Salesforce Koa Proves Enterprise Tool Calling Works: What Marketing Teams Should Deploy Today

Enterprise marketing workflows have outgrown standard chat completions. When a campaign automation agent runs, it cannot simply draft email copy; it must query lead scores in the CRM, match custom attributes, check suppression lists across an external warehouse, and trigger webhook payloads with precise JSON syntax.

Salesforce AI Research addressed this exact boundary in their technical report, Salesforce Koa: An Enterprise Language Model for Agentic Tool Use. By post-training an open-weight 120B foundation model using Group Relative Policy Optimization (GRPO) on specification-driven tasks, the Salesforce team achieved an 87% Task Success Rate on CRMAgentBench, beating general frontier baselines like GPT-4.1 (82%).

While running custom 120B parameter FP8 deployments requires dedicated GPU infrastructure, the architectural insights from Koa offer a blueprint for growth marketers and ops engineers. For teams building production agents today, combining those agentic design patterns with high-speed models like Gemini 3.7 Flash through the OneMux unified API delivers enterprise-grade execution without cluster overhead.

---

What Salesforce Koa Proves About Enterprise Tool Use

General foundation models often struggle inside CRM and marketing tech stacks because real workflows have overlapping actions, strict parameter typing, and severe penalties for hallucinations against live records.

The Koa paper introduces several critical findings that directly apply to marketing operations:

1. Specification-Driven Environments: Instead of unstructured multi-turn data, Salesforce derived environments from Agent Script—declarative graphs specifying routing topics, typed action interfaces, and subagent scopes. Expanding these into persona-driven simulated customer rollouts ensured the policy was evaluated on stateful task completion.
2. Grounded Resolution Rewards: Rather than relying purely on intermediate syntax checks, Koa's GRPO pipeline evaluated rollouts using a frozen coverage judge. Rewards scaled according to the ratio of resolved business sub-questions, gating duplicate actions with a zero reward if an identical tool call repeated consecutively.
3. Targeted In-Domain Adaptation: While open-domain reinforcement learning delivered modest lifts on public benchmarks like Tau2Bench and BFCL, continuing RL directly on the in-domain Salesforce CRM distribution sharply improved argument accuracy and full tool-call success on human-labeled enterprise datasets.

For a marketing team managing customer journeys, this highlights a clear reality: model scale alone does not prevent bad API payloads. Reliable execution requires declarative tool schemas, duplicate call checks, and models designed for responsive function routing.

---

Why Gemini 3.7 Flash Fits Modern Marketing Teams

Not every growth engineering team can fine-tune and host a custom Nemotron-120B checkpoint. For marketing stacks that ingest product telemetry, lead records, and content catalogs, Google's Gemini 3.7 Flash serves as a fast, cost-effective inference engine.

1. High-Throughput Lead Routing and Data Enrichment

Marketing agents frequently process asynchronous queues: categorizing inbound form submissions, evaluating firmographic signals, and writing structured updates back to HubSpot or Salesforce. Gemini 3.7 Flash provides the rapid token generation and low latency needed to clear massive batch jobs without stalling campaign delivery.

2. Large Working Memory for Multi-Touch Attribution

Gemini 3.7 Flash includes a 1M token context window. In marketing analytics, multi-touch attribution often requires feeding months of disparate session logs, email click histories, and customer support tickets into a single prompt. Instead of relying on aggressive vector pruning that drops contextual timestamps, the model can parse wide spans of structured customer histories directly.

3. Native Function Calling Precision

Marketing operations rely on deterministic outputs. Whether querying a Google Ads bid API or updating an email sequence stage, malformed arguments break execution. Gemini 3.7 Flash supports native structured tool calling, making it well-suited to enforce the parameter typing championed by the Koa architecture.

---

Production Pricing & Cost Analysis

For marketing teams scaling autonomous outreach or real-time web personalization, token spend compounds rapidly. Accessing Gemini 3.7 Flash via OneMux pricing gives growth teams transparent usage tracking and significant savings against official base rates.

