.
STEP 06
Compare Koa Against Your Control Group
Keep comparable workflows on the existing model so you can evaluate action accuracy, escalations, resolution time, and cost per resolved task. Compare similar task types and conditions rather than relying on subjective impressions.
STEP 07
Expand or Roll Back Based on Results
Expand from the subagent level to agent-level deployment, then consider an org-wide rollout only if the evidence supports it. Treat a broad model change as a controlled release with documented rollback procedures.
How Model Precedence Works in Agentforce
Salesforce's documentation describes a configuration hierarchy in which a more specific model setting overrides a broader default.
LEVEL 01
Org Level
Provides the default model for agents and subagents.
LEVEL 02
Agent Level
Overrides the org-level model default.
LEVEL 03
Subagent Level
Overrides both the agent-level and org-level defaults.
WHY THIS MATTERS
The narrowest applicable configuration wins. This lets you compare Koa with another model inside one Salesforce org. Salesforce also notes that certain templates, including Agentforce Service Agent, may use a Salesforce-owned classifier for routing that can be retained or replaced.
Common Mistakes When Testing Salesforce Koa
Switching Org-Wide First
You change every agent relying on the default and lose a useful control group. Start with one subagent instead.
Testing Wording Instead of Actions
Check which actions were called and which arguments were passed, not just whether the response sounds natural.
Blaming the Model for Upstream Problems
Topic classification and grounding retrieval may remain unchanged. A wrong topic or incorrect record can lead Koa toward the wrong result.
Skipping the Baseline
Without measurements from the existing model, your pilot produces opinions rather than a meaningful comparison.
Assuming Koa Appears in Every Model List
A third-party guide reported that Koa wasn't yet listed on Salesforce's Supported Models page. Check the latest documentation and your pilot entitlements.
GENERAL AVAILABILITY CHECKLIST
What to Verify When Salesforce Koa Reaches GA
- Koa's exact model identifier and where it appears in Agentforce Setup and Builder.
- Supported Salesforce editions, licenses, and geographic regions.
- Pricing and credit consumption per action.
- Whether model settings such as temperature are configurable. Salesforce currently describes Koa as hosted at temperature 0.
- The supported context limit when data masking is enabled.
FREQUENTLY ASKED QUESTIONS
Salesforce Koa Setup FAQs
Where do I enable Koa in Agentforce?
Salesforce says you can select Koa org-wide in Agentforce Setup, at the agent or subagent level, or through the Data Cloud generative models catalog, subject to availability.
Can I use Koa for just one agent?
Yes. Salesforce describes support for model selection at the agent, subagent, and agent-router levels.
Do I have to rebuild my agents to try Koa?
Model selection is a configuration setting, not a full rebuild. However, you should retest relevant prompts, tools, and custom actions after changing models.
Is the Koa model name public?
Not in the documentation reviewed for this guide. Ask your Salesforce account team for the supported model identifier.
COMING UP NEXT
How Salesforce Koa compares with frontier AI models, when a specialized CRM reasoning model makes sense, and how to route different types of Agentforce work to different models.
SALESFORCE KOA MODEL COMPARISON
Salesforce Koa vs Frontier Models: Which Tasks Should Go Where?
Salesforce Koa is best treated as one model among several inside Agentforce, not a replacement for every AI model your organization uses.
In Salesforce's technical report, Koa matches or nearly matches GPT-5.5 in two of three Tau2Bench domains, performs close to Claude Opus 4.8 on CRM Bench, and trails frontier models in function-call accuracy and telecom-style tasks.
That points toward a practical model-routing strategy: pilot Koa for repetitive, well-defined CRM actions, while testing whether more complex tasks should remain on a stronger general-purpose model.
MODEL SELECTION TAKEAWAY
Don't ask whether Salesforce Koa is the best AI model overall. Ask whether it performs better for a specific Agentforce workflow, under your security requirements, at an acceptable cost per successfully resolved task.
Where Salesforce Koa Is Strong and Where It Trails
| Benchmark Test |
Salesforce Koa |
Claude Opus 4.8 |
GPT-5.5 |
GPT-4.1 |
| Tau2Bench: Airline |
62.0 |
69.0 |
62.5 |
56.0 |
| Tau2Bench: Retail |
81.6 |
86.2 |
81.6 |
74.1 |
| Tau2Bench: Telecom |
60.5 |
64.0 |
95.8 |
34.2 |
| CRM Bench: Function-Call Accuracy |
0.77 |
0.83 |
0.82 |
0.85 |
| CRM Bench: Free-Text Accuracy |
0.85 |
0.79 |
0.89 |
0.60 |
| BFCL Overall Accuracy |
66.63% |
78.18% |
67.63% |
53.96% |
All figures are vendor-reported. Source: Salesforce Koa Technical Report, Table 1.
