In AI customer support, draft mode means the AI writes a reply that a human approves before it reaches the customer. Auto-reply means the AI sends the reply directly without human review. Both have legitimate uses — the right choice depends on your ticket volume, brand risk tolerance, and how well your AI has been calibrated on your specific content.
Understanding the Two Modes
Every AI customer support platform that generates replies offers some version of these two modes. The terminology varies — "draft mode", "suggest mode", or "co-pilot" versus "auto-pilot", "auto-send", or "autonomous mode" — but the underlying mechanism is identical.
Draft mode (human-in-the-loop):
- AI generates a reply
- Reply is held in the agent inbox as a draft
- Human reviews, edits if needed, then approves
- Customer receives the approved message
Auto-reply (AI-autonomous):
- AI generates a reply
- Reply is assessed against a confidence threshold
- If score meets or exceeds threshold: message sent directly to customer
- If score falls below threshold: reply held as draft for human review
The difference is not which mode is inherently better — it is which mode is appropriate for which situation, which ticket type, and which stage of your AI deployment.
When to Use Draft Mode
Draft mode is always the right starting point. It lets you observe how the AI performs on your specific ticket types, customer base, and product catalogue before trusting it to send messages independently.
Use draft mode when:
- You have just deployed the AI for the first time — always start here
- You are handling a new ticket category the AI has not encountered before
- The query involves a financial commitment such as a refund, partial credit, or compensation
- The query is from an escalating or clearly upset customer
- Your brand tone is highly specific and requires consistent enforcement by an editor
- You operate in a regulated industry where incorrect information carries legal risk
Draft mode is not a failure state or a consolation prize. Many mature support teams keep complaints and refund requests in permanent draft mode — not because the AI is poor, but because those query types carry sufficient risk that human review is always worth the time investment.
When to Use Auto-Reply
Auto-reply is appropriate once you have built sufficient evidence that the AI performs well on a given ticket category.
Use auto-reply when:
- You have run the AI in draft mode for at least one to two weeks and reviewed its accuracy systematically
- The ticket category is factual, low-risk, and consistent (order tracking, opening hours, standard shipping times)
- Your confidence threshold is set at 78 or above
- Your knowledge base has comprehensive, up-to-date articles on the topic
- You have a clear escalation path in place for queries the AI cannot confidently handle
A Shopify store handling 400 WISMO tickets per month where the AI is accurate on 88% of them is an obvious candidate for auto-reply on that category. The alternative — having an agent approve 400 near-identical correct drafts — is operational waste, not quality assurance.
A Decision Framework by Ticket Category
Ask four questions to determine the right mode for any given ticket category:
| Question | Draft Mode Indicator | Auto-Reply Indicator |
|---|---|---|
| Has the AI been reviewed on this category? | Not yet | Yes, for at least one week |
| Is AI accuracy above 80% on this category? | Below 80% | Above 80% |
| What is the cost of an error on this category? | High (financial, legal, emotional) | Low (factual, easily corrected) |
| How frequently does this category occur? | Low volume | High volume |
If any single answer points toward draft mode, use draft mode for that category. You can run different categories in different modes simultaneously. Auto-reply for order tracking does not require auto-reply for refund requests.
The Hybrid Approach: Category-Level Mode Selection
The most effective configuration is neither "all auto-reply" nor "all draft mode" but a calibrated mix by ticket type.
Example configuration for a mid-sized e-commerce store:
| Ticket Category | Mode | Notes |
|---|---|---|
| Order tracking (WISMO) | Auto-reply | Well-calibrated, factual, high volume |
| Delivery times | Auto-reply | Consistent answers, low risk |
| Stock availability | Auto-reply | Simple lookup, AI accuracy typically high |
| Returns policy | Draft mode | Must match policy precisely |
| Refund requests | Draft mode | Financial commitment — human confirmation |
| Complaints | Draft mode | Emotional, complex, relationship-critical |
| Account issues | Draft mode | Security-adjacent — err on the side of caution |
This configuration means the AI handles approximately 55–65% of tickets fully autonomously while humans review everything with higher stakes. The support team's workload concentrates on the queries that genuinely benefit from human judgement.
How Draft Mode Improves AI Accuracy Over Time
An often-overlooked benefit of draft mode is that it generates improvement signal. When agents edit an AI draft before approving it, that edit is data: the AI produced reply X, the correct reply was Y.
