Industry-Specific Use Cases Basics

9 examples to get you started with industry-specific agent use cases - 6 basic and 3 intermediate.

These sketches are conceptual and provider-agnostic. They show how industry context changes risk and control, not a production compliance program.

Prerequisites

Familiarity with a simple agent loop: goal, tools, observe, stop or escalate.
Skim Why Industry Context Changes an Agent's Risk Profile for vocabulary.
No framework required for this page.

Basic Examples

1. Same Support Goal, Two Industries

Start from one user intent and force the industry difference into the open.

goal = "Customer wants a refund for last week's order"
 
retail = {
    "data": ["order_id", "sku", "payment_last4"],
    "tools": ["lookup_order", "issue_refund"],
    "gate": "auto if amount < 50 else human",
}
 
bank = {
    "data": ["account_id", "txn_id", "kyc_flags"],
    "tools": ["lookup_txn", "open_dispute_case"],
    "gate": "human always for money movement",
}

The natural-language goal can look identical in a ticket UI.
Data classes and allowed actions diverge immediately.
Gates follow blast radius, not chatbot personality.

Related: Why Industry Context Changes an Agent's Risk Profile - risk levers

2. Sketch the Loop Once, Specialize the Policy

Keep architecture stable; parameterize policy.

def support_turn(state, tools, policy, llm):
    thought = llm.reason(state, policy.instructions)
    action = llm.choose_tool(thought, tools.allowed)
    if policy.requires_approval(action):
        return {"status": "needs_human", "action": action}
    obs = tools.run(action)
    return state + [thought, action, obs]

One loop shape can serve multiple industries.
Industry differences live in tools.allowed and policy.requires_approval.
Do not encode regulations only as soft wording in the system prompt.

Related: Inside the Agent Loop: Perceive, Reason, Act, Observe - loop primitives

3. Compare Tool Surfaces Side by Side

List tools for retail support vs clinical admin support.

retail_tools = ["search_kb", "lookup_order", "create_return_label", "issue_coupon"]
clinic_tools = ["search_policy", "lookup_appointment", "draft_message", "create_callback_task"]
 
# Deliberately missing on clinic agent:
# diagnose, prescribe, change_meds, send_results_without_review

Retail may execute low-risk actions after simple checks.
Clinical admin agents should prefer draft and task creation over clinical decisions.
Missing tools are a safety feature, not a product gap.

Related: Healthcare Agents: Clinical Documentation and Patient Triage Support - clinical assist patterns

4. Make "Done" Industry-Aware

Stopping conditions include escalation, not only a final answer.

def is_done(state, policy):
    if state.get("needs_human"):
        return True
    if state.get("final_answer") and policy.allows_unsupervised_answer(state):
        return True
    if state["turns"] >= policy.max_turns:
        return True
    return False

In regulated flows, "done" often means "queued for human review."
Max turns still apply so the agent cannot thrash while waiting for clarity.
Silent partial answers are failures, not soft successes.

Related: Stopping Conditions: How an Agent Knows When It's Done - stop rules

5. Evidence Fields Beat Confident Tone

Require structured outputs that carry sources or case IDs.

reply = {
    "message": "Your return label is ready.",
    "evidence": {"order_id": "A123", "policy_id": "RMA-14"},
    "confidence": "high",
    "next": "send_label",
}

Industry reviewers care about evidence more than eloquence.
Empty evidence should block unsupervised send for high-stakes domains.
Confidence scores help triage; they do not replace policy gates.

Related: Internal Knowledge-Base Agents with RAG Grounding - grounding pattern

6. Dual Paths: Draft vs Execute

Separate "prepare" from "commit" in the control flow.

def handle(action, role):
    if action.kind == "draft":
        return {"status": "draft_ready", "body": action.body}
    if action.kind == "execute" and role == "human_approver":
        return run_execute(action)
    return {"status": "blocked", "reason": "execute_requires_human"}

Most industry pilots should ship draft-first.
Execution rights belong to a narrower role or a separate approval step.
This split is how you get value without pretending the model is licensed staff.