Three Important Patterns in the Benchmark Results
PATTERN 01
Telecom Accounts for Most of Koa's Gap to GPT-5.5
Koa scores 60.5 on telecom tasks compared with GPT-5.5's 95.8. However, Koa matches GPT-5.5 on retail (81.6) and is within half a point on airline tasks (62.0 vs. 62.5).
INTERPRETATION
Performance varies significantly by workflow domain. The airline benchmark contains only 50 tasks, so small differences should be interpreted cautiously.
PATTERN 02
Function-Call Accuracy Needs the Most Attention
Koa scores 0.77 on CRM Bench function-call accuracy, below Claude Opus 4.8 (0.83), GPT-5.5 (0.82), and GPT-4.1 (0.85). Because reliable action execution is central to CRM agents, this should be one of the first metrics tested in your own pilot.
PATTERN 03
Free-Text Accuracy Is a Relative Strength
Koa scores 0.85 on CRM Bench free-text accuracy, outperforming Claude Opus 4.8 (0.79) and GPT-4.1 (0.60), although it remains behind GPT-5.5 (0.89).
IMPORTANT EVIDENCE GAP
Koa Hasn't Been Publicly Compared with Every Agentforce Model
Salesforce's technical report compares Koa with Claude Opus 4.8, GPT-5.5, and GPT-4.1. However, these may not be the models currently configured in your Agentforce organization.
According to SFDC Developers, Koa is positioned as an alternative to reasoning options such as Salesforce Default, Claude Haiku 4.5 on Bedrock, or Gemini.
The reviewed material did not identify a published head-to-head comparison against those specific options. Therefore, you cannot assume Koa outperforms your existing Agentforce model. You need a controlled test in your own environment.
PRACTICAL MODEL ROUTING STRATEGY
Which Agentforce Tasks Should Use Salesforce Koa?
Agentforce supports model selection at the agent and subagent levels, allowing organizations to test different models for different workflows. The following recommendations are starting hypotheses, not proven performance guarantees.
| Work Type |
Starting Hypothesis |
| Repetitive, Policy-Bound CRM Actions Case routing, lead qualification, and opportunity updates. |
Pilot Koa. Measure action accuracy and successful task completion. |
| Open-Ended or Judgment-Heavy Reasoning Unusual exceptions, ambiguous requests, and complex decisions. |
Keep a stronger model initially and test whether Koa can handle the workload. |
| Work Outside Typical CRM Workflows Long-form drafting, coding, and extensive document summarization. |
Keep your current model. The reviewed evidence doesn't establish an advantage for Koa on these tasks. |
| Data-Sensitive Enterprise Workflows Workflows where security or vendor reviews delay adoption. |
Evaluate Koa's Salesforce-hosted architecture as a potential governance benefit, while testing output quality and permissions. |
The Practical Case for Multi-Model Routing
This approach reflects Salesforce's broader strategy: use a specialized CRM model for repetitive, high-volume business actions and keep more capable general-purpose models for harder tasks.
It also reduces the risk of replacing a well-performing model across every Agentforce workflow before you have sufficient evidence.
THE MISSING VARIABLE: COST
Salesforce hasn't published Koa-specific pricing. Once pricing becomes available, compare models using cost per successfully resolved task, not just benchmark scores or estimated inference efficiency.
COMING UP NEXT
A practical, step-by-step Salesforce Koa pilot playbook covering workflow selection, baseline metrics, regression testing, success criteria, and rollout decisions.
SALESFORCE KOA IMPLEMENTATION GUIDE
Salesforce Koa Pilot Playbook: A Step-by-Step Plan
A successful Salesforce Koa pilot compares Koa with your current model on one real business workflow, using a fixed test set and clearly defined success metrics before testing begins.
The timelines below are suggestions rather than fixed Salesforce requirements. Adjust each phase based on your organization's risk level, workflow complexity, and access to the Koa pilot.
THE PILOT PRINCIPLE
Compare Outcomes, Not Just AI Responses
Your pilot should answer three questions: Does Koa choose the correct actions? Does it complete tasks without increasing risk? And does it improve the overall cost or quality of the workflow compared with your existing model?
FIVE-PHASE PILOT FRAMEWORK
How to Run a Salesforce Koa Pilot
PHASE 0 Before You Have Access
Prepare Your Workflow and Success Criteria
Preparation can begin before Salesforce grants Koa pilot access. Start by choosing a suitable workflow and checking whether the underlying Agentforce configuration is ready.
- Pick one workflow. Choose a high-volume, well-understood process with low risk if something goes wrong.
- Check your foundations. Review Salesforce data quality, permissions, and topic definitions. Changing the model won't fix upstream issues.
