AI Elder-Exploitation Detection 2026: The Duty That Attaches the Moment the Model Flags
Fraud teams bought the model to catch scams against older customers. What they also bought is machine-generated suspicion — and suspicion is the trigger word in every vulnerable-adult reporting statute, hold rule and trusted-contact provision on the books.
The Detector Was the Easy Half
Behavioural models for elder financial exploitation are now a standard line item for banks, credit unions, broker-dealers, advisers and the fintechs that serve them. They are genuinely useful: the pattern of a lonely retiree wiring escalating sums to a new overseas payee, or a newly added joint owner draining an account within weeks of being added, is exactly the kind of signal that a model reads better than a branch employee seeing one transaction at a time.
The compliance problem is not that the model is wrong. It is that the model is productive. An institution that previously formed suspicion about a handful of customers a month now forms it about hundreds, and every legal framework in this area — state mandatory reporting for vulnerable adults, exploitation-specific suspicious activity filings, temporary-hold authority, trusted-contact outreach, safe-harbour protection for good-faith disclosure — is keyed to suspicion, belief or reasonable cause. Volume was the constraint that kept those duties manageable, and the model removed it.
Two failure shapes follow. The first is under-response: alerts pile up in an unowned queue, and when one of them turns out to be a real case that ran for another eight months, the institution has a documented record showing it knew. The second is over-response: someone freezes an account on a score alone, without the statutory notice or the human determination, and the customer — who was not being exploited — has a straightforward claim about their own money.
The Five-Rung Escalation Ladder
Every alert should have exactly one rung, assigned by a person, with the duty attached to the rung rather than to the score. Ladders that skip rungs are how institutions end up either sitting on knowledge or freezing accounts without authority.
The model saw an out-of-pattern transaction. Nobody has inferred exploitation or impairment.
What attaches: Ordinary fraud-review handling. No protective duty attaches. Do not label the file with capacity language at this stage — the vocabulary you choose here follows the customer through every later review.
A human reviewed the alert and identified recognised indicators: a new payee with urgency, secrecy, a third party present during instructions, sudden changes to beneficiaries or authorised users.
What attaches: This is where reasonable cause starts forming. Open a documented case, assign an owner, and start the clock your own procedures define. Consider trusted-contact outreach within the scope you are permitted to discuss.
A designated reviewer has concluded that a specific transaction or pattern reflects likely exploitation of a covered adult.
What attaches: Reporting obligations to the applicable state agency and any exploitation-specific filing obligation are now live, on the shortest applicable clock. Temporary-hold authority, where it exists, becomes available — with its notice, escalation and expiry conditions.
The concern is not a third-party actor but the customer's own ability to understand the transaction.
What attaches: The most legally delicate rung. You are not a clinician and should not record a diagnosis. Restrict to observation: what was said, what was repeated, what was not recognised. Escalate to the internal function that owns capacity questions, and involve the customer's designated contacts or agents rather than improvising.
A guardianship or conservatorship petition exists, or an agent under a power of attorney is asserting authority over the account.
What attaches: Authority now comes from a document or an order, not from your risk model. Verify the instrument, scope the powers granted, record the effective date, and reconcile it with any hold or restriction already in place. Note that court-appointed fiduciaries themselves are a recognised exploitation vector, so monitoring does not stop.
Who Owes What, and to Whom
The duties in this area do not sit in one place. They are spread across state vulnerable-adult statutes, financial-crime filing obligations, securities rules for firms with brokerage or advisory relationships, and the private-law duties owed to the customer under the account agreement. An alert can trigger several at once, on different clocks, running to different recipients:
- To a state protective agency. Most states make some categories of financial-institution personnel mandatory or permissive reporters of suspected exploitation of an elder or vulnerable adult, with short deadlines and immunity for good-faith reports. Whether your staff are covered is a state-by-state question, and multi-state institutions frequently apply the strictest rule everywhere rather than maintaining fifty workflows.
- To financial-crime regulators. Elder exploitation has its own reporting typology in suspicious-activity filing guidance, which means the narrative quality matters: a filing that says "model score 0.87" is far less useful, and far less defensible, than one describing the observed pattern.
- To the customer. The account agreement, the duty to honour instructions, and consumer-protection expectations all run toward the customer. Every protective step is simultaneously an interference with their autonomy, and the defensibility of the step rests on it being narrow, documented and time-bound.
- To designated contacts and agents. A trusted contact, an agent under a power of attorney and a court-appointed fiduciary are three different things with three different scopes. Treating them interchangeably is the most common privacy failure in this workflow.
- Internally, to supervision. Firms with supervisory obligations must be able to show that alerts were reviewed by qualified people under a written procedure, not closed in bulk to clear a backlog.
Design Rules That Keep the Model Useful
These are the choices that determine whether the detector reduces your exposure or manufactures it. None of them require turning the model off.
