Wealth Management AI Intake: A Governance-First Design Guide

The phrase “compliant AI intake” sounds reassuring, but it collapses several decisions that belong to the firm. A vendor can describe security controls, configurable guardrails, or product capabilities. It cannot decide which rules apply to a particular investment adviser, broker-dealer, dual registrant, state-registered firm, communication, or customer journey. That determination belongs with the firm’s compliance and legal functions.

The useful question is narrower: can the firm define an administrative first-contact job that technology may perform under written controls? For many firms, that job may include acknowledging an inquiry, collecting limited routing information, explaining an approved process, offering an allowed calendar, creating a record, and escalating the conversation. Advice, recommendations, performance discussion, account instructions, and other protected decisions remain outside that job unless the firm has separately approved them.

Begin with the firm’s regulatory perimeter

A wealth-management label does not identify one regulatory model. An SEC-registered investment adviser, state-registered adviser, broker-dealer, bank, insurance organization, family office, and unaffiliated marketing firm may face different duties. The same sentence can also be treated differently depending on who says it, to whom, in what context, and whether it offers advisory services, responds to a current client, or recommends a securities transaction.

FINRA’s Artificial Intelligence topic page states that its technology-neutral rules and securities laws continue to apply when member firms use AI. The page also tells firms to review applicable laws, rules, and regulations when considering new technology. That is the correct starting posture: identify the firm, the communication, the data, and the responsible rule owner before approving a use case.

Do not ask whether AI may handle intake in the abstract. Ask which approved administrative actions this firm will permit in this exact customer journey.

Write the intake job before choosing the technology

A safe project does not begin with a voice demo. It begins with a written job. The document should describe the incoming customer moments, the information required for routing, the statements the system may make, the decisions it may not make, and the person or queue responsible for every exception.

A bounded administrative job may include

• Identifying whether the caller is a prospective client, current client, referral source, job candidate, vendor, or other contact.
• Collecting contact details and a broad reason for the inquiry when the firm has approved those fields.
• Explaining published services, locations, office hours, and the firm’s approved introductory process.
• Offering a calendar or callback request that the firm has approved for that inquiry type.
• Creating a structured record and notifying the responsible team without interpreting the financial facts.

The written boundary should address

• Questions about securities, investments, allocation, performance, tax treatment, legal matters, or the suitability of a strategy.
• Statements that could imply the firm has accepted a person as a client or agreed to provide a service.
• Existing-client instructions, complaints, account changes, trades, money movement, or urgent security concerns.
• Requests involving confidential records, identity data, credentials, account numbers, or other information the channel should not collect.
• Unclear intent, emotional distress, accessibility needs, language limitations, threats, or any situation requiring human judgment.

Treat every AI response as a firm communication

A polished voice does not reduce the firm’s responsibility for what is communicated. FINRA’s Rule 2210 frequently asked questions state that member firms are responsible for AI-created communications they distribute or make available. The same guidance explains that chatbot communications may fall into different communication categories and must be supervised under the applicable rules.

This means a project needs more than a list of forbidden words. The firm should approve the source material, message classes, disclosures, escalation language, response ownership, retention path, and change process. It should also decide whether a response may vary freely, use an approved library, or require a human before it is sent.

Separate qualification from professional judgment

Firms often want to protect advisor time by asking about service need, investable-asset range, location, timing, or current relationship. Those questions may be useful for routing, but the firm should decide whether each one is necessary, how it is worded, what happens when a person declines, and whether the answer changes access to a professional.

An intake system should not declare that a prospect is a perfect fit, promise that a named advisor can solve the problem, or turn an asset range into an implied recommendation. A more disciplined path records the answer, applies the firm’s approved administrative rule, and explains the next step without judging the person’s worth or financial situation.

A respectful routing design

1. Explain why the question is being asked. Connect the field to routing or preparation rather than making the person prove status.
2. Use the least sensitive useful answer. A broad range may be sufficient when a precise value is unnecessary.
3. Provide a decline path. Let the person continue to a human or alternative route when the firm’s policy requires one.
4. Avoid invented alternatives. Do not promise a partner firm, lower-minimum product, or referral unless that option is real and approved.
5. Preserve the reason for the route. The receiving person should see what was asked, what was answered, and which rule selected the next step.

