One of the highest-friction tasks in customer support is composing replies to tickets. An agent reads a new ticket, recalls the relevant information from their training and the knowledge base, and writes a clear, accurate, and helpful response. For a seasoned agent handling a familiar question, this takes two to four minutes. For a new agent, a complex question, or a topic that requires looking something up, it can take significantly longer.
Multiply that by 50–100 tickets per agent per day and the composing time alone represents a significant fraction of the workday. AI draft replies compress that time — for the right types of tickets — to under a minute.
This article explains how AI draft replies work, when they are most effective, and what to watch out for when deploying them.
How AI Draft Replies Work
The mechanism behind AI draft replies has three steps:
Step 1: Retrieve Relevant Context
When a ticket arrives, the AI first identifies what the ticket is about — the intent, the key question, the product area. It then searches the knowledge base for articles that are relevant to that intent.
The search can use:
- Full-text search: finding articles that contain the same words as the ticket
- Semantic search (more advanced): finding articles that address the same topic even if the wording is different — a ticket asking "why did my payment fail" surfaces articles about payment errors, billing issues, and card declined scenarios even if those exact words are not all present in the article titles
The AI returns the top 3–5 most relevant articles.
Step 2: Generate a Draft Based on Source Material
The AI uses the retrieved articles as source material to write a reply. This is called Retrieval-Augmented Generation (RAG) — the AI is not generating a reply from general knowledge, it is synthesizing an answer from your specific documentation.
This matters enormously. A pure LLM generating a reply from general knowledge might hallucinate your pricing, invent a feature that does not exist, or give advice that conflicts with your actual policies. A RAG-based draft is anchored in your documentation and substantially less likely to produce factually incorrect content about your product.
Step 3: Agent Reviews and Sends
The draft appears in the agent's reply panel. The agent reads it, edits as needed, and sends — or discards it if it misses the mark. The agent is always the final decision point. Nothing is sent automatically without agent review.
What AI Draft Replies Are Good At
Answering Documented Questions
If a customer asks "how do I enable two-factor authentication?", and you have a knowledge base article covering exactly that, the AI draft will be close to perfect. The agent may add a personal greeting or a closing note, but the substantive content will be accurate and complete.
For this category of question — well-defined, fully documented — AI draft replies typically save 80–90% of the composing time.
Combining Multiple Sources
A customer question that spans two or three topics — "how do I switch plans and how does that affect my billing cycle?" — requires the agent to mentally combine information from a plans article and a billing article. The AI does this synthesis automatically, producing a single cohesive reply from multiple sources.
Consistent Tone and Format
AI-drafted replies tend to be consistent in structure and tone. They do not vary in formality the way human-written replies do from agent to agent or from a morning shift to a tired Friday afternoon.
New Agent Onboarding
New agents who do not yet know the product well can use AI draft replies as a learning tool. They see the AI's suggested reply (grounded in the knowledge base), read the source articles it references, and build product knowledge through the act of reviewing and editing suggestions.
Where AI Draft Replies Fall Short
Questions Not in the Knowledge Base
If the customer's question is about something not documented — a new feature, an edge case, a question about your internal processes — the AI has nothing to reference and will either produce an unhelpful reply or, in poorly configured systems, hallucinate an answer.
The fix is not technical — it is editorial. Keeping the knowledge base current and comprehensive directly improves AI draft quality. The AI is only as good as its source material.
Emotionally Complex Situations
A customer who is angry, disappointed, or has experienced a service failure needs more than an accurate technical response. They need acknowledgment and empathy — elements that an AI can approximate but rarely hits exactly right in context.
Most agents handling emotionally charged tickets correctly learn to discard the AI draft and write a human reply that addresses the emotional dimension first. This is the right behavior — and it is available because the agent is in control.
Highly Specific Account Questions
"Why does my invoice show €287.50 this month when it was €240 last month?" is not a question a knowledge base article can answer. It requires looking at the customer's specific account, usage data, and billing history. The AI has no access to this information and should not attempt to answer it from general knowledge.
Most platforms handle this by not generating a draft when the intent is clearly account-specific — they flag it as "requires account review" instead.
Measuring the Impact on Your Team
Metrics to Track
Draft acceptance rate: what percentage of AI drafts do agents send with minimal editing? A high acceptance rate (60%+) on eligible tickets indicates the drafts are consistently useful. A low rate suggests the knowledge base quality or AI configuration needs attention.
Time-to-first-response: compare FRT before and after enabling AI drafts. For eligible ticket types, FRT should decrease.
Agent satisfaction: do agents find the drafts helpful? This is worth asking directly. Agents who feel AI is monitoring them or creating pressure to respond faster may resist the feature, reducing its benefit. The framing matters: AI drafts are a writing assistant, not a performance metric.
Segmenting Results
Not all tickets are equal candidates for AI drafts. Track results separately for:
- Simple, single-topic tickets (highest AI draft utility)
- Multi-topic tickets (moderate utility)
- Account-specific or emotionally complex tickets (low utility, drafts often discarded)
This segmentation tells you where the AI is genuinely helping and where it is being bypassed anyway.
Implementation Tips
Start with high-volume, well-documented question types. Pick your top five ticket categories, ensure the knowledge base articles for those topics are excellent, and measure the draft quality specifically for those categories before rolling out broadly.
Set accurate agent expectations. AI drafts are not perfect. Agents should expect to edit every draft before sending. The time savings comes from not writing from scratch — not from automating the full composing process.
Review discarded drafts. When agents discard an AI draft without using it, that is signal. A weekly review of discarded drafts — what they contained and why agents did not use them — is one of the most direct feedback loops for improving the knowledge base.
Do not use draft acceptance rate as a performance metric for agents. If agents feel their metrics depend on accepting AI suggestions, they will accept drafts they should edit — reducing response quality. The goal is better responses faster, not higher AI utilization.
How Nura24 Implements AI Draft Replies
Nura24's AI draft reply feature appears directly in the ticket reply panel when an agent opens a ticket. The system searches the tenant's knowledge base using a combination of full-text and semantic search, retrieves the top relevant articles, and generates a suggested reply using the configured AI model (Claude Sonnet by default). The draft is marked clearly as an AI suggestion. Agents can send it as-is, edit it before sending, or close it and write their own reply. The feature is opt-in per workspace and is available on paid plans. Draft quality improves directly as the tenant's knowledge base grows — making the knowledge base and AI draft features mutually reinforcing investments.