Procedures

Support teams that move past simple FAQ deflection hit the same obstacle: queries that require collecting information, checking a live system, applying a business rule, and either resolving the issue or handing off to a human — in the right order, every time. Static scripts fail when customers change their minds mid-conversation. Rigid rule-based bots cannot reason across multiple steps. AI Agent Procedures are the architecture built for exactly this gap.

What is an AI Agent Procedure?

An AI Agent Procedure is a structured, multi-step workflow that instructs an AI agent on how to handle a specific class of complex customer query from intake to resolution. Each Procedure defines when it activates, what steps the agent follows, how it branches based on conditions or live system data, and how it ends — either by resolving the issue autonomously or handing off to a human with full context preserved.

Procedures combine two complementary mechanisms: natural language instructions that give the agent flexibility to adapt as conversations evolve, and deterministic controls (if/else conditions, code-evaluated logic, data connector calls) that enforce rules with precision when accuracy is non-negotiable.

Key characteristics:

• Written in natural language, the same way you would train a new teammate on a documented SOP
• Structured with named steps that the agent follows sequentially but navigates non-linearly as conversations shift
• Extended with branching logic using natural language evaluation or Python code for decisions like eligibility checks or date calculations
• Connected to live systems via APIs, data connectors, or MCPs so the agent can read account status, subscription details, or transaction history before acting
• Tested through Simulations — AI-driven end-to-end test conversations — before going live with customers
• Completed either by autonomous resolution or by handing off to a human agent with all conversation context captured

Why AI Agent Procedures Matter

Support teams handling more than a few hundred conversations per week spend a disproportionate share of time on a predictable set of complex queries: cancellations, refund claims, account verification, subscription changes. Each requires multiple steps and system checks, but the underlying process is repeatable. Without Procedures, that work stays with human agents indefinitely.

The alternative — scripted bots with rigid button paths — fails because customers interrupt, add context, and change their minds. A customer asking for a refund mid-cancellation flow should not have to restart from scratch. A trained agent who follows SOPs with judgment handles that naturally. Procedures give AI agents the same adaptive capability.

Customers building Procedures for high-volume complex queries report meaningful improvement in automation rates. Intercom customers using Fin Procedures for use cases like prescription refill handling, subscription management, and account troubleshooting target 60-80% automation on covered query types — well above what knowledge-base-only automation achieves.

The organizational benefit extends beyond metrics. When agents stop handling the repeatable tier of complex queries, they redirect time to the genuinely novel, sensitive, or high-stakes conversations where human judgment is irreplaceable.

How AI Agent Procedures Work

1. Define the trigger: The Procedure specifies when it activates — which customer intent or query type it handles. The AI agent matches incoming conversations to active Procedures based on this description and conversation examples.
2. Write the steps in natural language: Each step is an instruction block describing a unit of work. Example: "Ask the customer for their order number. Check order status using the order lookup connector. If the order is in processing, offer to cancel. If it has shipped, explain the return window."
3. Add deterministic controls where precision matters: Use if/else conditions to branch based on system data or customer responses. Use Python code for calculations — date validation, eligibility checks, numeric comparisons — that need guaranteed accuracy.
4. Connect to external systems: Data connectors and MCPs allow the agent to read live data from Stripe, Shopify, or internal systems and take actions like updating records or triggering downstream workflows.
5. Define handoff behavior: Specify when the agent should escalate to a human teammate, what note to leave, and which team receives the conversation.
6. Test with Simulations before launch: AI-driven test conversations run the full Procedure end to end, judge results against success criteria, and surface edge cases. Saved Simulations can be rerun after any update to catch regressions before they reach customers.

Best Practices for AI Agent Procedures

Start with high-volume, structurally consistent queries. The best candidates have an existing SOP that human agents already follow consistently. If there is no agreed-upon process, build the process first. Procedures automate what your team already does well.

Keep each Procedure scoped to one query type. A Procedure that covers cancellations, refunds, and subscription upgrades in one flow is hard to test and maintain. Separate concerns at the Procedure level; use sub-procedures for shared logic reused across flows.

Use code conditions for anything that must be unambiguously correct. Natural language conditions work well for fuzzy evaluations ("the customer expressed frustration"), but not for date arithmetic, eligibility thresholds, or exact string matching. Use Python code blocks for these cases.

Simulate before every deployment, including after minor updates. Even small instruction changes can affect downstream steps. Run the full Simulation suite after every edit, not only for new builds.

Write trigger descriptions with negative constraints. Specify not just when the agent should use the Procedure, but when it should not. This reduces false positives where the Procedure activates on the wrong query type.

Define the handoff note explicitly. When a Procedure escalates to a human agent, the note it adds to the conversation determines how useful that handoff is. Specify what information to capture and summarize at the escalation step.

AI Agent Procedures vs. Workflow Automation

Procedures and workflow automations are complementary. A common architecture uses a workflow to triage and route conversations to the right team or channel, then activates a Procedure to handle the resolution for complex query types.

Frequently Asked Questions

What kinds of queries are best suited for AI Agent Procedures?

Queries that are high-volume, structurally consistent, and require more than a single knowledge base lookup. Canonical examples include: order cancellations, subscription changes, refund processing, account verification, identity checks, troubleshooting with live system lookups, and policy-based eligibility decisions.

Do AI Agent Procedures replace human agents?

No. Procedures automate the repeatable, multi-step work that currently consumes human agent time. They include explicit handoff steps for queries that require human judgment, and escalate with full conversation context so the agent who takes over has everything they need from the start.

How is a Procedure different from a chatbot flow?

Traditional chatbot flows are rigid: customers must follow button paths, and any deviation breaks the experience. Procedures use AI reasoning at every step, so the agent adapts when a customer changes their mind, asks an unexpected question, or provides information out of order — without losing track of where the process stands.

Can Procedures connect to external systems like Stripe or Shopify?

Yes. Data connectors and MCPs allow Procedures to read from and write to external systems in real time. The agent can check subscription status in Stripe, look up an order in Shopify, or verify eligibility in a custom internal API — and use that data to determine which path to take next.