Design privacy and data handling before launch

A financial-services intake path can touch contact information, financial context, call recordings, transcripts, and identity clues. The safest form is not the one that gathers the richest profile. It is the one that gathers the minimum information required for the approved administrative decision and sends it only to approved destinations.

The SEC’s Regulation S-P amendments address safeguards, incident response policies and procedures, service-provider oversight, records, and notification for covered institutions. Whether and how those requirements apply is a firm-specific determination, but the amendments make one design principle unavoidable: data flow and vendor oversight belong in the approval process, not in a footnote after launch.

The data map should show

• Which channel receives the inquiry and whether the person is told that automation or recording is in use where required.
• Which fields, audio, transcripts, summaries, metadata, and attachments are created.
• Where each item is processed, stored, retained, reviewed, exported, and deleted.
• Which employees, vendors, subprocessors, and integrations can access the information.
• What happens during an outage, suspected incident, deletion request, complaint, or vendor change.

Build supervision, testing, and monitoring into the system

FINRA’s 2026 GenAI regulatory oversight report identifies governance, testing, monitoring, documentation, privacy, integrity, reliability, accuracy, and human review as considerations member firms are using around GenAI. A one-time script approval does not address model changes, source-content changes, routing failures, or unexpected conversations.

A practical control cycle

1. Approve. Name the business owner, compliance reviewer, technology owner, and final approver for the use case.
2. Test. Use ordinary questions, prohibited questions, ambiguous requests, current-client instructions, complaints, privacy-sensitive prompts, interruptions, and system failures.
3. Launch narrowly. Limit channels, hours, audiences, or actions until the firm has evidence that the controls work in real conversations.
4. Monitor. Review a defined sample, every exception, every complaint, failed handoffs, changed outputs, and unauthorized information collection.
5. Change deliberately. Treat new scripts, services, models, data destinations, calendars, and routing rules as governed changes.

Keep public claims inside the same control system

The website and intake system should not make claims the firm cannot substantiate. The SEC’s Investment Adviser Marketing compliance guide describes general prohibitions for covered adviser advertisements, including untrue or misleading statements, material claims without a reasonable basis, and discussions of benefits without fair and balanced treatment of material risks or limitations.

The rule’s application depends on the communication and the adviser, so the firm should obtain its own review. Operationally, the lesson is simple: approved service descriptions, proof, testimonials, performance language, and disclosures should come from governed source material. The intake system should not improvise a stronger version of the website’s promise.

The wealth-management systems page shows how website trust, prospect intake, records, and human handoff can be considered together. Firms evaluating the operating layer can also review AI intake systems without assuming that one configuration fits every firm.

Measure movement, not hypothetical revenue

Do not build the business case from an invented close-rate lift or a universal value for advisor time. Use the firm’s own inquiry records. Establish how many prospective-client contacts arrive, how many receive a useful response, how many require rework, how often senior professionals screen obvious administrative mismatches, how many handoffs fail, and how many communications require correction.

A credible review compares the same measures before and after one defined change. It also tracks control outcomes: prohibited-topic attempts, human escalations, incomplete records, inappropriate data collection, complaints, recording failures, and exceptions resolved within the firm’s policy.

Use a launch packet that a reviewer can inspect

Before activation, assemble one packet containing the customer journey, approved questions, prohibited topics, source material, disclosures, data map, vendor review, retention instructions, role ownership, escalation paths, test cases, acceptance results, monitoring schedule, change log, and shutdown procedure. If a reviewer cannot understand what the system does from the packet, the system is not ready.

The packet should state that the firm’s compliance and legal professionals make the final determination. Technology and operations teams can document the behavior; they should not certify the legal conclusion. A Systems Review can map the operating path, while the firm retains responsibility for its approvals.

The same discipline should apply to public-facing AI claims. The Search and AI Readiness guide explains how clear services, named expertise, genuine proof, and structured information can improve machine understanding without promising a recommendation or ranking.