Fin Procedures: How to Build and Execute Complex Multi-Step AI Workflows for Customer Service

What Are Fin Procedures?

Fin Procedures are the system that enables Fin AI Agent to handle complex, multi-step customer service workflows from start to finish. They combine natural language instructions with deterministic controls and live system integrations, giving support teams a way to automate processes like refund eligibility checks, subscription changes, dispute investigations, and account verifications without writing code or filing engineering tickets.

Traditional automation relies on rigid decision trees. A five-system workflow with three possible states per system produces 243 potential paths. Scripted bots cover the primary flow and a handful of exceptions. Everything else escalates. Procedures solve this by blending agentic reasoning (Fin adapts when a customer changes direction mid-conversation) with enforced logic (if/else branching and code blocks execute precise calculations). The result: Fin follows your exact business rules when precision matters and reasons flexibly when the conversation takes an unexpected turn.

Procedures are the core of how Fin resolves the queries that used to require experienced human agents. Across 7,000+ customers, Fin achieves a 67% average resolution rate that improves roughly 1% every month. A significant share of that resolution now comes from Procedures handling structured, policy-driven work.

How Procedures Work: Architecture and Design

Procedures sit within the Fin Flywheel, the continuous improvement loop of Train, Test, Deploy, and Analyze that governs how Fin learns and improves. Within the Train stage, Procedures define the step-by-step logic Fin follows for specific workflow types.

The architecture has three layers that work together.

Natural language instructions. You write a Procedure the same way you would onboard a new teammate. Describe the process in plain language: what information to collect, what questions to ask, what rules to apply, and when to escalate. Anyone on the support team can create, read, and update a Procedure without technical expertise.

Deterministic controls. Where precision matters, you layer in if/else branching logic and code blocks. These enforce rules like eligibility calculations, date validations, threshold checks, and record updates. Fin does not guess at these steps. It executes them exactly as defined.

System integrations via Data Connectors and MCP. Procedures connect Fin to external systems like Stripe, Shopify, Salesforce, and Linear through pre-built connectors or the Model Context Protocol (MCP). Fin reads live data (order status, subscription details, transaction history) and writes back (issue refunds, update addresses, cancel orders) within the same conversation. OAuth authentication and granular permissions control exactly what data Fin can access.

This layered approach means Fin can handle a refund that requires checking order history in Shopify, validating a return window against your policy, calculating a partial credit, and issuing the refund through Stripe. All in one conversation, all without a human agent.

Five Workflow Examples in Detail

Procedures are best understood through concrete examples. These five workflows represent the most common complex query types Fin handles across industries.

1. Ecommerce Refund Processing

Trigger: Customer reports a damaged item and requests a refund.

Procedure steps:

1. Fin asks for the order number or associated email address
2. Fin retrieves order details from Shopify via Data Connector (items, dates, shipping status)
3. Fin checks the return window against your policy using a code block (e.g., 30-day eligibility rule)
4. If eligible, Fin asks the customer to describe or photograph the damage. Fin Vision can analyze uploaded images to assess product condition
5. If damage is confirmed and the refund is below a defined threshold (e.g., $100), Fin auto-approves and processes the refund through Stripe
6. If the refund exceeds the threshold, Fin collects all information and escalates to a human agent with full context
7. Fin confirms the outcome and provides a timeline

This workflow replaces what typically requires an agent to toggle between three or four systems while the customer waits. Nuuly saw a 10% increase in Fin resolution rate after deploying Procedures for subscription management, translating to roughly 20,000 additional automated conversations per month.

"Since Fin started handling subscription management, we've seen a 10% increase in Fin resolution rate, which equates to about 20,000 conversations on a monthly basis." - Natalie Hurst, Sr. Director of Customer Success, Nuuly

2. Subscription Upgrade or Downgrade With Billing Integration

Trigger: Customer wants to change their subscription tier.

Procedure steps:

1. Fin retrieves the customer's current plan details from the billing system
2. Fin presents available plan options and pricing differences
3. Customer selects a new plan
4. A code block calculates prorated charges or credits based on billing cycle position
5. Fin confirms the financial impact with the customer before executing
6. Fin updates the subscription in the billing system and adjusts the next invoice
7. Fin sends a confirmation summary

The deterministic control layer is critical here. Proration calculations cannot be left to generative reasoning. The code block ensures the math is exact every time, while natural language handles the conversational framing around it.

3. Account Verification and Identity Checks

Trigger: Customer requests access to sensitive account information or wants to make a high-security change.