Published Price Snapshot

Prices captured on 2026-10-15T10:20:07.502Z from published catalogue listings. Standard token prices correspond to per 1,000,000 tokens (1M).

Cache reads on Gemini 3.7 Flash are listed at an official price of $0.075 per 1M tokens, with OneMux cache read pricing at $0.01365 per 1M tokens.

Concrete Marketing Workload Calculation

Consider a mid-market automated lifecycle engine executing 50,000 account audits per month:

• Input: 12,000 tokens per account (firmographic data, last 5 user transcripts, historical purchases).
• Output: 800 tokens per account (structured JSON tagging, campaign assignment, payload generation).
• Total Volume: 600M input tokens, 40M output tokens.

Official Gemini 3.7 Flash Billing:

• Input: 600 × $0.75 = $450.00
• Output: 40 × $3.75 = $150.00
• Total Cost: $600.00

OneMux Gemini 3.7 Flash Billing:

• Input: 600 × $0.1365 = $81.90
• Output: 40 × $0.6825 = $27.30
• Total Cost: $109.20

Using OneMux, the marketing team runs the identical workload with an 81.8% reduction in inference expenditure, preserving budget for broader experiment testing across alternate models in the model catalogue.

---

Implementing a Koa-Style Tool Pipeline with OneMux

OneMux provides an OpenAI-compatible routing interface. By configuring declarative tools alongside strict schemas, marketing developers can replicate Koa's task-routing principles in a few lines of code.

Review the OneMux Quickstart Guide to obtain your API key, then invoke Gemini 3.7 Flash:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.onemux.net/v1",
    api_key=os.environ.get("ONEMUX_API_KEY"),
)

# Define declarative marketing tools following Koa's schema isolation
tools = [
    {
        "type": "function",
        "function": {
            "name": "update_lead_lifecycle_stage",
            "description": "Updates a prospect lifecycle stage in the marketing CRM.",
            "parameters": {
                "type": "object",
                "properties": {
                    "lead_id": {"type": "string"},
                    "lifecycle_stage": {
                        "type": "string",
                        "enum": ["marketing_qualified", "sales_accepted", "disqualified"]
                    },
                    "qualification_reason": {"type": "string"}
                },
                "required": ["lead_id", "lifecycle_stage", "qualification_reason"]
            }
        }
    }
]

response = client.chat.completions.create(
    model="gemini-3-7-flash",
    messages=[
        {
            "role": "system",
            "content": "You are an enterprise marketing ops router. You only execute verified transitions."
        },
        {
            "role": "user",
            "content": "Lead ID 98231 completed two product tours and booked an enterprise demo. Transition stage."
        }
    ],
    tools=tools,
    tool_choice="auto"
)

# Inspect the model call
tool_call = response.choices[0].message.tool_calls[0]
print(f"Action: {tool_call.function.name}")
print(f"Arguments: {tool_call.function.arguments}")

By centralizing calls through a unified API, engineering teams retain single-key spend visibility and failover capabilities across frontier models without refactoring their payload infrastructure.

---

Practical Recommendations for Marketing Operations

1. Adopt Declarative Sub-Tasks: Do not ask one model call to analyze an account, formulate a campaign, and push updates. Break workflows into deterministic subagents (e.g., Lead Verifier, Message Drafter, CRM Syncer) as demonstrated in the Koa paper.
2. Enforce Duplicate-Call Suppression: Mitigate infinite loop billing. If your agent calls the same CRM endpoint with the exact same parameters consecutively, drop the turn and flag for human review.
3. Balance Speed and Reasoning: Use Gemini 3.7 Flash for synchronous enrichment and operational routing. Reserve heavier reasoning engines for complex narrative strategy and multi-variable attribution modeling.

Sources

• Chen, Z., Niu, S., Liu, Y., Zhao, W., Prabhakar, A., Mehrotra, S., et al. (2026). Salesforce Koa: An Enterprise Language Model for Agentic Tool Use. arXiv:2609.15066v2. https://arxiv.org/html/2609.15066v2