Draft Mode or Auto-Send: How to Decide What Your AI May Reply Without You

A team switches on an AI assistant and argues for a week about whether it can be trusted. Half want it answering everything at night. Half want every word checked. Both are answering the wrong question. Whether to use draft mode or auto-send is not one decision for the whole business. It is a decision per type of message, and it can be made with two questions.

The two modes

• Draft. The AI prepares a reply. A person approves, edits or discards it. Nothing reaches the customer without a human pressing send.
• Auto-send. The AI replies directly. A person reviews afterwards, by sampling.

Draft mode is slower and safer. Auto-send is instant and depends entirely on the quality of the AI's information and limits. Most businesses need both, applied to different things.

The two-question test

For each type of message, ask:

1. What does a wrong answer cost? Could it lose money, break a promise, cause a legal problem, or embarrass the business in a screenshot? Or would it be a small inconvenience corrected with one more message?

2. How fixed is the right answer? Is there one correct reply that rarely changes and is written down? Or does it depend on stock, the customer's history, judgement or negotiation?

Then:

• Cheap to get wrong, fixed answer: auto-send.
• Cheap to get wrong, variable answer: auto-send with a tight scope, or draft.
• Costly to get wrong, fixed answer: draft at first, auto-send after testing.
• Costly to get wrong, variable answer: draft always, or a person from the start.

Common message types

Your table will differ. Building it takes an hour and replaces the argument about trust.

What draft mode is good for

• Learning. In the first weeks, every draft shows what the AI knows and where it guesses. Each edit is information about what to fix in its instructions or reference material.
• Speed without risk. Reading a draft takes seconds, and writing a reply takes minutes.
• Consistency. Drafts follow the same tone and policy whoever is on shift.
• Training new staff. A new admin sees a reasonable reply to adapt.

Its weakness is that it needs someone available. A draft waiting at 23:00 helps nobody until morning.

What auto-send is good for

• Out-of-hours coverage for simple questions.
• First response time. A customer who gets an instant, correct answer to "are you open on Sunday?" does not go elsewhere.
• Volume. Repetitive questions stop consuming the team's day.

Its weakness is that mistakes reach the customer before anyone sees them. This is why its scope must be narrow and its information current.

Moving a category from draft to automatic

Do not move things on a feeling. Use a small test.

1. Run the category in draft mode for two to four weeks.
2. Count the outcome of each draft: sent unchanged, sent with edits, discarded.
3. Read the edits. Are they corrections of fact, or matters of style? Factual corrections mean the AI's information is incomplete.
4. Set a bar. For example, at least 95 of every 100 drafts sent unchanged, and none of the others containing a costly error.
5. Fix and repeat until the bar is met.
6. Switch to auto-send for that category only.
7. Keep sampling. Read twenty of its automatic replies each week.
8. Move it back if error rates rise, for instance after a price change.

A category earns automation with evidence, and can lose it again.

Limits that apply in both modes

• A handover rule. If the customer is upset, asks for a person, or asks twice without a satisfactory answer, the AI stops and passes the chat on.
• "I will check" is allowed. The AI should say it will ask the team when the answer is not in its reference, not produce something plausible.
• No commitments beyond stated limits. Discounts, delivery dates and exceptions have ceilings.
• Honesty about being automated, where customers ask or where the context calls for it.
• A record. Every automatic reply and every approved draft should be traceable.

Mixed arrangements by time of day

Many small teams settle on a pattern like this:

• Working hours: draft mode for most things, auto-send for the simplest.
• After hours: auto-send for simple questions, and for everything else an honest holding message: "Thank you, our team will reply at 08:00 tomorrow", with the draft prepared for the morning.

The morning shift then starts with a queue of prepared replies instead of a blank inbox.

Common mistakes

• One setting for everything.
• Starting with auto-send to save time, and learning from customers what the AI did not know.
• Staying in draft mode forever for categories that have been approved unchanged for months.
• No sampling after automation.
• Forgetting that information goes stale. A category that was safe in June may not be after a price change in September.

In Polynusa

Polynusa's AI agents support both arrangements: replies prepared as drafts for approval, and replies sent directly within what the workspace allows. We suggest every new workspace start in draft mode, build the table above, and promote categories one at a time. Plan details are on the pricing page.

Summary

Decide per message type, not for the business as a whole. Ask what an error would cost and how fixed the right answer is. Send automatically only where errors are cheap and answers are fixed, draft where either is in doubt, and hand over complaints and commitments to a person. Promote a category after weeks of drafts sent unchanged, keep sampling, and demote it when the facts change.

FAQ

What is draft mode for an AI assistant?

In draft mode the AI writes a suggested reply and a person reads it, edits if needed, and sends it. The customer only receives what a person approved. It keeps most of the time saving, because reading and approving is faster than writing, while a human remains responsible for what goes out.

Is it safe to let AI send replies automatically?

For questions with fixed, published answers, such as opening hours, location, or how to order, automatic replies are low risk once tested. For anything that commits the business, such as prices outside the list, refunds, delivery promises or complaints, keep a person in the loop.

How do we know when a category is ready for auto-send?

Track how often the team sends the AI's drafts for that category without changes. When a few weeks of drafts go out unedited at a very high rate, and the category's errors would be cheap to correct, it is a candidate. Move it, then keep sampling.