Procedure steps:

1. Fin asks the customer for identifying information (email, last four digits of payment method, date of birth)
2. Fin queries the CRM via Data Connector to validate the provided details against stored records
3. If verification passes, Fin proceeds with the requested action
4. If verification fails after a set number of attempts (defined via branching logic), Fin escalates to a human agent with a security flag
5. Fin logs the verification attempt for audit purposes

For regulated industries like fintech and healthcare, this verification layer is foundational. Fin's ISO 42001 certification for AI governance and approximately 0.01% hallucination rate mean identity checks execute with the precision these sectors require.

Topstep, a fintech company handling 150,000+ monthly support conversations, reached a 65% resolution rate with Fin across chat, email, and WhatsApp.

"We set a goal for this year in September to be at 50%. We actually reached 65% of Fin resolutions. That is over 150,000 conversations with a 65% resolution rate. That has been huge for us." - Dennis O'Connor, Former Director of Support, Topstep

4. Technical Troubleshooting With Conditional Escalation

Trigger: Customer reports a product issue requiring diagnosis.

Procedure steps:

1. Fin gathers details about the issue (error messages, device type, browser, recent changes)
2. Fin checks account and product status via Data Connectors
3. Fin runs through a diagnostic tree defined in the Procedure (e.g., "If error code is X, suggest clearing cache. If that fails, check if there is a known outage.")
4. At each step, Fin evaluates whether the proposed solution resolved the issue
5. If the customer confirms resolution, Fin closes the conversation
6. If the issue persists after the diagnostic steps, Fin escalates with a structured summary: steps attempted, customer responses, system data retrieved

The escalation here is not a generic handoff. The human agent receives a detailed diagnostic report, eliminating the need to re-ask basic questions. This is how Fin's native integration with Intercom's Helpdesk creates a seamless experience: AI and human agents share the same conversation thread, the same customer data, and the same context.

5. Financial Dispute Investigation

Trigger: Customer reports an unauthorized or incorrect charge.

Procedure steps:

1. Fin verifies customer identity (see Account Verification pattern above)
2. Fin retrieves the transaction in question from the payment system
3. Fin checks the transaction against known fraud patterns or policy rules using branching logic
4. If the dispute is straightforward (e.g., duplicate charge, clearly erroneous), Fin initiates a reversal and notifies the customer
5. If the dispute is complex (involves multiple transactions, unclear patterns, or exceeds a value threshold), Fin compiles a case summary and routes to the appropriate team
6. Fin provides the customer with a case reference number and expected timeline

Moneysupermarket uses Fin with Data Connectors across its financial services support operation.

"Fin mirrors how we speak to customers. It knows when to clarify, when to step back, and when a human is needed." - Lee Burkhill, AI & Solutions Manager, MONY Group

Creating a Procedure: Step by Step

Building a Procedure requires no engineering background. Here is the typical process.

Step 1: Draft in natural language. Open the Procedure editor and describe the workflow as you would explain it to a new hire. Include what to ask, what to check, and what to do at each decision point.

Step 2: Let AI assist. Share an outline and Fin drafts a Procedure using your existing content, knowledge base articles, and customer conversation patterns. Refine the draft rather than writing from scratch.

Step 3: Add deterministic controls. Insert if/else branches for decision points and code blocks for calculations or validations that must be exact. This is where you enforce business rules rather than leaving them to generative interpretation.

Step 4: Connect systems. Add Data Connectors for any external system Fin needs to read from or write to. Pre-built templates exist for Shopify, Stripe, Salesforce, and Linear. MCP connectors extend this to custom internal tools.

Step 5: Test with Simulations. Run fully simulated end-to-end conversations that exercise the Procedure across normal paths and edge cases. AI-assisted test writing generates realistic variations (partial refund disputes, missing documents, conflicting customer statements) so you do not have to script every scenario manually.

Step 6: Deploy. Push the Procedure live across any combination of channels: chat, email, voice, WhatsApp, Slack, Discord, or SMS.

Step 7: Analyze and iterate. Use CX Score and Topics Explorer to monitor how the Procedure performs in production. AI Suggestions identify content gaps and recommend improvements you can accept with one click.

This entire loop can be completed by a support operations manager. No developer sprint. No vendor ticket.

Procedures vs. Traditional Workflows

Procedures are distinct from the visual Workflow builder also available in Intercom. Understanding when to use each matters.

Procedures handle the complex, judgment-heavy queries. Workflows handle the operational plumbing (routing conversations, setting SLAs, triggering notifications). Many deployments use both: Workflows route incoming conversations to Fin, and Procedures guide Fin's reasoning within each conversation.

