Implementation Guide | Financial Services

    Detecting and Routing Elder Financial Exploitation Signals in Customer Service AI Conversations

    Define signal classes from regulator-identified behavioral and transactional red flags, then adapt them to conversational cues. Map each class to a runtime action (allow, redact, block, or escalate for human approval) that is enforced in the interaction path. Route escalations to trained human reviewers, and retain a record of the policy in force, the action taken and the reviewer decision. As of November 2026, model risk expectations follow the April 2026 interagency guidance, which rescinds earlier model risk management guidance.

    Specify the evidence record

    • Conversation and interaction identifiers with timestamps
    • Signal class and confidence value, with the detection component version
    • Policy version in force and the threshold that applied
    • Action taken (allow, redact, block or escalate) and the tool calls permitted or denied
    • Reviewer identity, decision and decision time
    • Outcome or referral reference, such as a case or escalation number

    Start the taxonomy from regulator red flags

    Regulator and agency materials group elder financial exploitation indicators into two families. Behavioral indicators include third-party pressure, new "friends" or romantic interests pressing for financial decisions or account access, isolation by caregivers, and cues that suggest confusion. Transactional indicators include large or unusual withdrawals, sudden balance changes, unusual changes to wills or powers of attorney, and title transfers with no apparent reason. The interagency statement on elder financial exploitation (SR 24-8 / CA 24-6) describes training customer-facing staff to recognize both kinds, including when transactions are conducted through powers of attorney or other agents.

    These lists were not written for chat or voice transcripts, so the translation to conversational cues is your design decision. Document the rationale for each class and each threshold, and keep the link back to the underlying red flag visible. That link is what lets a compliance reviewer or examiner see why a cue is treated as a signal.

    • Third-party influence: Indications that someone else is directing, coaching or pressuring the customer, or is present on the interaction.
    • Urgency and secrecy: Pressure to act immediately or to keep a request from family or the institution. This is a conversational adaptation of the behavioral red flags.
    • Unusual requests: Large or atypical transfers, sudden account changes, and changes to powers of attorney or beneficiaries.
    • Confusion or undue influence: Cues that the customer does not understand the request or is acting under pressure.

    Design confidence and thresholds for this risk, not generic fraud

    The CFPB has noted that elder fraud indicators may differ from conventionally accepted suspicious-activity patterns, and it encourages institutions that use predictive analytics to explore additional elder-specific risk factors. Reusing generic fraud thresholds is therefore a weak default. Set thresholds per signal class and per requested action.

    The tradeoff is deliberate. A low-confidence cue combined with a high-risk request, such as an unusual transfer, can justify blocking completion and escalating. The same cue during a balance inquiry may justify only a scripted clarifying question. Co-occurring signals, for example third-party coaching together with secrecy cues, should raise the tier. Expect over-escalation at first, and size reviewer capacity accordingly, because a human reviewer is the safeguard against both missed signals and unnecessary friction. Record who approved each threshold and when, and re-approve when you change one.

    Map each signal tier to a runtime action

    The mapping below is a design pattern, not a regulatory requirement. Its purpose is that the agent's behavior after a signal is decided by policy in advance, not improvised by the model mid-conversation.

    Two constraints apply across every tier. First, the agent must not tip off a suspected exploiter, since a third party may be on the line or coaching the customer. Responses should use approved neutral scripts and should never state or imply that a suspicion or a SAR exists. Second, SAR existence is strictly confidential and generally limited to law enforcement and financial regulators and institutions, so the agent must never disclose it or reveal related information.

    • No signal: allow: Normal service continues under standard policy.
    • Weak or isolated cue: allow with scripted questions: The agent asks approved questions asking the customer to explain and confirm the request, a practice the CFPB describes for frontline staff. Sensitive account details are redacted where policy requires.
    • Sensitive disclosure request after a signal: redact or block: Balances, account identifiers and other sensitive data are withheld from responses until a reviewer clears the interaction.
    • High-risk request with a signal: block and escalate: The agent does not complete the transfer, account change or access change. Tool calls that would execute it are denied and routed for human approval.
    • Strong or combined signals: escalate with priority: The interaction goes to a trained reviewer at once. If the customer appears to be in immediate danger, follow the emergency procedure, which the CFPB describes as calling 911.

    Design the human review handoff

    Escalation should reach reviewers who are trained for this work. Both the CFPB and the interagency statement point to recurring, role-based training. The interagency statement also describes open lines of communication among the departments that research and respond to unusual activity, such as BSA compliance, fraud prevention and consumer protection. Route escalations so that these functions can see them, and assign an accountable owner in each.

    The handoff should carry the documented conversation details, because the CFPB recommends documenting specifics. Include the signals that fired, the requested action, what the agent said and what it was prevented from doing. The review playbook then covers the human decisions. The SAR decision stays with the institution's designated human process. Mandatory and voluntary scenarios should be handled distinctly, because voluntary SARs are permitted for exploitation below regulatory dollar thresholds, and FinCEN provides an elder financial exploitation category on the electronic SAR form. The playbook should also cover referral to Adult Protective Services and law enforcement, and trusted-contact outreach. The CFPB describes trusted-contact alerts as voluntary and lists example triggers: being unable to contact the account holder, concerns that persist after a discussion, confirming health status, or confirming the identity of a new guardian, trustee, executor or power of attorney holder.

    Treat detection and enforcement as separate components

    In April 2026 the OCC issued revised interagency model risk management guidance (OCC Bulletin 2026-13), which rescinds prior model risk management guidance. Cite the current guidance in your documentation, not the earlier issuances, and confirm the transition timing with your regulator. The guidance is risk-based and says practices suited to one institution may be unsuitable for another with a different risk profile or model purpose. It defines a model as a complex quantitative method, system or approach that applies statistical, economic or financial theories to process input data into quantitative estimates. Whether your detection component meets that definition is a determination your institution must make and record. Scale validation, monitoring and effective challenge to the result and to the component's role.

    Separating the detection component from the enforcement point makes this easier to defend. The detector produces a classification. The enforcement point applies a fixed, versioned policy to each interaction and tool call. Each can be tested and logged independently, and a change to one does not silently alter the other. Trussed operates only at the enforcement point. It enforces policy in the execution path on individual interactions and tool calls, with allow, redact, block and escalate for human approval, and it produces audit-ready evidence from that enforcement. It is not the detection model, and it does not provide transaction-level or workflow-level governance.

    From conversation signal to retained evidence

    Detect

    A detection component classifies conversational cues into defined signal classes with a confidence value.

    Enforce

    Policy applies allow, redact, block or escalate to each agent response and tool call.

    Review

    Trained reviewers receive the handoff and make the decision, including any SAR determination.

    Retain

    A durable record shows the policy in force, the action taken and the reviewer outcome.

    Review your enforcement and evidence design

    Walk through how runtime enforcement and audit-ready evidence can support your escalation policy for AI agent conversations.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.