AI support agents that need approval for refunds
Let support agents triage, draft replies and look up accounts, and route refunds, credits and account changes to a person. With NobleCloak, policy is set per tool, so a copilot can draft a reply for a human to post and a refund waits for an approver. Every decision is recorded.
100 scenarios in 5 areas
These are scenarios for how teams govern AI agents in Customer Support & Success. Some depend on capabilities still in development. Products named in a scenario are systems your team runs, named to make the example concrete. The policy applies to the AI agent's calls into them.
Ticket Triage & Responses20 examples
Copilot drafts, human posts
A support rep's Claude drafts a Zendesk response freely, but the public-reply tool is set to ask, so the rep confirms in-session before the message goes out under the company's name.
Ticket merges need a click
Merging duplicate Freshdesk tickets is set to ask, because a wrong merge buries one customer's issue inside another customer's thread.
Extra gate for strategic accounts
Replies to tickets from accounts flagged Strategic in Salesforce require confirmation, while replies to everyone else post directly.
Shared Slack channels gated
A support agent posting into a customer's Slack Connect channel must confirm each message, because that channel is the customer's own workspace.
Status page posts by support ops only
Only Support Ops can reach the Statuspage incident-update tool, and even then each post requires confirmation before it goes public.
Triage agent tags, never replies
The overnight Zendesk triage agent may read, tag, and route tickets, but NobleCloak denies the public-reply tool on its allow-list, so nothing reaches a customer until a person is on shift.
Macro edits limited to leads
Only the Support Lead role, synced from Okta, may call the Zendesk macro-update tool; for Tier 1 reps the tool is not entitled and never appears to their AI.
No bulk-close from an agent
An agent cleaning up a noisy Intercom backlog is denied the bulk-close tool outright, after a past script closed 3,000 open conversations from a single misread filter.
Tier 1 sees the knowledge base only
Tier 1 reps' assistants are entitled to Confluence and Guru search, while the Jira create-issue tool stays invisible until a ticket moves to Tier 2.
Tier 2 files the bug
Tier 2 engineers' AI may create Jira issues from a Zendesk ticket with no prompt, keeping escalation fast for the people who own it.
Translation through the governed proxy
A triage agent translating Portuguese tickets may reach the translation API only through NobleCloak's egress proxy, so ticket text never leaves through an unlisted endpoint.
Sub-agent can't change priority
When the triage agent spawns a sub-agent to summarize long threads, the sub-agent does not inherit the parent's set-priority tool and can only read.
AI draft cost per queue
Token and tool-call usage is attributed per Zendesk queue, so the support director can see that billing tickets cost three times more AI effort than login tickets.
Block replies containing card numbers
A reply that includes a full card PAN, pasted in by a frustrated customer and echoed back by the AI, is blocked before it posts.
Throttle the autonomous responder
The after-hours auto-responder is limited to 200 replies per hour, so a looping prompt can't flood the customer base overnight.
BPO staff are read-only
Outsourced BPO agents get a role that reads tickets and suggests drafts, and a per-user override removes reply rights from two contractors still in their training period.
QA sampling from the ledger
The QA manager pulls every AI-assisted reply from last week, with which agent drafted it and which rep confirmed it, straight from the NobleCloak audit ledger.
Patient messages: sort, don't advise
At a health system, an assistant may categorize patient portal messages in the EHR in-basket, but the send-to-patient tool is denied so clinical answers come only from licensed staff.
Read the full examplePreview the new Tier 2 policy
Before saving a revised Tier 2 policy, the support ops admin uses "Inspect as user" on a named Tier 2 rep to confirm Jira and Statuspage access look right.
Same rule in every AI app
A rep who switches from the in-house support bot to Claude Desktop hits the same ask-on-reply rule, because NobleCloak enforces it at the gateway rather than in each app.
Refunds, Credits & Account Changes20 examples
Refunds always confirmed
A support copilot may look up Stripe charges freely, but the refund tool is set to ask, so the rep confirms each refund in-session.
Refund ceiling by amount
Refunds under $100 run without a prompt, while anything larger requires confirmation, so small goodwill refunds don't slow the queue.
Goodwill gift cards capped
Gift cards sent through Tango Card come out of a $2,500 monthly budget for the support team's AI, and any single card over $50 needs approval.
