01
Overview
Salesforce is making unusually specific performance claims for Koa, its new CRM reasoning model for Agentforce. The headline is that Koa matches or exceeds leading model performance on CRM actions with three times fewer errors.
That sounds impressive, but the useful question is what Salesforce actually measured — and what the launch-day evidence cannot yet tell us.
02
The headline Koa benchmark claim
Salesforce says its CRM Benchmark includes real-world-style tasks such as updating an opportunity, routing a case and scheduling a follow-up. On that benchmark, Salesforce reports that Koa matches or exceeds leading model performance while producing three times fewer errors.
The benchmark is focused on CRM actions and tool use, not general knowledge, coding, mathematics or creative tasks. Koa is a domain-specific reasoning model, so its most relevant evaluation is whether it chooses and executes the correct actions inside enterprise workflows.
03
Salesforce's additional Koa performance numbers
Salesforce's Koa product page reports three additional comparisons against what it calls today's default general-intelligence models:
- 11% more precise at calling the right action.
- 2.1 times greater reliability.
- 15% better at remembering context in long back-and-forth conversations.
Salesforce connects these metrics to practical agent behavior: fewer incorrect tool calls, more reliable answers and less need for users to repeat context during longer conversations.
04
What does '3x fewer errors' mean?
Salesforce frames the result around CRM action errors. The launch material says the benchmark includes tasks such as updating opportunities, routing cases and scheduling follow-ups.
It should not be interpreted as 'Koa makes one-third as many errors on every AI task.' The result belongs to Salesforce's CRM evaluation environment and the exact comparison depends on the benchmark design and reference models.
05
Why action precision matters for AI agents
A conversational model can give a plausible answer and still choose the wrong tool or business action. In CRM workflows, that difference is important: routing a case incorrectly or updating the wrong field can create an operational failure even when the generated explanation sounds convincing.
Koa was post-trained on synthetic workflows that explicitly map tasks to sequences of actions and tool calls. Salesforce says this training included supervised fine-tuning and reinforcement learning using Group Relative Policy Optimization, with NVIDIA NeMo tooling.
That architecture helps explain why Salesforce is emphasizing action selection rather than conventional language-model leaderboards.
06
Is Koa better than Claude, GPT or Gemini?
The available evidence does not justify that conclusion in general.
Salesforce says Koa matches or exceeds leading models on its CRM Benchmark, but the launch materials do not establish a universal ranking across coding, reasoning, research, writing or other workloads. Nor should a CRM-specialized benchmark be used to declare a general winner.
A fair comparison would need the same models, prompts, tools, Salesforce environment, success criteria and cost/latency measurements evaluated under transparent conditions.
Until independent or reproducible evaluations appear, the strongest defensible statement is narrower: Salesforce reports that Koa performs very well on its own CRM-focused action benchmark.
07
Who produced the benchmark?
The launch performance figures come from Salesforce. They are vendor-reported claims, not an independent third-party certification.
That does not make them useless. They provide an early view of what Salesforce optimized Koa for and which metrics the company considers important. But buyers should distinguish a vendor benchmark from independent production evidence.
08
What is missing from the launch-day benchmark picture?
Several details would make Koa easier to evaluate independently:
- complete results for every comparison model;
- cost and token usage under identical workloads;
- latency measurements;
- failure categories and severity;
- reproducible evaluation instructions;
- performance outside CRM workflows;
- production-scale customer results after general availability.
Those gaps are especially important because Koa remains in select customer pilots.
09
Why the benchmark still matters
Koa's evaluation strategy is notable even before independent results arrive. Salesforce is measuring the model around business actions rather than only abstract reasoning tests.
That reflects a larger shift in agent evaluation: an enterprise agent needs to select the correct tool, follow policy, maintain context and complete a workflow reliably. A model can score well on general benchmarks and still fail at those operational jobs.
For teams considering Agentforce, Koa's real test will therefore be production task completion, error rates, latency and cost after broader access becomes available.
10
FAQs
What is Salesforce's Koa CRM Benchmark?
Salesforce describes it as a suite of real-world CRM tasks including actions such as updating opportunities, routing cases and scheduling follow-ups.
Does Koa have three times better accuracy?
That is not Salesforce's wording. Salesforce says Koa produces three times fewer errors while matching or exceeding leading model performance on CRM actions.
Is the Koa benchmark independent?
No. The launch figures discussed here are Salesforce-reported.
Does the benchmark prove Koa is better than GPT, Claude or Gemini?
No. It supports a narrower claim about Salesforce's CRM-focused evaluation. A general comparison would require broader and preferably independent testing.
Is Koa generally available?
Not yet. Salesforce says Koa is in select Agentforce customer pilots, with U.S. general availability expected in winter 2026.
11
Bottom line
Koa's early benchmark story is promising for CRM agents, particularly around tool selection and workflow reliability. But the numbers should be read as vendor-reported CRM-specific evidence, not a universal AI leaderboard. The most valuable next evidence will come from reproducible evaluations and real production deployments after Koa expands beyond its pilot.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.