- Define success in writing. Document what would justify expansion and what would require stopping the pilot.
- Confirm access and region. Contact your Salesforce account team about pilot eligibility.
- Create a budget scenario. Use current Flex Credits pricing as a placeholder until Koa-specific rates are available.
PHASE 1 Suggested Duration: 1–2 Weeks
Establish a Baseline Using Your Current Model
Record how your existing model performs before making any changes. This creates a fair comparison point for Koa.
One independent analysis recommends reviewing 20 to 40 real transcripts per high-traffic topic. Another suggests building a test set around the ten actions your agent gets wrong most often.
EXPECTED OUTPUT
A fixed evaluation set containing representative workflows, frequent failure cases, and baseline performance metrics.
PHASE 2 Sandbox Evaluation
Run the Same Test Cases Against Salesforce Koa
Run your fixed test set against Koa in a sandbox. Include both routine cases and situations where your existing agent has struggled.
Normal and Worst Cases
Typical tasks alongside known failures.
Missing or Conflicting Data
Incomplete fields and inconsistent records.
Permission Restrictions
Requests involving records the agent shouldn't access.
Failed Tools and Escalations
Tool errors, fallback behavior, and handoffs.
Long Conversations
Multi-turn interactions as a stress test.
Evaluation rule: Compare the actions selected, the arguments passed, and the resulting record changes—not how polished the response sounds.
PHASE 3 Controlled Live Pilot
Deploy Koa to One Subagent
Assign Koa to a single subagent while the rest of your organization continues using its existing model. This provides a useful comparison group and limits the impact of unexpected behavior.
Keep human approval in place for actions involving money, contracts, customer commitments, or compliance-sensitive decisions.
Review the pilot metrics weekly and investigate unexpected errors immediately rather than waiting until the end of the trial.
PHASE 4 Go / Hold / Stop Decision
Decide Whether to Expand, Hold, or Stop
Use the criteria you defined before testing. Don't rewrite your success thresholds after seeing the results.
EXPAND
The model meets quality, safety, and business thresholds.
HOLD
Evidence is insufficient or specific issues need further testing.
STOP
The model fails predefined safety or performance criteria.
If you expand, move one scope at a time—from subagent to agent and eventually org-wide. Treat each expansion as a controlled release with a documented rollback plan.
PILOT PERFORMANCE SCORECARD
Salesforce Koa Metrics You Should Track
| Metric |
What It Tells You |
| Action Accuracy |
Whether Koa chooses the right action with the correct arguments. |
| Wrong State-Changing Actions |
Whether the agent makes incorrect changes to data. This is the most important safety metric. |
| Escalation and Handoff Rate |
Whether the agent recognizes its limits and escalates appropriately. |
| Loops and Timeouts |
Whether the agent becomes stuck or fails to finish. |
| Task Completion and Resolution Rate |
Whether the workflow is successfully completed. |
| Time to Resolution |
How quickly the agent resolves a task. |
| Human Correction Rate |
The hidden workload created by manual corrections and reviews. |
| Cost per Resolved Task |
The most useful cost comparison across different models. |
DEFINE YOUR GO / NO-GO THRESHOLDS
Set Success Criteria Before Testing
If you decide what "good enough" means after reviewing results, the pilot may end up confirming your original expectations rather than objectively measuring performance.
A sensible minimum rule is that Koa must not increase incorrect state-changing actions, regardless of improvements in response quality, speed, or cost.
SECURITY & GOVERNANCE
Salesforce Koa Pilot Governance Checklist
- Named accountability: Assign an owner for the pilot and a clearly identified person responsible for rollback.
- Human approval: Require review before high-risk or irreversible actions.
- Audit logging: Enable appropriate logging for actions, outcomes, errors, and escalations.
- Written rollback plan: Document how to return to the previous model before the live pilot starts.
- Security and legal sign-off: Complete the necessary review using the questions covered in the Salesforce Koa security section.
PILOT TAKEAWAY
A Successful Pilot Produces Evidence, Not Just Feedback
A Koa pilot aims to determine whether the model fits your organization's specific workflows. A fixed test set, clear performance thresholds, limited rollout, and documented governance controls make that decision more reliable.
COMING UP NEXT
Which organizations and Agentforce use cases are the best candidates for Salesforce Koa, and which teams should wait for more evidence before adopting it?
SALESFORCE KOA ADOPTION GUIDE
Who Should Use Salesforce Koa, and Who Should Wait?
Salesforce Koa is best suited to teams already using Agentforce, operating in US regions, and automating repeatable, policy-bound CRM tasks where data control matters.
It's a weaker fit for organizations outside the US, teams handling primarily open-ended work, and businesses that need to self-host or fine-tune their AI models.