Courtesy fee reversals at a credit union
A member-service assistant may reverse courtesy fees in the core banking system up to $35, and larger reversals need confirmation from the member-service rep.
Login email changes gated
Changing the login email on a customer account in Auth0 requires in-session confirmation, so a social-engineered request isn't completed by an agent on its own.
Insurance endorsements need a click
A service rep's assistant can prepare a policy endorsement in Guidewire PolicyCenter, but issuing it is set to ask, because it changes coverage and premium.
Billing approves large refunds
Refunds over $5,000 go to a named approver on the billing team.
Pause is fine, cancel asks
An agent may pause a Recurly subscription without a prompt, but cancellation requires confirmation because it ends revenue and may trigger data deletion.
Replacement shipping within budget
An agent buying replacement-shipment labels through EasyPost is held to an allowlisted vendor, a $40 per-call cap, and a $6,000 monthly budget.
Mailing address changes stay human
At a bank, the AI is denied the core banking change-address tool entirely, because an address change can be an early step in account-takeover fraud.
MFA resets reserved for identity support
Only the Identity Support role may call the customer-MFA reset tool; general reps' assistants don't see it at all.
Patient balance write-offs to revenue cycle
At a hospital, write-off tools in the patient billing system are entitled only to Revenue Cycle staff, so front-desk support AI cannot forgive a balance.
Senior rep exception
One senior billing specialist gets a per-user override that removes the refund prompt for that one person.
Month-end credit reconciliation
Finance reconciles every AI-issued credit and refund against the ledger, matching each to the rep who confirmed it and the ticket it came from.
Contract amendments stay with sales
Support assistants are denied the Salesforce CPQ amend-contract tool, so seat or price changes go through account management.
Card disputes belong to the disputes team
Chargeback and dispute tools at the card processor are entitled only to the Disputes role, keeping the general support AI out of Reg Z timelines.
Refund only after a verified return
The refund tool is callable only inside the returns workflow, after the warehouse has marked the item received in the order system.
Proration helper can't pay out
A sub-agent spawned to compute proration for a mid-cycle downgrade gets read-only billing tools and cannot issue the credit it calculates.
Plan changes for billing ops
Downgrades and upgrades in Chargebee are entitled to the Billing Ops role; Tier 1 reps' AI can explain plans but cannot change them.
Refund volume per agent per day
Each rep's assistant is limited to 25 refunds per day, so one compromised session can't drain the refund pool.
Customer Data & Privacy Requests20 examples
Erasure requires confirmation
The delete-customer-record tool used to fulfill erasure requests is set to ask, so the privacy analyst confirms each deletion.
Opt-out updates at a bank
A service assistant may set the GLBA information-sharing opt-out flag in the CRM, but the change is confirmed in-session because it governs what the bank can share with affiliates.
Do-not-sell flags are open, profile deletes ask
An agent may set a CCPA do-not-sell flag in Segment without a prompt, while deleting the profile asks first.
No exporting to personal email
Sending a customer data export by email is gated, and exports to non-company domains are denied.
DSAR compilation for privacy only
An agent assembling a data subject access request across Salesforce, Zendesk, and Snowflake is entitled only to the Privacy Team role, so general support can't run cross-system lookups.
No bulk deletes from Snowflake
Agents are denied the bulk DELETE tool on the Snowflake production connection; erasure goes through the per-record tool only.
Prod read denied, staging allowed
Support engineers' assistants may query the staging Postgres connection to reproduce bugs, while the production connection holding customer data is denied.
PHI stays inside the perimeter
At a digital health company, an assistant reading patient records in Epic cannot reach the internet except through the governed proxy, which allows only approved endpoints.
Erasure only inside the OneTrust workflow
The deletion tool is callable only as a step in the OneTrust DSAR workflow, after identity verification is recorded.
Scrub PII before the model sees it
Social Security and account numbers in ticket bodies are masked before ticket text is passed to the support copilot.
Show last four only
Account lookups return only the last four digits of an SSN or account number to the support AI, even though the underlying API returns the full value.
DSAR clock evidence
The privacy team shows a regulator that a request was fulfilled within the required window using ledger entries showing each lookup, export, and deletion with timestamps.
One BPO, many clients, no crossover
A support outsourcer serving several client companies runs each client as its own NobleCloak org, so an agent working one client's tickets can't see another client's customers.
