Customer support agent with approved refunds

Routes support requests, answers from a policy knowledge base, and issues refunds only after a person approves.

router risk high v1 customer support seed

Why this topology

interactive-latency: only workflow or router (interactive latency)
high-risk-no-autonomy: autonomous topology removed (risk is high)
prefer-workflow: agentic topologies removed (the task is not open-ended)
autonomous-needs-consent: autonomous topology removed (autonomy is act_with_approval)
multiple-task-kinds: router (3 request kinds)

Mandatory controls: human_gate input_filter audit_log retrieval

Flow

classify → Request router · 600 in / 50 out
answer → Answer writer · 2500 in / 400 out

Diagram source (Mermaid)

flowchart LR
    io_in(["request"])
    io_out(["response"])
    c_router_model(["Request router"])
    c_answer_model(["Answer writer"])
    c_policy_kb[("Policy knowledge base")]
    c_orders_tool[["Orders API"]]
    c_refund_gate{{"Refund approval"}}
    g_pii_filter>"input_filter"]
    g_audit>"audit_log"]
    io_in --> c_router_model
    c_router_model --> c_answer_model
    c_answer_model --> io_out
    g_pii_filter -.-> c_router_model
    g_pii_filter -.-> c_answer_model
    g_audit -.-> c_orders_tool

Components

id	name	kind	role	model / tool / framework
router-model	Request router	model	classify the request as order help, policy question or refund	claude-haiku-4-5
answer-model	Answer writer	model	write the customer answer from retrieved policy text	claude-sonnet-5-5
policy-kb	Policy knowledge base	retrieval	policy and order-help articles
orders-tool	Orders API	tool	look up orders and issue refunds
refund-gate	Refund approval	human_gate	a support lead approves every refund

Guardrails

id	kind	protects	what it does
pii-filter	input_filter	router-model, answer-model	Redacts card numbers and national ids before any model sees the text.
audit	audit_log	orders-tool	Every order lookup and refund is logged with the request id and the approver.

Failure modes

failure	component	mitigated by
A refund request is routed to the plain-answer path and the customer is told it is done.	Request router	refund-gate
An answer quotes a policy that has since changed.	Policy knowledge base	audit
Personal data reaches a model or a log.	Answer writer	pii-filter

Threats

OWASP	component	mitigated by	note
LLM01 Prompt Injection	Answer writer	pii-filter	customer text is untrusted input
LLM02 Sensitive Information Disclosure	Answer writer	pii-filter	personal data in prompts
LLM06 Excessive Agency	Orders API	refund-gate	write access to orders

Eval plan

metric	method	threshold
routing accuracy	200 labelled tickets, rerun on every prompt change	>= 0.95
refund-gate coverage	replay of refund requests; no refund without an approval record	100%

Observability and deployment

traces per request
cost per request
approval latency
refund approvals and rejections

container behind the support portal · The gate is a queue in the support tool, not a model.

Estimated cost

$0.0098 per request · $1477.50 / month at 5000 requests/day · ~6.63 s end to end

Prices as of 2026-10-01 · source · latency is assumed.

token counts per step are estimates written with the blueprint and approved by the owner, not measurements
steps run one after another, so parallel branches make the latency an overestimate
latency uses an assumed time to first token and output speed per model
30 days per month and the stated requests per day

Volume Tokens

Runnable starter

No starter is offered: no API data for langgraph from Radar yet.

Changelog

v1: seed