There are two popular ways to introduce AI into customer conversations, and both of them fail.
The first is to never let it speak. The AI summarizes, suggests, tags — anything but actually answering the customer. This feels safe, but it quietly costs you the thing you bought the AI for: the reply at 11 pm, the answer during the lunch rush, the follow-up nobody had time to send. You've hired a capable assistant and assigned it to filing.
The second is to flip the switch on day one. Connect the channels, turn on auto-reply, hope for the best. This is how businesses end up with screenshots of their chatbot confidently inventing a discount, or answering a cancellation request with a cheerful upsell. One bad reply to the wrong customer and the whole project gets shelved — usually for a year.
The problem isn't the AI. It's that both approaches skip the step every human hire goes through: a probation period where you read their work before it ships.
Think of it as a new hire, not a feature toggle
When you hire a new front-desk person, you don't hand them the phone on day one and walk away. You also don't ban them from speaking to customers forever. You sit next to them. They draft, you review, you correct. Over a few weeks, you stop checking the easy stuff — directions, opening hours, "do you have availability Tuesday" — and keep an eye on the hard stuff a while longer.
AI deserves exactly the same on-ramp. In Cura, that on-ramp has three rungs, set per conversation type — not one global switch:
- Off. The AI stays silent for this kind of conversation. It still reads, tags, and routes — it just never speaks.
- Draft. The AI writes a full reply and holds it for you. You read it, edit it if you want, and send with one tap. The customer never sees anything you didn't approve.
- Auto. The AI sends on its own — but only after a separate check (more on that below) verifies the reply against your policies.
The "per conversation type" part matters more than people expect. A reasonable setup for a service business in month one looks like this: hours, location, and service questions on Auto; booking requests and reschedules on Draft; complaints, refunds, and anything involving money on Off, routed straight to a human. Nobody has to make one big scary decision. You make ten small reversible ones.
What Draft mode actually buys you
Draft mode looks like a half-measure. It's actually where most of the value shows up, for three reasons.
You see exactly what you'd be automating. After two weeks of reading drafts, you don't have to guess whether the AI handles reschedules well — you've watched it handle forty of them. The decision to move something to Auto stops being a leap of faith and becomes a review of evidence you already have.
The AI is learning your voice while you review. Every edit you make is signal. If you keep softening its openings or adding "see you soon!" to confirmations, that pattern feeds back into how it drafts. The probation period isn't just you evaluating the AI — it's the AI absorbing the house style.
Your response time collapses immediately. Even with a human approving every message, replying becomes a ten-second read-and-tap instead of a three-minute compose. For a lot of teams, Draft mode alone takes "we'll get back to you tomorrow" down to "answered within the hour" — before anything is automated at all.
The graduation rule
So when do you move a conversation type from Draft to Auto? The rule we recommend is boring on purpose:
If you're still rewriting the AI's answers about your cancellation policy, that type isn't ready — and that's fine. Hours and directions might graduate in week one while booking changes take a month. The ladder isn't a race.
And graduation isn't permanent. Launching a new service? New pricing? Drop the affected conversation types back to Draft for a week. It costs you almost nothing and catches the period when the AI's knowledge is most likely to lag reality.
Auto mode still shouldn't mean unsupervised
Here's the part that makes the top rung of the ladder safe to stand on: in Cura, Auto never means the first draft goes straight to the customer. Every outbound reply — drafted or automated — passes through a judge: a separate check that reads the reply against your knowledge base and policies before anything sends. Wrong language for this customer? Blocked. Invented price? Blocked and redrafted. Promise you don't actually make? Blocked.
The judge is what changes Auto from "hope the model is having a good day" to "the same review you were doing in Draft mode, done consistently, on every message." (It's a deep enough topic that we wrote about it separately — see what a presend judge is and why your AI agent needs one.)
A 30-day on-ramp you can steal
If you're rolling this out, here's the shape that works:
- Week 1 — everything on Draft
Read every reply before it sends. Edit freely; the edits are training data for your setup, not wasted work. Keep money and complaints on Off.
- Week 2 — graduate the obvious
Hours, location, parking, "what services do you offer" — these are usually edit-free within days. Move them to Auto. Everything else stays on Draft.
- Weeks 3–4 — graduate by evidence
Watch your edit rate per conversation type. Whatever you haven't touched in two weeks moves up. Whatever you're still rewriting stays put, and that's information too: it usually means your knowledge base is missing something, not that the AI is failing.
- Forever — keep a floor
Some conversations should stay on Draft or Off indefinitely. Complaints. Anything legal. Anything where the relationship is worth more than the response time. Knowing what not to automate is part of doing this well.
The businesses that get durable value from AI in their inbox aren't the boldest ones. They're the ones that treated trust as something the system earns one conversation type at a time — the same way they'd treat anyone else who talks to their customers.