Analytics deletion for privacy engineering
The Amplitude user-deletion API is invisible to everyone except the Privacy Engineering role.
Check what Tier 1 can see
The privacy lead uses "Inspect as user" on a new Tier 1 hire to confirm their AI cannot reach the DSAR export or Snowflake raw tables.
No disclosure before identity check
The full account-detail tool unlocks only after the identity-verification step in the same session has succeeded.
Call recordings for QA only
Talkdesk call-recording retrieval is entitled to the QA role, so a general rep's AI can't pull recordings of other agents' calls.
Backup restores denied
Customer requests to "restore my deleted data" can't be fulfilled by an agent; the database snapshot-restore tool is denied for all support roles.
Deletion blocked under legal hold
An erasure request against a customer whose account is on legal hold is refused, rather than relying on the analyst to notice the flag.
Privacy vendor's agent calls in
A third-party DSAR fulfillment vendor's agent may call our records agent over A2A only because it is on the peer allowlist, and only for the read-export tools.
Escalations & Customer Communications20 examples
Paging engineering asks first
A support agent may page the on-call engineer through PagerDuty, but the trigger tool is set to ask so reps confirm it's a real incident before waking someone at 3 a.m.
Mass incident email for comms only
Only the Customer Communications role may call the SendGrid bulk-send tool, and every send requires confirmation.
Outage notice routed to the comms lead
When a rep's assistant drafts an outage notice, the approval goes to the communications lead.
Claim status letters confirmed
At an insurer, an assistant generating claim status letters in the claims system must get confirmation before sending, since letters carry regulatory timelines.
Shared mailbox sends confirmed
Sending from the shared support@ Gmail mailbox requires confirmation, while sending from a rep's own mailbox does not.
Ask only for external recipients
Email sends to colleagues go through freely, but any email with an external recipient asks first.
Health system outreach confirmed
At a hospital, an assistant drafting patient notifications about a scheduling-system outage must confirm before sending through the patient messaging platform.
SMS within a Twilio budget
Outage SMS through Twilio is restricted to that allowlisted vendor with a $1,500 monthly budget and a per-call cap that stops a 50,000-recipient blast.
Exec escalations posted by managers
Only Support Managers' assistants may post to the #exec-escalations Slack channel, keeping executives from being pinged by every frustrated ticket.
No social replies from agents
Posting to X and LinkedIn is denied for every support agent, because a public reply on a complaint thread needs a human voice.
Bank complaints logged, not closed
A bank's service assistant may create a complaint record in the complaint management system, but closing it is denied so every complaint gets human review under UDAAP expectations.
Drafting without browsing the web
A reply-drafting agent can reach only the knowledge base and status page through the egress proxy, so it can't quote unvetted forum answers to customers.
Reseller agent escalates in
A reseller's support agent may escalate tickets into ours over A2A because it is a named peer, while unknown agents calling the same endpoint are refused.
Calling the processor's agent
Our billing agent may call the payment processor's support agent over A2A to check a payout delay, and no other external agent is allowed as a callee.
What did the AI tell customers
After an outage, the post-incident review reconstructs every customer-facing message the AI sent during the window from the audit ledger.
Bulk send for the incident window only
During a declared major incident, the comms agent gets bulk-send rights for four hours, then the grant expires on its own.
P1 belongs to escalation managers
Raising a Jira ticket to P1 is entitled only to the Escalation Manager role, preventing priority inflation from AI-drafted escalations.
Incident comms cost attributed
Tokens and tool calls spent drafting and sending incident communications are attributed to the incident team, so the cost of each outage includes its AI work.
The CEO's mailbox is off-limits
Agents are denied sending from executive mailboxes even when a delegate grants mailbox access, so "letter from the CEO" apologies are always sent by a person.
Callback scheduling is free
Booking a customer callback on a rep's Google Calendar runs without a prompt, because it's low-risk and saves the most back-and-forth.
Customer Success & Health20 examples
Stage changes ask first
When a CSM's assistant moves a Salesforce renewal opportunity to Closed Lost, the stage-change tool asks for confirmation.
Customer gifts with a cap
Sendoso gifts triggered by the CS agent have a $75 per-call cap and need confirmation above it.
Trial extensions confirmed
Extending a trial in the billing system asks first, because extensions delay revenue and are tracked against a quarterly limit.