Separate the exploitation signal from the capacity signal
They imply different duties, different recipients and different remedies, and one is about a third party's conduct while the other is about your own customer's mind. A single blended risk score forces reviewers to guess which one they are looking at, and pushes capacity language into files where only conduct was observed.
Make every alert closeable only with a recorded human disposition
The disposition field is the artefact that shows a duty was discharged rather than ignored. Bulk-closing is the single most damaging pattern available in this workflow because it converts a defensible triage decision into documented inattention.
Write the alert taxonomy for a courtroom, not a dashboard
Labels travel. 'Possible cognitive decline' is a clinical-sounding conclusion from a system that never examined anyone; 'repeated same-day duplicate transfer instructions' is an observation you can stand behind. Rename the categories before the record grows, because retrofitting language across historical alerts is not possible.
Keep protective interventions transaction-scoped and expiring
A hold on a specific transfer with a stated basis and an expiry date is defensible. A standing downgrade of the customer's account privileges based on an inferred condition is a status-based decision about a perceived impairment, which is exactly the shape of claim you do not want.
Decide retention deliberately
Impairment-adjacent scores are long-lived, sensitive, and subpoena-attractive in later capacity or inheritance disputes. Choose how long they live and why, rather than inheriting a vendor default that keeps everything forever.
Test for false-positive concentration by age and geography
A detector aimed at older customers will, by construction, flag older customers — but if it also concentrates on particular branches, languages or transaction rails, you are looking at a proxy problem that will surface as a fair-treatment issue rather than a fraud one.
Frequently Asked Questions
Our vendor says the model is 'decision support only.' Does that limit our duty?
It limits the vendor's exposure, not yours. 'Decision support' is a description of where the model sits in the workflow, and it is a useful description — a human making the determination is exactly the design you want. But the duty in this area attaches to what the institution knows and believes, and a support tool that reliably surfaces exploitation indicators to a reviewer is a mechanism for the institution to acquire knowledge. The phrase becomes actively misleading when it is used to justify not staffing the review, because at that point the tool is neither supporting a decision nor making one; it is generating a record of unexamined suspicion. Treat the label as a statement about who decides, and then make sure someone actually does.
Can we use the model's output as the reason in a suspicious activity narrative?
Use it as context, never as the substance. A narrative built around a score tells the reader nothing they can act on and signals that no human analysed the case. The useful narrative describes the observed pattern in plain terms: the timing and amounts, the relationship of the payee to the customer, when it started, what changed shortly before, what the customer said when asked, and who else was present or involved. Mentioning that the case originated from an automated monitoring alert is fine and often helpful for explaining how it was detected. What you want to avoid is a filing where removing the score would leave nothing behind.
The suspected exploiter is the customer's agent under a power of attorney. What changes?
Almost everything about who you can talk to, and nothing about your duty to report. An agent under a power of attorney has authority to transact but owes fiduciary duties to the principal, and abuse by an agent is one of the most common exploitation patterns precisely because the authority is genuine. Practically: your ordinary escalation path — calling the person on file who handles the account — routes directly to the suspected exploiter, so the outreach plan has to be built around the customer themselves, the designated trusted contact if any, and the protective agency. Documenting the powers actually granted matters too, since agents frequently assert authority the instrument does not contain, and transactions outside its scope are a separate and cleaner problem to act on.
How do we handle a customer who insists the transfers are voluntary?
Take it seriously and keep the record, because both parts matter. Adults are entitled to make decisions others consider unwise, including generous or foolish ones, and an institution that overrides that on suspicion alone is on weak ground. But scam victims routinely and sincerely defend the transfers while the scam is running, often having been coached on what to say — so the customer's assurance resolves the autonomy question, not the exploitation question. The workable posture is to document the conversation verbatim rather than summarising it, apply the protective steps you have actual authority for, make the required reports, and let the agency with investigative power investigate. Substituting your judgment for the customer's is not one of your options; failing to report because they reassured you is not one either.
We are a fintech, not a bank. Do these frameworks reach us?
Partly, and the gaps are the dangerous part. State vulnerable-adult reporting statutes are often drafted around enumerated categories of covered persons, and a non-bank platform may sit outside the enumeration in some states and inside it in others. Filing obligations follow your regulatory status and your partner-bank arrangements, and the partner agreement itself usually pushes monitoring and escalation duties onto you contractually regardless of whether the statute does. The result is that fintechs frequently have the detection capability, contractual duties to escalate, and no clear statutory safe harbour for good-faith disclosure — which is the worst combination, because the immunity provisions that protect banks for reporting may not be written to cover you. Map that exposure with counsel before you turn the alerts on, not after the first case.
The One-Hour Version of This Audit
Pull the last ninety days of alerts from your elder-exploitation or vulnerable-customer model and answer four questions: how many were closed without a recorded human disposition, how many used clinical or capacity language in a file where only transactions were observed, how many resulted in a restriction that has no expiry date, and how many were escalated to a designated contact whose permission scope was never verified.
Every count above zero is a duty either unmet or exceeded — and both directions are findings.