GOOD FIT
Who Should Consider Salesforce Koa?
Existing Agentforce Customers in US Regions
Koa is a model option within Agentforce, making it relevant to organizations already using Salesforce's agent platform.
Regulated or Security-Sensitive Teams
Salesforce claims that Koa runs within its infrastructure and wasn't trained on customer data, which may simplify some security reviews, subject to verification.
High-Volume, Repetitive CRM Workflows
Case routing, lead qualification, and record updates are practical candidates for a controlled Koa pilot.
Teams Prepared for Structured Testing
Organizations with baseline metrics and regression tests can evaluate Koa against their current model more reliably.
WAIT OR LOOK ELSEWHERE
Who Should Hold Off on Salesforce Koa?
Organizations Outside the US
Salesforce hasn't published a general availability date for non-US regions.
Teams Not Yet Using Agentforce
Koa isn't a standalone product. Start by evaluating Agentforce and your use cases.
Open-Ended or Non-CRM Workloads
Published evidence doesn't establish Koa's strength in judgment-heavy reasoning or work outside CRM.
Teams Requiring Self-Hosting or Fine-Tuning
Salesforce controls Koa's weights and hosts the model within its own infrastructure.
Teams Needing the Strongest Frontier Reasoning
Salesforce's technical report acknowledges that Koa remains below the strongest frontier models.
SALESFORCE KOA READINESS CHECK
Quick Self-Check: Is Salesforce Koa Right for You?
Use these five questions to assess whether Koa deserves a place in your Agentforce testing roadmap.
| Question |
If Yes |
| Are you using Agentforce today? |
Koa is relevant. |
| Is your Salesforce org in a US region? |
You can plan for GA. |
| Is a repeatable workflow one of your highest operational costs? |
Good pilot candidate. |
| Has security review delayed your agent rollout? |
Koa may help. |
| Can you measure results against an existing baseline? |
You're ready to pilot. |
FINAL TAKEAWAY
Start with the Right Workflow, Not a Company-Wide Model Switch
Salesforce Koa may be a strong candidate for teams that value Salesforce-controlled hosting and need reliable automation for repeatable CRM processes. But its suitability depends on your region, Agentforce setup, security requirements, and measured results. Start small, compare it with your current model, and expand only when the evidence supports the decision.
FREQUENTLY ASKED QUESTIONS
Salesforce Koa FAQs
Find answers to common questions about Salesforce Koa, including its features, pricing, availability, security, and Agentforce integration. Click any question to view the answer.
What is Salesforce Koa?
Salesforce Koa is Salesforce's first CRM reasoning model for Agentforce, built by post-training NVIDIA Nemotron 3 Super on synthetic CRM scenarios. It's designed for multi-step workflows such as qualifying leads, routing cases, and updating opportunities.
Was customer data used to train Salesforce Koa?
No. Salesforce states that Koa was trained entirely on synthetic CRM scenarios, without customer data. Salesforce also says customer data doesn't cross its trust boundary during training or inference.
Is Salesforce Koa better than GPT-5.5 or Claude Opus 4.8?
Not overall, according to Salesforce's technical report. Koa outperforms GPT-4.1 and scores close to Claude Opus 4.8 on CRM Bench, but it trails the strongest frontier models on several benchmarks. Performance depends on the task being evaluated.
Is Salesforce Koa generally available?
Not as of October 26, 2026. Koa is available to select pilot customers, with US general availability expected in winter 2026. Salesforce plans an open beta shortly after GA.
How much does Salesforce Koa cost?
Salesforce hasn't published Koa-specific pricing or a credit consumption rate. As a reference, the standard Agentforce action rate in Salesforce's published Flex Credits pricing is 20 credits, equivalent to $0.10 per action. This isn't a confirmed Koa price.
Where do I enable Salesforce Koa in Agentforce?
Salesforce says Koa can be selected org-wide in Agentforce Setup, at the agent or subagent level, or as a managed model in the Data Cloud generative models catalogue. Availability depends on your pilot access and supported configuration.
Can I fine-tune or self-host Salesforce Koa?
No. Salesforce controls Koa's model weights and hosts the model within its own infrastructure. Customer self-hosting and fine-tuning aren't offered.
Should I switch my entire Salesforce org to Koa?
Not initially. Start with a controlled pilot on one subagent, compare its results against your existing model, and expand only when action accuracy, safety, resolution rates, and business outcomes justify a broader rollout.
0
About the Author
Neerav Ahuja
Neerav Ahuja is a 7x Salesforce certified Admin specialist with an experience of 5 years in the Salesforce ecosystem. He currently heads a team of admins and is extensively involved in client engagement & problem solving.
16192 Coastal Highway Lewes, Delaware 19958 USA
