How to Set Escalation Rules When AI Monitoring Finds a Problem

Why does a broker-dealer need written escalation rules for AI monitoring?

Because FINRA Rule 3110 looks past the review itself to what the firm does with the results, and written escalation rules are how a firm shows its AI incident response actually happens. The rule expects follow-up, and its evidence-of-review standard asks for the actions the firm took on significant regulatory issues along with a record that someone looked.

FINRA Rule 3110 requires a supervisory system reasonably designed to achieve compliance and procedures for reviewing correspondence and internal communications. Supplementary Material .06 lets a firm use risk-based review, but only with "surveillance and follow-up to ensure that such procedures are implemented and followed." Supplementary Material .07 says evidence of review must identify the reviewer, what was reviewed, the date, and "the actions taken by the member as a result of any significant regulatory issues identified during the review."

Regulatory Notice 24-09 reminded firms that these technology-neutral rules apply to generative AI. FINRA's 2026 Annual Regulatory Oversight Report lists ongoing monitoring of prompts, responses and outputs, and human-in-the-loop review, as effective practices. Monitoring that produces alerts with no defined next step leaves a gap an examiner can see.

What kinds of problems does AI monitoring find?

Most findings fall into four groups: policy misuse, communications issues, customer data exposure, and possible violations by a registered person. Each group needs a different owner and a different clock.

• Policy misuse: an unapproved AI tool, a personal account used for firm work, or an approved tool used for a task the policy does not allow, with no customer data involved.
• Communications issues: an AI-drafted email or post that may be promissory, unbalanced or missing required content, or that went out without required approval.
• Customer data exposure: account numbers, Social Security numbers or other customer information entered into a tool that is not approved to hold it.
• Possible violations: a prompt or output that suggests a registered person may have broken a securities rule, such as discussing an unapproved product or a trade outside the firm's systems.

Our earlier articles cover how to sample AI prompts for supervisory review and how to test whether the monitoring works. This one covers what the reviewer does next.

How should you tier AI incident response and escalation?

Use three tiers tied to those groups: supervisor handling, compliance review, and incident or regulatory response. Write down who owns each tier, what triggers it, and how fast it must move.

Tier 1: supervisor handling

Policy misuse with no customer data and no communications issue stays with the designated principal. The principal talks to the person, records the conversation, and confirms the behavior stopped. Repeat findings for the same person move to Tier 2.

Tier 2: compliance review

Communications issues, repeat policy misuse, and anything the principal is unsure about go to the CCO or a delegate. Compliance decides whether the communication needs correction, whether it should have been approved under the firm's communications procedures, and whether records were captured. If the AI output was sent to a customer, it is likely a business communication subject to SEC Rule 17a-4(b)(4) retention, and the record of the review should be kept with it.

Tier 3: incident or regulatory response

Customer data exposure and possible violations go to the CCO immediately and start the firm's formal processes. Set a same-day target for the first escalation, because the downstream deadlines are short.

When does a finding become a Reg S-P incident?

When customer information may have been accessed or used without authorization. Customer data in an unapproved AI tool should be routed to your Regulation S-P incident response program on the day it is found.

Regulation S-P requires a written incident response program to assess the nature and scope of an incident, contain and control it, and notify affected individuals where sensitive customer information was or is reasonably likely to have been accessed or used without authorization. Under paragraph (a)(4), notice is due as soon as practicable and no later than 30 days after the firm becomes aware of the incident. Larger firms had to comply by December 3, 2025 and smaller firms by June 3, 2026.

The 30 days run from when the firm becomes aware of the incident, so a finding that waits a week in a supervisor's queue spends time the firm may need for its investigation. Your escalation rule should name the trigger plainly: any customer account number, Social Security number or similar identifier in a tool not approved for customer data goes to incident response the same day.

When might a finding need a report to FINRA?

When the firm concludes, or reasonably should have concluded, that an associated person or the firm violated securities rules in a way FINRA expects to be reported. That decision belongs to compliance and counsel, and the escalation rule's job is to get it to them quickly.

FINRA Rule 4530(b) requires a member to report within 30 calendar days after it has concluded or reasonably should have concluded that an associated person or the member itself violated securities laws, rules or regulations. Supplementary Material .01 narrows what FINRA expects to see. For the firm, that means conduct with widespread or potential widespread impact, or conduct arising from a material failure of systems, policies or practices. For associated persons, it also includes conduct with significant monetary results or multiple instances of violative conduct. Supplementary Material .02 limits paragraph (b) to conclusions the member reaches on its own. Written customer complaints found in AI-related reviews also feed the quarterly statistics due under Rule 4530(d) by the 15th day of the month after the quarter.

The point for AI monitoring is simple: the clock under Rule 4530(b) runs from when the firm reasonably should have concluded, so a finding that sits unreviewed does not stop it.

What should each escalation record include?

Enough for an examiner to see the alert, the decision and the outcome without asking anyone. Rule 3110 .07 sets the floor: reviewer, item reviewed, date and actions taken.

• The alert or sample item, with the prompt and output if captured.
• The reviewer's name, the review date and the tier assigned.
• The reason for the tier, in one or two sentences.
• Each handoff, with names and timestamps.
• The outcome: coaching, correction, incident response, regulatory filing decision or no action, and why.
• Links to any related incident file or Rule 4530 analysis.

Keep these records with your other supervisory evidence, and include the escalation process in your annual Rule 3120 testing so the report shows the tiers worked in practice.

What should you do this quarter?

• Write the three tiers into your written supervisory procedures, with owners, triggers and target times.
• Name a backup for each owner so escalation does not stall during vacations or exams.
• Add a same-day trigger for customer data in unapproved AI tools that routes straight to your Reg S-P incident response program.
• Build a simple escalation log, or use your surveillance tool's case feature, with the fields listed above.
• Run one tabletop using a realistic AI finding and time how long each handoff takes.
• Add the escalation test to this year's Rule 3120 testing plan.

How can s90 help?

s90 provides prompt monitoring and AI guardrails for broker-dealers and dual-registered firms, and we set alert routing to match your escalation tiers. Customer data alerts can go straight to your incident response owner, with the prompt, user and timestamp attached, and we keep the alert history your Rule 3120 testing will need. If you want to pressure-test your current escalation path, we can run a tabletop with your compliance team.