Skip to main content

Securities Compliance AI: MiFID II and Dodd-Frank Policy Mapping

The safest way to talk about securities-compliance AI is to avoid claiming that the gateway has a built-in MiFID II or Dodd-Frank switch. It does not. What Keeptrusts gives you is a set of enforceable runtime controls that can be mapped to the parts of a securities workflow you actually need to govern: who used the assistant, which prompts carried sensitive client or market information, which outputs looked like unauthorized recommendation language, which routes required human review, and how the team exported evidence for later examination.

That makes the platform useful for securities firms, even though it is not a trade-surveillance archive or a regulatory recordkeeping system by itself. The operational win comes from using current policies such as rbac, pii-detector, mnpi-filter, financial-compliance, human-oversight, and audit-logger in a disciplined way. Those controls help firms map AI use to their own supervisory procedures instead of relying on vague internal guidance about how staff should prompt the model.

Use this page when​

  • You are using AI for securities research support, supervisory review, trade commentary drafting, sales-assist workflows, or client communication preparation.
  • You need a realistic mapping between gateway controls and the governance requirements your firm already carries under securities regulation.
  • You want the rollout to align with Finance, Investment Banking & Capital Markets, and Pass Compliance Audits.

Primary audience​

  • Primary: Technical Leaders
  • Secondary: Technical Engineers, AI Agents

The problem​

Securities workflows are full of text that looks operational but carries regulatory weight. A sales assistant drafts client commentary. A supervisor asks AI to summarize exceptions for a desk review. Research support staff use a model to turn source notes into a briefing. Trade surveillance teams write narratives about unusual activity. In each case, the AI output can shape client communications, supervisory decisions, or downstream records even if the model never executes a trade or makes a formal compliance determination.

There are three recurring risks. The first is restricted-information handling. If prompts or outputs contain non-public market information, issuer-sensitive material, or confidential desk commentary, the assistant has crossed a line regardless of how useful the text was. MNPI Filter helps on the output side, but prompt-side control still depends on role attribution, input minimization, and local restricted-language blocking.

The second risk is recommendation language. Financial Compliance is useful when research, sales, or client-service workflows must avoid advice-style or guaranteed-outcome phrasing. It does not perform a full securities suitability analysis. What it does provide is a deterministic way to prevent clearly unacceptable language from leaving the route unchecked.

The third risk is weak reviewability. Securities programs eventually need to explain who used the assistant, which control route applied, whether human review happened, and what evidence was exported. That is why human-oversight and audit-logger matter. The route has to be governable after the fact, not just careful in the moment.

The solution​

The practical mapping starts with attribution. rbac is how the firm turns a generic assistant into a role-bound workflow. That matters because supervisory, sales, research, and operations roles should not share one undifferentiated route.

Next, control the content boundary. pii-detector helps minimize client and account identifiers. mnpi-filter adds an output-side backstop for non-public market-sensitive phrases. financial-compliance constrains recommendation or guarantee-like language on routes that could influence clients or registered representatives.

Then separate reviewer-required routes from ordinary assistance. human-oversight belongs on drafts that are expected to stop for a supervisor or compliance reviewer. It is not a universal default, but it is the right control where the organization wants a deliberate sign-off point.

Finally, keep the evidence path alive with audit-logger and export workflows. That is the bridge between policy execution and the supervisory procedures your securities program already maintains.

Implementation​

This example is a conservative securities-review route where attribution, restricted-language control, and supervisor review are all required.

pack:
  name: securities-review-lane
  version: 1.0.0
  enabled: true

policies:
  chain:
    - rbac
    - pii-detector
    - mnpi-filter
    - financial-compliance
    - human-oversight
    - audit-logger

policy:
  rbac:
    deny_if_missing:
      - X-User-ID
      - X-User-Role
      - X-Supervisory-Lane
    require_auth: true

  pii-detector:
    action: redact
    detect_patterns:
      - 'CLIENT-[A-Z0-9]{8}'
      - 'ACCOUNT-[0-9]{8,12}'
    redaction:
      marker_format: label
      include_metadata: true

  mnpi-filter:
    detect_patterns:
      - merger not public
      - earnings before announcement
      - non-public trading activity

  financial-compliance:
    blocked_patterns:
      - guaranteed return
      - you should buy
      - you should sell
    required_disclaimers:
      - This is not financial advice.

  human-oversight:
    action: escalate

  audit-logger: {}

The important thing to notice is what this route does not claim. It does not claim to encode MiFID II or Dodd-Frank directly. It gives the firm a place to enforce the local supervisory controls that those regimes make important: attributable usage, controlled sensitive language, hard review checkpoints, and evidence export.

kt policy lint --file ./securities-review-lane.yaml
kt gateway run --policy-config ./securities-review-lane.yaml --port 41002
kt export create --format json --filter "policy=mnpi-filter,financial-compliance,human-oversight,audit-logger"

That gives compliance and supervision teams a concrete way to review whether the route behaves like the firm expects before it supports live desks or client communications.

Results and impact​

The result is not automatic securities compliance. The result is a cleaner mapping between AI use and the supervisory controls the firm already understands.

  • Role attribution becomes explicit because rbac separates supervisory and user lanes.
  • Client identifiers are less likely to leak into provider calls because pii-detector minimizes them.
  • Restricted and recommendation-like language are easier to block because mnpi-filter and financial-compliance enforce explicit checks.
  • Reviewer accountability becomes more defensible because human-oversight and audit-logger keep the route reviewable.

For securities teams, that often matters more than any single model improvement. A governable assistant is usually more valuable than a slightly more fluent one.

Key takeaways​

Next steps​