How Synthflow Eliminated Integration Maintenance to Scale Voice AI Agents

Company: Synthflow
Industry: AI Voice Agents
Challenge: Self-built integrations “always broke,” blocking enterprise sales and draining engineering resources

Goals

• Build reliable OAuth integrations for Salesforce, Zoho, and NetSuite
• Serve enterprise clients with demanding integration requirements
• Eliminate maintenance burden on engineering team

Highlights

• Real-time AI agent infrastructure: Voice AI agents authenticate and execute CRM actions during live conversations
• Per-user authentication at scale: 2000+ end users each authenticate their own accounts — automated token refresh across all instances
• OpenAI-compatible tool calling: LLM-generated actions trigger authenticated integrations without custom parsing
• Zero maintenance: From “always broke” to fully automated OAuth management

Company Overview

Founded: 2023
Industry: AI Voice Agents
HQ: Europe
Website: https://synthflow.ai/

Synthflow is the leading AI Voice Platform for enterprises. Their AI voice agents automate phone conversations at scale – replacing manual call handling with fast, reliable, and human-like interactions. For Synthflow, integrations aren’t just a feature – they’re the foundation. A voice AI agent that can’t create tickets, update CRM records, or retrieve customer data during live conversations is just an expensive chatbot.

The Challenge

Having recently raised Series A funding, Synthflow was scaling fast. But their integration infrastructure was becoming a critical bottleneck.

The cost of building integrations in-house:

• Constant breakages: Integrations with HubSpot and Go High Level were hard to maintain and often broke
• Enterprise deals at risk: Upper-tier clients had demanding integration requirements; without reliable integrations, they risked losing valuable enterprise customers
• Engineering resource drain: Development team stuck maintaining broken integrations instead of building voice AI
• Complex OAuth management: Managing token refresh across thousands of users where each end user needed to authenticate their own Salesforce account

These technical problems created a strategic crisis. Synthflow couldn’t quickly add new integrations based on customer demand, putting them at risk of losing deals to competitors.

Synthflow’s evaluation criteria:

• OAuth integration for Salesforce, Zoho, and NetSuite with per-user authentication
• OpenAI-compatible tool calling with JSON structure
• Automatic token refresh and authentication flows
• GDPR and SOC2 compliance (Frankfurt-based servers)
• Predictable pricing across 2000+ customer base

Why Other Solutions Failed

Synthflow was using Activepieces as a white-labeled workflow automation provider for internal processes, but was experiencing reliability issues. There was also a fundamental category mismatch: Activepieces is built for internal workflow automation – not designed for embedded integrations where each customer needs to authenticate their own third-party accounts.

Synthflow needed an integration platform – infrastructure purpose-built for embedded use cases where thousands of end users each authenticate their own Salesforce, HubSpot, or CRM accounts, and where OAuth token management, per-user authentication, and multi-tenant architecture are core requirements.

Why Membrane

Synthflow’s voice AI agents need to authenticate and execute CRM actions in real-time during live customer conversations. This requires infrastructure purpose-built for AI agents, not workflow automation.

Membrane offered embedded integration infrastructure with per-user authentication and AI-native architecture:

Per-user OAuth at scale: Membrane manages token refresh automatically across 2000+ customers. Each end user authenticates their own Salesforce account – Membrane handles both third-party tokens and customer authentication tokens. Embedded integration infrastructure, not workflow automation repurposed for multi-tenant use.

OpenAI-compatible tool calling: Actions use standardized JSON matching OpenAI’s specification. Synthflow’s LLM dynamically selects which integrations to trigger during live conversations without custom parsing layers. Infrastructure designed for AI agents from the ground up.

Real-time authentication: When a voice AI agent creates a Salesforce case during a customer call, it authenticates using that customer’s credentials in real-time. Membrane’s Proxy API handles authentication, token management, and API execution while maintaining conversation flow.

Custom object support: Supports unique Salesforce configurations with custom fields and objects – critical when enterprise customers each have different CRM configurations.

White-label authentication: Customers authenticate their Salesforce accounts without leaving Synthflow’s platform.

Most importantly, Membrane meant engineers could stop maintaining integrations and focus on building voice AI agents.

Implementation Journey

Synthflow launched their Salesforce integration first, focusing on core operations their voice AI agents needed during live conversations.

Synthflow uses OpenAI’s structured JSON for tool calling. During a conversation, their AI agent dynamically selects which tools to use based on customer input. Actions are LLM-generated and triggered in real-time.

Example workflow:

Trigger: Customer calls and reports an issue

Processing: Synthflow’s LLM:

• Understands this requires creating a support case
• Extracts key information (urgency, issue type, customer details)
• Generates tool call in OpenAI JSON format
• Calls Membrane’s authenticated Salesforce integration

Output: Case created in real-time using customer’s Salesforce credentials; AI agent responds: “I’ve created case #12345 for you”

Core Salesforce actions built:

• Create Case: Voice AI creates support cases during conversations
• Create Action: Logs follow-up actions from call interactions
• Create Task: Assigns tasks based on conversation outcomes

Membrane enabled universal actions that work across all CRM systems. Synthflow’s “Create Task” action:

• Takes standardized input (subject, content, status, priority, ownerId)
• Maps to Salesforce’s Task object structure
• Works consistently across Salesforce, HubSpot, and other CRMs
• Same field mapping across all customer instances

It meant that Synthflow could build once, and deploy multiple times.

The developer experience transformed:

• Pre-configured settings: Developers could use pre-configured settings without additional setup
• Automated authentication: No more building OAuth flows or managing token refresh
• Custom compatibility: Their non-standard action parsing worked seamlessly with Membrane’s Proxy API

The Result

With Membrane, Synthflow gained infrastructure that enables AI agents to act on behalf of users in real-time at scale.

Immediate wins:

• Zero maintenance: From “always broke” to fully automated OAuth management across 2000+ customers
• Enterprise deals unlocked: Serves upper-tier clients with demanding integration requirements
• Engineering focus restored: Team freed from maintaining OAuth flows to build voice AI capabilities
• Real-time actions: Voice agents authenticate and execute CRM actions during live conversations

Architectural transformation:

• Per-user authentication at scale: 2000+ users each authenticate their own Salesforce account – automated token refresh across all instances
• AI-native infrastructure: OpenAI-compatible tool calling means LLMs generate integration actions natively
• Universal action patterns: Build actions once that work across Salesforce, HubSpot, and other CRMs
• White-label experience: Customers never leave Synthflow’s platform when authenticating third-party systems

What’s Next

With Membrane’s infrastructure, Synthflow can rapidly expand their integration catalog – each new integration follows the same pattern: OpenAI-compatible actions, per-user OAuth, universal field mappings.

The vision: AI voice agents that can connect to any enterprise system, authenticate on behalf of any user, and take any action – all during natural conversation flow.

As Synthflow scales to serve more enterprise customers, Membrane’s architecture scales with them: predictable per-customer pricing, automated token management, and zero maintenance burden.