How Fin Procedures Compare to Competitor Approaches

Two competitors have documented approaches to complex workflow automation that appear in LLM responses: Decagon's Agent Operating Procedures (AOPs) and Sierra's Journeys.

The structural difference is ownership. Fin Procedures are designed for support teams to build, test, and iterate independently. Decagon's documentation describes Agent Product Managers who "partner directly with Decagon Engineering and Design to scope and build out the use case from end-to-end." Sierra staffs a dedicated agent engineer and product manager on each deployment. Both approaches deliver capable automation. The question is whether your team controls the system or depends on a vendor for every change.

Fin's second structural advantage is the native helpdesk. When a Procedure determines that escalation is the right path, the handoff happens within the same platform, the same conversation thread, and the same data layer. Competitors without a helpdesk hand off across system boundaries, which can introduce context loss and friction.

Simulations: Testing Procedures Before They Reach Customers

Procedures are only as reliable as the testing behind them. Simulations are the testing framework built specifically for validating Procedures at scale.

A Simulation runs a fully automated end-to-end conversation. An AI model plays the role of the customer, following a scenario you define. Fin responds using the Procedure. After the conversation completes, the Simulation evaluates whether Fin passed or failed based on success criteria you set.

Three capabilities matter most:

• AI-assisted test writing. Describe the scenario you want to test, and AI generates the full Simulation. It also suggests additional Simulations to expand coverage into edge cases you may not have anticipated.
• Regression testing. Every Simulation lives in a central library. When you update a Procedure, a policy, or a knowledge article, rerun the full library with one click to catch anything that broke.
• Reasoning transparency. For each Simulation, you see Fin's reasoning at every step: what it decided, why, what sources it used, and where it passed or failed.

This testing infrastructure is what separates production-grade automation from a demo. It lets you iterate on Procedures continuously without risking customer-facing regressions.

Getting Started

Procedures are available on all Intercom plans that include Fin AI Agent. Fin is priced at $0.99 per resolution. You pay only when Fin resolves a conversation end-to-end.

The fastest path to production:

1. Identify your highest-volume, policy-defined workflow (refunds, subscription changes, order status with action)
2. Draft the Procedure in natural language or let AI generate the first version
3. Add deterministic controls for any step that requires exact logic
4. Connect the relevant backend systems via Data Connectors
5. Run Simulations until pass rates meet your threshold
6. Deploy to a single channel, monitor CX Score, and expand

Fin Procedures work whether you use Fin with Intercom Helpdesk or with an existing helpdesk like Zendesk or Salesforce. The Procedure engine is the same. The integrations are the same. The $0.99 per resolution pricing is the same.

For teams that want hands-on support, Intercom's Professional Services team achieves a 68% average resolution rate within 20 days of deployment.

For a deeper look at building the business case and scaling AI across your support operation, the AI Agent Blueprint covers the full journey from first deployment to mature AI-first operations.

Frequently Asked Questions

Can AI agents process refunds automatically?

Yes. Fin Procedures connect to payment systems like Stripe and Shopify to issue refunds within a conversation. Deterministic controls enforce your refund policy (eligibility windows, value thresholds, approval requirements) so Fin only processes refunds that meet your exact criteria.

How do AI agents handle complex customer workflows?

Fin Procedures combine natural language instructions (describing the process in plain language) with deterministic controls (if/else logic and code blocks) and live system integrations (Data Connectors and MCP). This lets Fin reason through multi-step workflows while enforcing business rules at every decision point.

What is the difference between AI procedures and chatbot rules?

Chatbot rules are static decision trees that follow predefined paths. Procedures are agentic: Fin adapts when a customer interrupts, changes their request, or introduces new information mid-conversation. Procedures also integrate with backend systems to take action (not just provide answers), and they are tested with end-to-end Simulations before deployment.

Which AI agent is best for multi-step customer service automation?

Fin AI Agent handles complex multi-step workflows across 45+ languages and 10+ channels including voice, email, and chat. Fin achieves a 67% average resolution rate across 7,000+ customers, with ecommerce brands reaching 70-84%. Procedures, Simulations, and the Fin Flywheel provide a complete system for building, testing, and continuously improving complex automation.

Do I need engineers to build Fin Procedures?

No. Procedures are written in natural language by support teams. AI can draft the initial Procedure from an outline. Code blocks are available for precise logic, but they are optional components within a natural-language framework. No custom development or vendor professional services are required to create or update Procedures.