A customer support agent should solve a defined set of intents using current approved knowledge and narrowly permissioned account tools. It must verify identity before protected data or actions, distinguish policy from recommendation, and escalate unsupported, sensitive, or high-impact cases. Measure correct resolution and customer harm, not just response speed or containment.
Who this is for: Support leaders, product teams, and engineers automating customer help while protecting accounts, policy, and escalation rights.
- Launch with bounded intents whose policies, tools, and escalation destinations are explicit.
- Enforce authentication, authorization, and transaction limits in code for every account operation.
- Evaluate policy fidelity, action correctness, escalation quality, and repeat contact by intent.
Select supportable intents
Start with requests that have stable evidence and a clear resolution path, such as order status or documented setup guidance. For each intent, define required information, approved knowledge, allowed actions, prohibited promises, and escalation triggers. Do not launch a general solve anything goal that quietly extends into legal, safety, fraud, or exceptional refund decisions.
Map the customer journey after the answer. A helpful explanation that leaves the account unchanged may not resolve the problem. Conversely, an action without explanation can confuse the customer. Define completion from verified system state and customer-facing confirmation, while keeping escalation available when the supported workflow does not fit.
Ground answers and protect identity
Retrieve current approved policy and product guidance with version and locale metadata. Cite or internally trace claims to the source section. If sources conflict or omit the case, say what is unknown and escalate. Do not let customer-provided text redefine policy, even when it quotes an apparent internal instruction.
Before exposing account data or changing anything, use the product's normal authentication and authorization. The model should not collect more sensitive information than needed or decide that a persuasive message proves identity. Tool code scopes each read and write to the authenticated account and returns only fields necessary for the current resolution.
Constrain actions and escalation
Represent support operations as narrow tools such as resend_receipt or cancel_pending_order, each with business-rule checks and idempotency. Preview consequential actions and ask the customer to confirm when appropriate. Financial exceptions, account access recovery, threats, safety concerns, and unsupported policy situations should move to trained people with useful context.
Escalation is an outcome, not a failure to hide. Pass the customer's stated need, authenticated context, attempted steps, relevant sources, tool outcomes, and unresolved question without forcing them to repeat everything. Avoid generated labels that bias the human agent against the customer. Preserve the original wording for sensitive complaints.
Evaluate service outcomes
Build tests by intent, policy boundary, customer wording, account state, and tool failure. Include requests that should be refused or escalated. Score answer support, identity handling, correct tool and arguments, verified completion, and quality of handoff. Repeat important cases because generated choices can vary.
In production, review repeat contact, reopened cases, incorrect actions, complaints, escalations, and human corrections alongside response time. Segment results so common easy questions do not hide harm in rare cases. Update the source pack and tests when policy changes. Check current platform APIs quarterly rather than depending on old tool behavior.
Handle a duplicate shipment request
A customer reports receiving two shipping notices and asks the agent to cancel the extra order.
- Authenticate the customer and retrieve only their recent orders and shipment states through a scoped read tool.
- Compare order identifiers and explain whether there are two orders or duplicate notices, citing current cancellation policy.
- If one eligible pending order is clearly identified, preview the exact order, items, and consequence before requesting confirmation.
- Call a narrow cancellation tool that rechecks status and idempotency, then report the actual result rather than a predicted success.
- Escalate with the evidence if both orders shipped, identity is uncertain, or policy does not cover the situation.
Support intent specification
Complete one specification for every intent the agent is allowed to handle.
- Intent boundary: qualifying requests, confusing neighbors, excluded cases, and completion definition.
- Evidence: approved sources, version filters, required account fields, and unknown behavior.
- Identity: authentication level, resource authorization, sensitive fields, and disclosure limits.
- Actions: tools, parameter limits, confirmation, idempotency, prohibited promises, and rollback.
- Escalation and tests: trigger, destination, handoff fields, expected outcomes, and monitored harms.
Common mistakes
- Optimizing for fewer human conversations while ignoring incorrect resolution and customers who return for the same issue.
- Letting the model infer account ownership or policy exceptions from the customer's persuasive wording.
- Giving one broad account-management tool permission to perform unrelated reads, credits, cancellations, and access changes.
Try one
A customer requests a refund outside the documented window and describes a serious hardship. How should the agent respond?
The agent should acknowledge the request, state the applicable policy accurately without inventing an exception, and avoid asking for unnecessary sensitive details. If an authorized exception process exists, it should collect only required facts and escalate to the designated reviewer with the customer's wording and account evidence. It must not promise approval or issue funds through an ordinary tool. Evaluation rewards respectful clarity and correct authority boundaries.
Sources
- OpenAI agents guideOfficial guidance for building agent workflows with current OpenAI platform capabilities.
- OpenAI safety best practicesOfficial recommendations for testing, constraints, and human oversight.