Drafting a reply with AI, without letting it send
AI that drafts and never sends is a different proposition from a bot. This is the one an agent will actually use.
The Suggest button drafts a reply from your own knowledge base and puts it in the message box for you to edit. It never sends anything. A person always reads it, changes what needs changing, and presses send.
That constraint is the whole design. An AI that drafts is useful to an agent in a way an AI that sends is not, because the agent stays accountable for what the customer receives.
Where the answer comes from
Your material — the pages you added and the question-and-answer pairs you wrote. Not the open internet, and not the model's general knowledge about businesses like yours.
That matters because it is the difference between a draft you can trust to be about your delivery times and one that is plausibly about somebody else's. If your knowledge base says nothing about a topic, the button will tell you so rather than improvise.
Why it sometimes declines
It is deliberately willing to produce nothing, and it tells you which reason applies:
| What it says | What to do |
|---|---|
| No knowledge base yet | Add your sources — nothing can be drafted from nothing |
| No customer question to answer yet | It drafts a reply to something; there is nothing to reply to |
| Nothing close enough in your knowledge base | Answer by hand, then add that answer so next time it can |
| The AI wouldn't answer this from your knowledge base | Usually the right call — answer it yourself |
| Too many suggestions just now | Wait a moment |
| Not available on this channel yet | Nothing to do |
The third row is the useful loop. A refusal is a list of the gaps in your knowledge base, generated by real customer questions, and the fix takes two minutes while the question is in front of you.
How to use it well
- Read the draft properly. Not skim — read. You are sending it.
- Cut what does not apply. Drafts tend to be complete rather than concise, and customers prefer concise.
- Add the specific detail the knowledge base cannot know — their order number, the actual date.
- If it is wrong, fix the source rather than just the message.
Step four is what makes the tool improve over months rather than staying where it started.
What it is not
It is not the automated answering that happens before an agent is involved. That is a step inside a flow, it runs without anybody watching, and it has its own handover path for questions it cannot answer — covered in AI answers from your own website.
Both read the same knowledge base, which is the point: one set of answers, used automatically where that is safe and as a draft where a person should decide.
Why drafts are tracked separately
A small detail with a real consequence for your reporting. A suggestion an agent edited and sent is not the bot answering a customer, and it is recorded as what it is.
Without that separation, a busy team using the draft button would make automated resolution look far higher than it is, and you would draw the wrong conclusion about how much work the bot is absorbing. The measurement side is in measuring response time and agent workload.
Where it earns its place
- A long factual answer — policies, specifications, process steps — where typing is the slow part.
- A new agent who does not yet know your policies by heart.
- A busy hour, where the difference between drafting and typing is several minutes per conversation.
- A question you know you have answered before and cannot remember how.
Where it does not: anything about money, anything about a specific order, and anything a customer is upset about. Those are conversations to have in your own words.
Frequently asked
Can it send automatically if we want it to?
Not from this button. Automatic answering is a separate feature you configure deliberately in a flow, which is the right place for that decision.
Does the customer know a draft was AI-assisted?
They see whatever you send. Since you edit it and it comes from your own material, what they get is your answer.
Will it invent a delivery time?
It answers from your knowledge base, and it declines when the material does not cover the question. That is the guard.
What if it keeps declining?
Your knowledge base is thin. Each refusal names a gap worth filling, and filling them is the fastest way to make the button useful.
The loop that makes it better
Every refusal is a gap in your knowledge base, named by a real customer at the moment it mattered. Treat them as a queue:
- Answer the customer yourself, properly.
- Paste that answer into your knowledge base as a question-and-answer pair while it is in front of you.
- Next time the same question arrives, the draft is there.
Teams that do this for a month end up with a knowledge base built from real questions rather than from what somebody guessed customers would ask — which is also the material the automated answering uses, so the benefit lands twice.
Where it fits with quick replies
| Quick reply | AI draft | |
|---|---|---|
| Answer is | Fixed, word for word | Composed for this question |
| You are | Inserting a known answer | Editing a proposed one |
| Best for | Policy, details, boilerplate | Questions phrased in a new way |
| Speed | Instant | A moment |
Use the saved answer where one exists, and the draft where none does. They are not competing — the fastest agents reach for the shortcut first and the button when the question is unfamiliar, as described in quick replies for a team.
What it does not change
Nothing about Meta's rules. A drafted reply is an ordinary message, so it still needs an open 24-hour customer service window under Meta's sending messages documentation — the button will happily draft an answer into a locked composer, and you will still need a template to send anything at all.
Inside the window it costs nothing, because service conversations have been free and unlimited since 1 November 2024 under Meta's pricing documentation. So the draft button saves time rather than money, and the states of the lock are in the 24-hour countdown.
A realistic expectation
It will be useful on perhaps 3 or 4 questions in 10 at first, and more as your knowledge base grows. That is a reasonable return for a button, and it is a much better outcome than a system that answers all 10 with 2 of them subtly wrong.
The refusals are the feature. A tool that declines when it does not know is a tool an agent can trust the rest of the time, which is the only basis on which anybody uses one of these for longer than a week.
Who should use it, and when
| Agent | Use it for |
|---|---|
| New this week | Almost everything factual — it teaches your policies as it drafts |
| Experienced | Long answers, and questions phrased in an unfamiliar way |
| Anyone, at peak | Clearing a backlog of factual questions quickly |
| Anyone, on a complaint | Nothing — write those yourself |
The last row is the one to be firm about. A customer who is upset can tell the difference between an answer and a generated one, and that is the conversation where it matters most.
The short version
It drafts from your own knowledge base and never sends. Read what it produces, cut what does not apply, add the specifics it cannot know, and press send yourself. When it declines, that is a gap worth filling while the question is still in front of you — and filling it improves both the draft button and the automated answering, because they read the same material.
Use it for long factual answers and unfamiliar phrasings. Do not use it for money, for a specific order, or for anybody who is upset.
Reviewing what it has been drafting
Worth doing once a month for the first few months. Read a handful of conversations where an agent used the button and ask two questions: was the draft accurate, and did the agent change it before sending?
An agent sending drafts unedited is a signal, not a crime — usually it means the drafts are good, occasionally it means somebody has stopped reading. Either way it is better to know, and the answer changes how much you lean on the feature.
Where the drafts come from, restated
Worth being unambiguous, because "AI" covers a lot of very different things. This button reads the material you gave it — your pages and your written answers — and composes a reply from that. It is not searching the web, it is not drawing on general knowledge about businesses like yours, and it is not learning from your conversations.
Which means the quality of the drafts is entirely downstream of the quality of your knowledge base. A thin knowledge base produces refusals; a good one produces drafts worth sending. There is no third option where the tool covers for material you never wrote.