AI-Powered Marketing Compliance Automation System

remitly · Finance · year n/a · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Financial-services marketing compliance

Remitly needed to reduce slow, repetitive legal review cycles for marketing content across regulated money-transfer markets while preserving compliance with country-specific rules and keeping legal review as the final approval step.

Constraints:

• Remitly operates in 170 countries with different money-transmission regulations
• customer communications span more than 18 languages and multiple channels
• marketing content can be text, image, video, or audio, though the first release handled English text only
• manual legal feedback was scattered across Google Docs, Slides, Slack, and meetings
• compliance misses are higher risk than false positives
• marketers still need judgment and legal retains final approval

Architecture

Remitly built a self-service GenAI compliance checker for English marketing text. Marketers upload content, and the system analyzes it against internal compliance guidelines derived from regulations and best practices. The analysis identifies the exact text that may violate a guideline, names the violated guideline, and suggests a compliant revision. Marketers iterate on the content until issues are resolved, then submit to legal for approval. Under the hood, the system combines Confluence/doc guideline sources, engineered prompts, and an AI analysis layer. A structured training/evaluation corpus was created from previously unstructured legal feedback, and the team optimized precision/recall through model iterations and beta usability feedback.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  historical-feedback-sources["historical-feedback-sources"] --> training-evaluation-corpus["training-evaluation-corpus"]
  training-evaluation-corpus --> evaluation-loop["evaluation-loop"]
  compliance-guidelines["compliance-guidelines"] --> prompt-guidance-layer["prompt-guidance-layer"]
  compliance-guidelines --> ai-analysis-engine["ai-analysis-engine"]
  prompt-guidance-layer --> ai-analysis-engine
  marketing-content-upload["marketing-content-upload"] --> ai-analysis-engine
  ai-analysis-engine --> structured-guidance-output["structured-guidance-output"]
  structured-guidance-output --> marketer-review-loop["marketer-review-loop"]
  marketer-review-loop --> legal-approval["legal-approval"]
  evaluation-loop --> ai-analysis-engine
  future-native-tool-integrations["future-native-tool-integrations"] --> marketing-content-upload

Components

Data flow

A marketer uploads English text content. The system combines engineered prompts with internal compliance guidelines stored in Confluence pages and documents. The AI analysis engine checks claims, disclaimers, IP, spelling, grammar, and other compliance categories. It returns structured guidance: offending text, violated guideline, and a suggested change. The marketer edits and reruns checks until confident, then sends the content to legal for final approval. Training and evaluation data come from structured pairs of marketing content and historical legal feedback collected from prior review channels.

Why this architecture

Remitly scoped the first version to English text because the full marketing surface includes audio, images, video, and 18+ languages. A self-service checker addressed the biggest workflow pain: marketers could resolve likely issues before legal submission rather than waiting through multiple review rounds. The system returns structured guidance instead of a binary pass/fail so marketers can understand and fix issues. Historical legal feedback had to be transformed into a structured corpus because the old process stored expertise in scattered comments, Slack messages, and meetings. Legal approval remains in place because compliance accountability cannot be delegated entirely to the model.

Models used: generative AI · large language models

Scale

• Remitly operates in 170 countries.
  “We are in 170 countries.”transcript, stated
• Remitly reported around 8 million active customers.
  “We have around 8 million active customers”transcript, stated
• Remitly communicates with customers in over 18 languages.
  “we communicate to our customers in over 18 languages.”transcript, stated
• The manual compliance workflow usually repeated two to four review cycles.
  “this cycle goes on for like you know two to four times.”transcript, stated
• The first implementation took 15 development days spread over 3 months.
  “It took us 15 days to develop this spread over 3 months.”transcript, stated
• The team reported 95% accuracy.
  “we achieved 95% accuracy”transcript, stated
• The team reported around 97% issue identification.
  “around 97% uh of identification of the issues.”transcript, stated
• Future multimodal expansion needs to handle text, audio, image, and video.
  “we want to expand to audio, image and video.”transcript, stated

Alternatives considered (roads not taken)

Continue manual legal review loops

Rejected because: The two-to-four-cycle process slowed communications and forced marketers to context-switch back to prior projects.