In systems like Kriseena, the confidence score distribution for each ticket category is tracked over time. Queries where AI drafts are frequently edited — indicating systematic shortfalls — surface as knowledge base improvement opportunities. Adding or updating a knowledge base article on that topic typically raises confidence scores on similar future queries within days.
Draft mode is not only a safety net. It is a feedback loop that actively improves the AI's accuracy on your specific content over time.
Auto-Reply and the Confidence Threshold: Why They Must Go Together
Auto-reply without a confidence threshold is dangerous. If the AI sends every reply regardless of certainty, it will inevitably send wrong answers with confidence — and a confident wrong answer causes more damage than a delayed correct one.
A properly configured auto-reply system works as follows:
- AI generates reply and calculates confidence score
- Score meets or exceeds threshold: message sent to customer automatically
- Score falls below threshold: reply held as draft for human review
- Human resolves the query
- If the same low-confidence query recurs repeatedly, a new knowledge base article is added
The threshold acts as a quality gate. The goal is not to maximise the percentage of messages sent automatically — it is to maximise the percentage of correct messages sent automatically. That distinction matters in practice.
Common Deployment Mistakes
Switching to full auto-reply too early Enthusiasm for automation leads teams to enable auto-send before the AI has been properly calibrated. Errors reach customers. Trust in the system deteriorates. Teams revert to fully manual handling — discarding the efficiency gains they had achieved. The faster path is slower: draft mode first, evidence second, auto-reply third.
Leaving everything in draft mode indefinitely Draft mode without any auto-send means every message still requires a human touch. The AI saves perhaps 30–40% of agent time (less writing) rather than 70–80% (fewer tickets handled manually at all). This under-utilises the technology and understates its ROI to stakeholders.
Treating all ticket types identically WISMO queries and refund requests have different risk profiles. Using the same mode for both means either over-automating (sending uncertain refund replies) or under-automating (reviewing hundreds of correct WISMO drafts every day).
Rubber-stamping drafts without reading them Draft mode only improves AI accuracy if agents actually engage with what they are reviewing. If agents approve every draft in under two seconds without reading, the feedback loop breaks and the AI cannot improve. Invest time in early review — it pays compound returns over weeks and months.
Key Takeaways
- Draft mode is always the right starting point; auto-reply is earned through demonstrated evidence
- Use auto-reply for high-volume, low-risk, factual categories once accuracy is confirmed for each category
- Maintain draft mode permanently for financial decisions, complaints, and escalations
- Run different ticket categories in different modes simultaneously — hybrid is the most effective configuration
- Auto-reply always requires a confidence threshold; enabling it without one introduces unnecessary risk
- Agent edits to drafts are feedback — they actively improve AI accuracy over time when acted upon
Frequently Asked Questions
Can I switch between draft mode and auto-reply without losing historical data? Yes. Switching modes is a configuration change, not a data operation. All historical messages, confidence scores, and draft records are retained. You can switch back to draft mode at any time if you want to review AI performance more closely after a knowledge base change.
What happens to a message that falls below the auto-reply threshold? It is added to the agent inbox as a draft with the AI's suggested reply pre-populated. The agent sees the confidence score alongside the draft and can approve it with one click, edit it, or write from scratch. No message reaches the customer until an agent approves it.
Is auto-reply suitable for B2B SaaS companies, not just e-commerce? Yes. Auto-reply is suitable for any repetitive, factual ticket category in B2B SaaS — billing questions, feature availability, integration compatibility, known bugs with documented workarounds. The principle is the same: validate accuracy by category, then enable auto-reply for that category. B2B companies typically have longer customer lifetimes and higher relationship stakes, so the recommended starting threshold is higher: 82–85 for B2B versus 75–80 for e-commerce WISMO.
Can customers tell whether a message was sent by AI or approved by a human? With well-calibrated auto-reply and a knowledge base that reflects your brand voice, customers typically cannot distinguish AI-sent from human-approved messages. Some companies disclose AI involvement in their support; others do not. This is a brand and transparency decision rather than a functional one.
How long should I run draft mode before enabling auto-reply for a given category? A minimum of one week on a live sample of at least 50 tickets in that category. Smaller samples do not give sufficient signal to assess accuracy reliably. Two weeks is better. After that, review accuracy specifically for that category, set your threshold, and monitor closely for the first week of auto-reply operation before widening its scope.