AI Customer Support Specialist Job Description Template (2026)

Full JD (copy-ready)

Paste this into your ATS or careers page. Edit the company name and any bracketed placeholders.

# AI Customer Support Specialist — Job Description

## Role summary
An AI Customer Support Specialist designs and operates the AI-augmented support stack: training Intercom Fin, Ada, or Zendesk AI on your product; engineering a knowledge base structured for RAG retrieval; tuning prompts and escalation logic; auditing AI conversations daily to catch hallucinations; and measuring deflection rate, AI-resolved CSAT, and cost per ticket to prove the system is actually reducing human load instead of just generating plausible-sounding nonsense.

## Responsibilities
- Train and tune production AI support agents (Intercom Fin, Ada, Zendesk AI, Drift, Kustomer IQ) against real ticket history and product documentation.
- Engineer the knowledge base for retrieval: chunk sizing, metadata tagging, heading hierarchy, and content freshness — not just writing articles for humans.
- Audit 5–10% of AI conversations daily, tag failure modes (hallucination, missing context, wrong escalation, tone mismatch), and ship fixes within 48 hours.
- Design escalation and human-handoff flows: sentiment-triggered routing, complexity triggers, topic blocklists (billing, cancellations, legal, safety) that always route to a human.
- Write system prompts, few-shot examples, and guardrails that shape AI behavior within your brand voice and compliance constraints.
- Maintain a deflection and quality dashboard tracking AI resolution rate, CSAT delta (AI vs human), first-response time, cost per ticket, and escalation accuracy.
- Run A/B tests on prompts, retrieval strategy, and escalation thresholds with statistical rigor — not one-day vibes checks.
- For custom RAG deployments: maintain the embedding pipeline, vector store (Pinecone, Typesense, pgvector), and answer-layer LLM calls (OpenAI, Anthropic).
- Document every AI failure with root cause and prevention — build a weekly failure report that shrinks over time.
- Coordinate with human support reps during transitions: what the AI handles, what escalates, what context gets passed on handoff.
- Monitor LLM cost: token usage, caching hit rate, model selection (GPT-4o-mini vs Claude Haiku vs Sonnet) based on query complexity.
- Stay current on new AI support tools and models, run quarterly evaluations, and recommend migrations when something materially better ships.

## Must-have skills
- 2+ years in customer support AT LEAST 1 year operating a production AI chatbot (Intercom Fin, Ada, Zendesk AI, Drift, or a custom RAG build).
- Hands-on experience tuning prompts, not just configuring SaaS workflows — understands how wording a system prompt changes output quality.
- Knowledge base authoring with RAG in mind: chunk boundaries, metadata, avoiding contradictory content across articles.
- Fluency with at least one helpdesk (Zendesk, Intercom, HelpScout) at an admin level — not just answering tickets.
- Comfort reading basic conversation logs and JSON payloads to trace why the AI responded a certain way.
- Metrics literacy: can define deflection rate, calculate it, explain why AI-resolved CSAT needs a separate cohort.
- Written English strong enough to author help center content, system prompts, and failure reports at a US-native standard.

## Nice-to-have skills
- Custom RAG build experience: embeddings, vector stores (Pinecone, Typesense, Weaviate, pgvector), retrieval tuning.
- OpenAI API or Anthropic API hands-on — function calling, structured outputs, evaluations.
- LangChain, LlamaIndex, or similar orchestration frameworks.
- Prompt evaluation frameworks (Promptfoo, LangSmith, Braintrust) and regression testing for prompts.
- SQL for pulling conversation data from a warehouse (BigQuery, Snowflake, Redshift).
- Light Python or TypeScript for scripting data prep, log analysis, or webhook integrations.
- Experience with voice AI (Retell, Vapi) for deflection on phone.

## Tools and technology
- Intercom Fin
- Ada
- Zendesk AI / Advanced AI
- Drift
- OpenAI API (GPT-4o, GPT-4o-mini)
- Anthropic API (Claude Sonnet, Haiku)
- Pinecone / Typesense / pgvector
- LangSmith / Braintrust / Promptfoo
- Notion / Confluence
- Looker / Metabase

## Reporting structure
Reports to the Head of Support, Head of CX, or VP of Customer Operations. Collaborates with human support reps on handoff design, with product on feature-signal aggregation, and with engineering on API integrations and custom RAG infrastructure where applicable.

## Success metrics (KPIs)
- AI deflection rate 30–60% (tickets resolved without human touch).
- AI-resolved CSAT within 5 points of human-resolved CSAT.
- Escalation accuracy 95%+ (AI routes to human when it should, handles when it should).
- Weekly failure count trending down; zero repeat failures on same root cause.
- Cost per resolved ticket reduced 40–70% vs human-only baseline.
- Knowledge base coverage: 90%+ of top-50 question types have a dedicated article.