Start with all content formats and languages

Rejected because: The team deliberately scoped the first version to English text before expanding to audio, images, video, and more languages.

Use a separate compliance tool forever

Rejected because: The roadmap is to integrate checks into native creative tools such as Adobe and Figma to remove tool context switching.

Let AI replace legal approval

Rejected because: The target workflow keeps legal review as a final quality check even after marketer self-service iteration.

Patterns

safety-policy-enforcement (safety)

The tool checks marketing content against internal compliance guidelines distilled from regulations and best practices.

structured-output-contract (contracts)

The model output is structured as offending text, violated guideline, and suggested change.

human-in-the-loop-gate (control)

Marketers apply judgment to AI guidance, and legal remains the final approval gate.

eval-driven-development (evaluation)

The team used precision, recall, beta usability feedback, and multiple model iterations before rollout.

data-quality-loop (learning)

Unstructured legal comments were converted into content-feedback pairs for training and evaluation.

confidence-thresholding (evidence)

The team explicitly tuned the precision/recall balance for a compliance use case where recall matters strongly.

Tradeoffs

Start with English text

Chose narrow text-only MVP over all languages and multimodal marketing assets on day one.

Rationale: The full surface was too broad for the first release, and images/video introduce extra checks such as disclaimer font size.

Use self-service marketer checks before legal review

Chose marketer iteration with AI guidance over legal-only review after creative completion.

Rationale: Self-service checks reduce the back-and-forth and context switching before final legal approval.

Keep legal as final quality gate

Chose AI-assisted precheck plus legal approval over full AI approval of marketing compliance.

Rationale: The desired state is a lighter legal review, not removal of accountability from legal.

Invest in structured examples

Chose content-feedback training/evaluation corpus over training directly from scattered historical comments.

Rationale: The model and evaluator needed explicit examples of content and corresponding legal guideline feedback.

Gotchas (what bit them)

The most time-consuming part was not model coding but training-data creation.

Root cause: Legal feedback lived in Google Docs, Slides, Slack, and meetings rather than structured examples.

Resolution: The team converted historical content and legal feedback into a corpus usable for both training and evaluation.

The first version deliberately excluded most of Remitly's communication surface.

Root cause: Audio, image, video, and non-English support add additional compliance and perception problems, such as disclaimer font size.

Resolution: Start with English text, then expand to multimodal and multilingual checks once core behavior works.

Compliance automation must balance false positives and missed issues.

Root cause: Over-flagging creates alert fatigue and inefficiency, but missed violations create regulatory risk.

Resolution: The team measured precision and recall and iterated to reach the reported 95% accuracy and 97% issue identification.

A separate compliance tool still creates workflow friction.

Root cause: Marketers would have to leave their native creative environment to run checks.

Resolution: The roadmap is to integrate compliance checking directly into tools such as Adobe and Figma.

Evaluation

Remitly evaluated the system with quantitative precision/recall metrics, a structured content-and-legal-feedback corpus, several model iterations before beta, and usability feedback from marketers about language and issue-presentation structure.

Metrics: accuracy · issue identification rate · precision · recall · marketer usability feedback · beta readiness

Outcomes

• Remitly built a self-service marketing compliance checker for English text.
• The team reported 95% accuracy and around 97% issue identification.
• The system was being rolled out to the full marketing team with training.
• The intended workflow reduces legal review effort by moving many checks earlier into marketer iteration.
• Future plans include audio, image, video, non-English language support, and integration into Adobe or Figma.

The transcript reports 170 countries, around 8 million active customers, more than 18 languages, two-to-four manual review cycles, English text as first scope, 15 development days over 3 months, 95% accuracy, around 97% issue identification, and planned expansion to audio, image, video, and native tools.