Related: Regulated-Industry Agent Checklist: Compliance, Audit, and Human Sign-Off - launch checklist

Intermediate Examples

7. Mini Case: Retail Support Agent Policy

A concrete policy object for a mid-risk retail agent.

retail_policy = {
    "max_turns": 6,
    "auto_refund_limit": 50,
    "pii_fields_logged": ["order_id"],  # not full card data
    "must_cite": ["return_policy"],
    "escalate_if": ["chargeback_threat", "fraud_flag", "vip_account"],
}

Numeric limits are enforceable in code.
Logging policy is part of the use case, not an ops afterthought.
Escalation triggers prevent the agent from negotiating high-risk edge cases alone.

Related: E-Commerce Agents: Product Search, Support, and Merchandising - commerce patterns

8. Mini Case: Bank Support Agent Policy

Same ticket shape, stricter defaults.

bank_policy = {
    "max_turns": 8,
    "auto_money_movement": False,
    "tools": ["lookup_txn", "explain_fee", "open_case", "draft_reply"],
    "human_required_for": ["dispute", "limit_change", "beneficiary_add"],
    "audit": ["prompt_version", "tool_args", "approver_id"],
}

Fee explanation can be assisted; beneficiary changes cannot be casual tool calls.
Audit fields are required outputs of the runtime, not optional debug logs.
"Open case" is often the correct autonomous action instead of "fix money."

Related: Financial Services Agents: Fraud Triage, Reporting, and Trading Support - finance use cases

9. Industry Basics Smoke Checklist

Before building, answer these on one page.

checklist = {
    "industry_stakes": "what fails if the agent is wrong?",
    "data_class": "public | customer | regulated | privileged",
    "autonomy": "draft | recommend | supervised_execute | narrow_auto",
    "tools": "least privilege list",
    "human_gates": "which actions need sign-off?",
    "audit": "what must be reconstructable later?",
    "non_goals": "diagnosis? legal advice? trading decisions?",
}

If non-goals are empty, scope is not finished.
If autonomy is "full" and data is regulated, redesign before coding.
If audit is "whatever the LLM gateway keeps," compliance will fail later.

Related: Industry-Specific Use Cases Best Practices - ten practices | A Decision Framework: Is This Problem a Good Fit for an Agent? - fit check

Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana @solana/web3.js + @solana/spl-token.

Industry-Specific Use Cases Basics

9 examples to get you started with industry-specific agent use cases - 6 basic and 3 intermediate.

These sketches are conceptual and provider-agnostic. They show how industry context changes risk and control, not a production compliance program.

Prerequisites

Familiarity with a simple agent loop: goal, tools, observe, stop or escalate.
Skim Why Industry Context Changes an Agent's Risk Profile for vocabulary.
No framework required for this page.

Basic Examples

1. Same Support Goal, Two Industries

Start from one user intent and force the industry difference into the open.

goal = "Customer wants a refund for last week's order"
 
retail = {
    "data": ["order_id", "sku", "payment_last4"],
    "tools": ["lookup_order", "issue_refund"],
    "gate": "auto if amount < 50 else human",
}
 
bank = {
    "data": ["account_id", "txn_id", "kyc_flags"],
    "tools": ["lookup_txn", "open_dispute_case"],
    "gate": "human always for money movement",
}

The natural-language goal can look identical in a ticket UI.
Data classes and allowed actions diverge immediately.
Gates follow blast radius, not chatbot personality.

Related: Why Industry Context Changes an Agent's Risk Profile - risk levers

2. Sketch the Loop Once, Specialize the Policy

Keep architecture stable; parameterize policy.

def support_turn(state, tools, policy, llm):
    thought = llm.reason(state, policy.instructions)
    action = llm.choose_tool(thought, tools.allowed)
    if policy.requires_approval(action):
        return {"status": "needs_human", "action": action}
    obs = tools.run(action)
    return state + [thought, action, obs]

One loop shape can serve multiple industries.
Industry differences live in tools.allowed and policy.requires_approval.
Do not encode regulations only as soft wording in the system prompt.

Related: Inside the Agent Loop: Perceive, Reason, Act, Observe - loop primitives

3. Compare Tool Surfaces Side by Side

List tools for retail support vs clinical admin support.

retail_tools = ["search_kb", "lookup_order", "create_return_label", "issue_coupon"]
clinic_tools = ["search_policy", "lookup_appointment", "draft_message", "create_callback_task"]
 
# Deliberately missing on clinic agent:
# diagnose, prescribe, change_meds, send_results_without_review

Retail may execute low-risk actions after simple checks.
Clinical admin agents should prefer draft and task creation over clinical decisions.
Missing tools are a safety feature, not a product gap.

