Customer support · Production

GoDaddy Digital Care: 10 lessons learned operationalizing LLMs in customer support messaging

The problem

GoDaddy's Digital Care team handles over 60,000 customer contacts per day across messaging channels. Their initial mega-prompt AI assistant became unwieldy as prompts grew past 1,500 tokens, accuracy declined with each new instruction added, and LLMs proved slow and unreliable in production.

First attempt

The mega-prompt caused the LLM to ask questions unrelated to the target topic. Users were sometimes stuck with a bot that refused to transfer them to a human agent. Early RAG implementations queried on every prompt invocation and retrieved irrelevant content before customer intent was established.

Workflow diagram · grounded in source
1
Customer contacts via messaging
trigger
“My team, Digital Care, leverages LLMs to provide an exceptional customer experience in our messaging channels (SMS, WhatsApp, and web)”
2
AI classifies support topic
ai_action
“The assistant would communicate with the customer until it could classify the conversation into one of twenty support topics we use to bucket our support inquiries”
3
Guardrails check content
validation
“our chat systems (and other tools) use controls to check for personally identifiable information and offensive content in AI responses, user messages, and prompt instructions”
4
Intent-aware RAG retrieval
integration
“Subsequent implementations involved switching to a specialized RAG prompt after we determined the intent of the conversation”
5
Topic-specific question collection
ai_action
“the assistant would ask a set of questions specific to that topic that would help our Guides (human support agents) accelerate the support process”
6
Route to support queue
routing
“route the conversation to a support queue matching the topic”
7
Summary extracted for human Guide
output
“when a user needed to transfer to a human agent, we would execute a task to extract vital information to provide a summary for the human agent”
Reported outcome

GoDaddy evolved to a multi-agent Controller-Delegate architecture, implemented deterministic guardrails for human transfers, and adopted intent-aware RAG with LLM Agents.
Early SPR experiments showed over 50% reduction in token usage.

Reported metrics
Daily customer contacts in messaging channelsover 60,000
Prompt size at second experiment launchover 1500 tokens
Chat completion failure rate at model provider1%
invalid output rate on ChatGPT 3.51%
Show all 8 reported metrics
daily customer contacts in messaging channelsover 60,000
prompt size at second experiment launchover 1500 tokens
chat completion failure rate at model provider1%
invalid output rate on ChatGPT 3.51%
invalid output rate on ChatGPT 40.25%
ChatGPT 4.0 average response latency for completions under 1000 tokens3-5 seconds
maximum observed LLM call durationup to 30 seconds
token usage reduction with SPR techniqueover 50%
Reported stack
ChatGPT 3.5 TurboChatGPT 4.0ChatGPT functionsRAGSPRs
Source
https://www.godaddy.com/resources/news/llm-from-the-trenches-10-lessons-learned-operationalizing-models-at-godaddy
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

GoDaddy evolved to a multi-agent Controller-Delegate architecture, implemented deterministic guardrails for human transfers, and adopted intent-aware RAG with LLM Agents.

What tools did this team use?

ChatGPT 3.5 Turbo, ChatGPT 4.0, ChatGPT functions, RAG, SPRs.

What results were reported?

Daily customer contacts in messaging channels: over 60,000; Prompt size at second experiment launch: over 1500 tokens; Chat completion failure rate at model provider: 1%; invalid output rate on ChatGPT 3.5: 1% (source-reported, not independently verified).

What failed first in this deployment?

The mega-prompt caused the LLM to ask questions unrelated to the target topic.

How is this customer support AI workflow structured?

Customer contacts via messaging → AI classifies support topic → Guardrails check content → Intent-aware RAG retrieval → Topic-specific question collection → Route to support queue → Summary extracted for human Guide.