Health scores read, renewals untouched
A CSM copilot may read Gainsight health scores and create CTAs, but the renewal-forecast edit tool is denied so forecasts stay owned by CS leadership.
Discount caps on renewal quotes
A renewal quote in Salesforce CPQ with a discount over 15% is denied for the CSM's AI and must go through deal desk.
Usage data through a read-only connection
The CS agent may query Snowflake only through the read-only analytics connection, never the production write connection.
QBR prep cost per CSM
Tokens spent generating QBR decks are attributed per CSM and per account, so CS leadership can see which books of business are expensive to serve.
Email drafter can't send
A churn-risk agent spawns a sub-agent to draft outreach emails; the sub-agent gets drafting tools only and cannot send.
Enrichment within budget
Contact enrichment through ZoomInfo is held to an allowlisted vendor and a $1,000 monthly budget for the CS team's agents.
Only managers reassign books
Reassigning account ownership in Salesforce is entitled to the CS Manager role only.
Feature flags stay with engineering
The LaunchDarkly flag-toggle tool is not entitled to any CS role, so a CSM's AI can't switch on a pre-release feature for a customer.
Health score overrides on record
Every manual health-score override made through the CSM's assistant is recorded with who made it and why, for the quarterly churn review.
Bank relationship managers see CRM, not balances
A relationship manager's copilot reads Salesforce Financial Services Cloud, but core-banking balance lookups are denied for that role.
Customer's agent asks for usage
A customer's own internal agent may query our CS agent over A2A for their usage stats, because their org is on the peer allowlist and only the usage-report tool is exposed.
Survey sends limited
The NPS agent can send Qualtrics surveys to each account no more than once a quarter.
Departing CSM loses reach
When a CSM gives notice, a per-user override removes their AI's export and email-send tools before their directory role changes.
No upsell during an open P1
Expansion outreach is blocked for any account with an open P1 ticket in Zendesk.
Call summaries for the account team
Gong call-summary tools are entitled to CSMs and account executives, not to support reps.
Cost to serve by account
AI usage across support, CS, and onboarding agents is rolled up per customer account, giving finance a real cost-to-serve figure at renewal.
Check a new CSM's setup
Before a new CSM's first day, the CS ops admin uses "Inspect as user" to confirm their AI sees Gainsight and Gong but not billing or CPQ.
| Action | Who | Decision |
|---|---|---|
| tag_ticket | triage agent | allowed |
| draft_reply | support copilot | allowed |
| issue_refund | support copilot | needs approval |
| change_mailing_address | any agent | denied |
| erase_customer_data | privacy team's assistant | needs approval |
| bulk_close_tickets | any agent | denied |
Approval requests go to the support lead in Slack or the web inbox (Teams and email coming soon).
Illustrative: names no package, import or API call
AI agents in Customer Support & Success: common questions
Can an AI agent issue a refund?
With approval, if you set it that way: the refund tool pauses the run until the support lead says yes. Letting small refunds through and routing large ones by amount is coming soon.
Can an agent reply to customers directly?
That is a per-tool choice. Many teams let the copilot draft and keep posting with a person, or mark the reply tool as needs approval for a given agent.
Who handles privacy and erasure requests?
Only the roles you scope those tools to, usually the privacy team, and an erasure can be held for approval. Inspect as user shows what each person's agent would get.
What does the support lead see afterward?
A record of each decision: which agent called which tool from which person's session, what policy decided (allowed, held for approval, or denied) and how any approval request was answered. It covers the calls policy blocked as well as the ones it allowed.
Set this policy up for your own agents.
Start with your own agents, your tools and the approvers you already have.
Other functions
Finance & Accounting
Payments, the close, treasury, budgets and payroll.
Software Development & Engineering
Code review, releases, production systems and on-call.
IT & Security Operations
Identity, devices, threat response, SaaS admin and cloud.
People & HR
Hiring, onboarding, compensation, employee relations and training.
Sales & Revenue
Outreach, the CRM, quotes, contracts and renewals.
Marketing & Communications
Content, campaigns, paid media, social and the website.
Legal, Risk & Compliance
Contracts, regulatory work, audit evidence, vendor risk and records.
Operations, Procurement & Facilities
Purchasing, vendors, inventory, facilities and projects.
Data, Analytics & AI Operations
Warehouse access, pipelines, agent-to-agent work, models and AI cost.