Related: Healthcare Agents: Clinical Documentation and Patient Triage Support - clinical assist patterns

4. Make "Done" Industry-Aware

Stopping conditions include escalation, not only a final answer.

def is_done(state, policy):
    if state.get("needs_human"):
        return True
    if state.get("final_answer") and policy.allows_unsupervised_answer(state):
        return True
    if state["turns"] >= policy.max_turns:
        return True
    return False

In regulated flows, "done" often means "queued for human review."
Max turns still apply so the agent cannot thrash while waiting for clarity.
Silent partial answers are failures, not soft successes.

Related: Stopping Conditions: How an Agent Knows When It's Done - stop rules

5. Evidence Fields Beat Confident Tone

Require structured outputs that carry sources or case IDs.

reply = {
    "message": "Your return label is ready.",
    "evidence": {"order_id": "A123", "policy_id": "RMA-14"},
    "confidence": "high",
    "next": "send_label",
}

Industry reviewers care about evidence more than eloquence.
Empty evidence should block unsupervised send for high-stakes domains.
Confidence scores help triage; they do not replace policy gates.

Related: Internal Knowledge-Base Agents with RAG Grounding - grounding pattern

6. Dual Paths: Draft vs Execute

Separate "prepare" from "commit" in the control flow.

def handle(action, role):
    if action.kind == "draft":
        return {"status": "draft_ready", "body": action.body}
    if action.kind == "execute" and role == "human_approver":
        return run_execute(action)
    return {"status": "blocked", "reason": "execute_requires_human"}

Most industry pilots should ship draft-first.
Execution rights belong to a narrower role or a separate approval step.
This split is how you get value without pretending the model is licensed staff.

Related: Regulated-Industry Agent Checklist: Compliance, Audit, and Human Sign-Off - launch checklist

Intermediate Examples

7. Mini Case: Retail Support Agent Policy

A concrete policy object for a mid-risk retail agent.

retail_policy = {
    "max_turns": 6,
    "auto_refund_limit": 50,
    "pii_fields_logged": ["order_id"],  # not full card data
    "must_cite": ["return_policy"],
    "escalate_if": ["chargeback_threat", "fraud_flag", "vip_account"],
}

Numeric limits are enforceable in code.
Logging policy is part of the use case, not an ops afterthought.
Escalation triggers prevent the agent from negotiating high-risk edge cases alone.

Related: E-Commerce Agents: Product Search, Support, and Merchandising - commerce patterns

8. Mini Case: Bank Support Agent Policy

Same ticket shape, stricter defaults.

bank_policy = {
    "max_turns": 8,
    "auto_money_movement": False,
    "tools": ["lookup_txn", "explain_fee", "open_case", "draft_reply"],
    "human_required_for": ["dispute", "limit_change", "beneficiary_add"],
    "audit": ["prompt_version", "tool_args", "approver_id"],
}

Fee explanation can be assisted; beneficiary changes cannot be casual tool calls.
Audit fields are required outputs of the runtime, not optional debug logs.
"Open case" is often the correct autonomous action instead of "fix money."

Related: Financial Services Agents: Fraud Triage, Reporting, and Trading Support - finance use cases

9. Industry Basics Smoke Checklist

Before building, answer these on one page.

checklist = {
    "industry_stakes": "what fails if the agent is wrong?",
    "data_class": "public | customer | regulated | privileged",
    "autonomy": "draft | recommend | supervised_execute | narrow_auto",
    "tools": "least privilege list",
    "human_gates": "which actions need sign-off?",
    "audit": "what must be reconstructable later?",
    "non_goals": "diagnosis? legal advice? trading decisions?",
}

If non-goals are empty, scope is not finished.
If autonomy is "full" and data is regulated, redesign before coding.
If audit is "whatever the LLM gateway keeps," compliance will fail later.

Related: Industry-Specific Use Cases Best Practices - ten practices | A Decision Framework: Is This Problem a Good Fit for an Agent? - fit check

Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana @solana/web3.js + @solana/spl-token.

Industry-Specific Use Cases Basics

Search Documentation

Industry-Specific Use Cases Basics