How to Require Approval Before an AI Agent Changes a Customer's Address or Phone Number
Declare address and phone number as gated fields on the agent's write tool, hold every call that touches them in the execution path, and route the proposed change to an approver who is independent of the requestor. Allow the call only on approval, block it on denial or silence, and retain the requester, prior and proposed values, approver, decision, time and policy applied as audit evidence.
Step 2: Place enforcement in the execution path
- 1
Intercept
The enforcement layer receives the tool call before it executes and evaluates it against policy.
- 2
Match
A call that writes to a gated field is held. Other calls follow their own allow, redact or block rules.
- 3
Escalate
A request containing the payload above is sent to the approver role defined in policy.
- 4
Decide
The call is released only on explicit approval. Denial or no response blocks it.
- 5
Record
The decision and the resulting allow or block are written as evidence at the moment they occur.
Step 4: Decide what evidence to retain
- Requestor identity, including the agent and the initiating person or session
- Originating input that led to the proposed change
- Prior and proposed values, handled according to your data protection rules
- Approver identity and role, with the decision and any stated rationale
- Timestamps for request, decision and execution or block
- Policy and version applied, and the resulting allow or block outcome
Why approval has to apply at the tool call
Address and phone number look like routine record maintenance, but they feed other processes. Mailed notices follow the address. Verification steps and one-time passcodes often follow the phone number. A wrong value, whether from a mistake or a deliberate attempt to deceive the agent, has effects well beyond the record itself.
Authoritative agent security guidance points to the same control. OWASP's guidance on excessive agency recommends human-in-the-loop control so that a person approves high-impact actions before they are taken, and its example places the approval routine inside the extension that performs the action. That placement matters. Review of logs after the fact tells you a change happened. An approval gate at the call decides whether it happens.
The sources do not name address or phone fields specifically. Applying the pattern to them is a design decision: you classify the action by risk and reversibility, then gate it. That is the walkthrough below.
Step 1: Define which fields and actions require approval
Start with the write action, not the agent. OWASP advises that the reversibility class of an action is declared on the action rather than inferred by the agent at run time. Read-only inspection can run unattended. Changes that are hard to undo externally require approval. Apply that to your tool definitions.
Then make access default-deny for writes. The agent's identity should have read access to customer records by default, and write access to gated fields should exist only through the approval path. OWASP also recommends enforcing this through the permissions of the identity the tool uses, and applying least privilege at the action level, not just the credential level.
- List gated fields: Name the fields explicitly, here postal address and phone number, and add others such as email only by deliberate decision.
- Declare risk on the tool: Mark the update action with its risk and reversibility class in the tool definition, not in the prompt.
- Scope credentials: Use separate, minimal credentials per integrated system so the agent cannot write to gated fields outside the gate.
- Treat untrusted input as a trigger: Content supplied by the customer or retrieved from outside the trusted prompt should never authorize a gated change on its own.
- Rate limit attempts: Limit gated write attempts per period so repeated tries are visible and bounded.
Step 3: Specify approvers, denials and timeouts
Define approvers as roles, not named people, in the same way NIST SP 800-53 refers to organization-defined personnel or roles. NIST SP 800-128 adds that best practice is review by at least one authorized individual who is independent of the requestor, which preserves separation of duties. For an agent, the requestor is the agent plus the person or session that initiated the task, so the approver should be neither. In a customer service setting that usually means a supervisor or identity-verification role rather than the same agent handling the conversation.
Timeouts and denials are design decisions that the standards do not dictate, but the safe default is clear: fail closed. If no approver responds within a defined window, the call is blocked and the request expires rather than waiting open indefinitely. A denial should return a clear result to the agent so it can tell the customer the change is pending review, and should not invite a retry with different wording. Expired and denied requests are evidence too, and repeated ones for the same account are a signal worth monitoring.
Electronic routing is acceptable. SP 800-128 notes that change requests can be tracked and routed through automated tools that allow electronic approvals, provided approvals, disapprovals and rationale are recorded.
Tradeoffs, tuning and where Trussed fits
Approval adds latency and work for approvers. Keep the gated field list short and explicit so approvers see only changes that matter, and make the request payload complete so each decision takes seconds rather than a follow-up investigation. Widening the list to every field invites approvers to confirm without reading. Start narrow, review block and approval patterns, and extend deliberately.
Separate obligations should stay separate. If you are a card issuer, 17 CFR Part 248, as of November 2026, requires reasonable written policies to assess the validity of a change of address when a replacement or additional card is requested soon after, and the card may not be issued until that assessment is done. That is a distinct requirement and does not mandate approval of every contact-detail change.
Trussed enforces policy in the execution path of agents and tool calls, with actions that include allow, redact, block, rate limit and escalate for human approval, scoped to individual interactions and tool calls. Enforced at runtime, not reviewed after the fact, a gated write is held until the decision returns, and the enforcement itself generates audit-ready evidence mapped to frameworks including HIPAA, NYDFS and NIST. The pattern in this guide works regardless of which tooling you use to implement it.
Approval gate for a gated field write
Agent proposes
The agent calls the update tool with a new address or phone number.
Policy matches
The call is matched to a gated field rule before it reaches the CRM.
Approver decides
An authorized human, independent of the requestor, approves or denies.
Enforce
The call is allowed or blocked at runtime.
Evidence retained
Requester, values, approver, decision and policy are recorded.
Put an approval gate on high-risk tool calls
See how policy enforcement and escalation for human approval work in the path of an agent's tool calls.
Explore Agent and Tool-Call Enforcement