Scaling Customer Support, Compliance, and Developer Productivity with Gen AI

coinbase · Finance · 2024 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (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: enterprise GenAI platform for support, compliance, and engineering

Coinbase needed to scale customer support, regulated compliance investigations, and developer productivity across volatile crypto markets and many geographies.

Constraints:

• Customer support demand fluctuates with market volatility and must handle many geographies and languages.
• Compliance workflows require explainable, human-reviewed decisions for high-risk cases.
• Developer productivity tools must fit existing engineering workflows while leaving final responsibility with human developers.

Architecture

Coinbase built an internal GenAI platform with standardized LLM and data interfaces, then used it for an AI support chatbot, agent-assist feedback loops, compliance auto-resolution workflows, and developer tools such as PR review and UI testing. The architecture combines RAG, vector memory, Bedrock Knowledge Bases, Cohere reranking, model-backed response generation, Bedrock guardrails, MCP data connectors, ML risk models, and human decision gates.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  A["Internal GenAI platform"] --> B["LLM interface"]
  A --> C["MCP data connectors"]
  A --> D["Customer support chatbot"]
  D --> E["Vector memory layer"]
  D --> F["Bedrock knowledge-base retriever"]
  F --> G["Cohere reranker"]
  G --> H["Response generation sub-agent"]
  H --> I["Guardrail layer"]
  D --> O["Human support agents"]
  J["Compliance risk models"] --> K["Compliance auto-resolution engine"]
  C --> K
  K --> L["Human compliance agents"]
  K --> M["Customer RFI channel"]
  A --> N["Developer productivity tools"]
  B --> N

Components

Data flow

Support users first interact with an AI chatbot backed by conversation memory, Bedrock knowledge bases, reranking, response generation, and input/output guardrails. Complex support cases escalate to human agents. Compliance risk models trigger high-risk alerts, a compliance auto-resolution engine gathers internal/external evidence through MCP data connectors, drafts a narrative summary, and leaves final SAR decisions to human compliance agents. Developer tools use Bedrock-served Claude models for coding assistance, PR review, and UI testing.

Why this architecture

Coinbase chose a shared platform plus domain-specific workflows: standard LLM and MCP interfaces make reuse possible, while support, compliance, and developer tools keep their own risk controls, data connectors, and human review loops.

Models used: OpenAI-compatible LLMs · Claude models through Bedrock · Bedrock Knowledge Bases · Bedrock Guardrails · Cohere reranker · traditional compliance ML models

Scale

• Coinbase serves customers across 100-plus geographies.
  “Coinbase serves customers across 100-plus geographies globally.”transcript, stated
• Close to 65% of customer contacts are handled automatically by AI systems.
  “So today, close to 65% of our customer contacts are handled automatically by our AI systems.”transcript, stated
• Automation saves nearly five million employee hours annually.
  “This automation, consequently, saves us nearly five million employee hours every year.”transcript, stated
• Automated cases resolve in under 10 minutes, while human-agent cases can take up to 40 minutes.
  “When cases are handled by automation, they get resolved in less than 10 minutes. But when cases go to a human agent, they can be long-drawn.”transcript, stated
• Nearly 40% of daily code at Coinbase is AI-generated or AI-influenced.
  “So today, nearly 40% of all code written daily at Coinbase is AI-generated, or AI-influenced, I should say.”transcript, stated

Alternatives considered (roads not taken)

Bespoke integrations per data source and model

Rejected because: Coinbase standardized on OpenAI API and MCP interfaces so every team could reuse and extend the platform.

Fully automated compliance filing decisions

Rejected because: The company remains liable for SAR decisions, so human compliance agents retain final authority.

AI replacing manual code review

Rejected because: The automated reviewer handles routine checks, while humans still review nuanced code issues.

Patterns

tool-use-orchestration (agent-architecture)

Coinbase uses a GenAI platform and compliance engine to coordinate LLM calls, MCP data connectors, risk models, customer RFIs, and human review.

rag-retrieval (retrieval)

The support chatbot retrieves vectorized Coinbase help articles from Bedrock Knowledge Bases before generating answers.

agent-memory (agent-architecture)

Prior conversation history is stored in a vector database and used as memory for the support chatbot.

guardrails-io-filtering (safety)

Bedrock Guardrails and domain filters protect support interactions from harmful content, PII leakage, prompt injection, and multilingual policy violations.

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

Compliance workflows synthesize evidence but leave final SAR filing decisions to human compliance agents.

Tradeoffs

Platform interface design

Chose standard OpenAI API and MCP interfaces over one-off model and data integrations.

Rationale: Shared interfaces make the platform reusable across support, compliance, and developer-productivity teams.

Support automation boundary

Chose chatbot plus escalation and agent assist over chatbot-only support.

Rationale: Automation handles simple cases instantly while complex cases escalate to human agents with AI assistance.

Developer-tool rollout

Chose multiple paved-path coding tools over one mandated assistant.

Rationale: Coinbase lets developers choose best-in-class tools that fit their workflows while serving the models through Bedrock.

Gotchas (what bit them)

Compliance automation still needs explainable human decisions.

Root cause: Regulators require detailed investigation records and the company remains accountable for filing decisions.

Resolution: Use AI to gather evidence and draft narrative summaries, then route final judgment to compliance operations agents.

One compliance workflow cannot cover all jurisdictions.

Root cause: Coinbase operates across many countries with diverse regulatory requirements.

Resolution: Adapt compliance processes by workflow and geography while using shared MCP-based platform access.

AI-generated code does not remove developer accountability.

Root cause: Some code paths are unsuitable for AI generation and every generated line still needs human understanding.

Resolution: Use coding tools as paved paths while keeping human review and judgment in the software process.

Evaluation

Coinbase tracks automation rates, support resolution time, compliance explainability, developer code influence, PR-review assistance, and UI testing automation.

Metrics: automated customer-contact share · employee hours saved · support resolution time · code generated or influenced by AI · unit-test coverage gaps · CI failure debugging support

Outcomes

• Close to 65% of customer contacts are handled automatically.
• Automation saves nearly five million employee hours annually.
• Automated cases resolve in less than 10 minutes compared with human-agent cases that can take up to 40 minutes.
• Nearly 40% of daily code is AI-generated or AI-influenced.

The platform spans customer support, compliance, and developer productivity rather than a single GenAI use case.