When a business decides to use AI in support, the choice is usually framed as "deploy a chatbot or not." But that's the wrong fork. The real question is different: should AI answer the customer itself, or should it help a human answer? The difference between an autonomous chatbot and an AI draft is the difference in who's accountable for every word that reaches the customer.
This article walks through both approaches in detail: how they work technically, where each one is risky and where it's useful, and what actually makes sense for a small team that can't afford to lose customer trust to a failed automation.
How an autonomous chatbot works
A chatbot gets a message and answers immediately, on its own, with no human involved. Its strength is being always-on and instant: a simple, often-repeated question gets an answer at 3 a.m. But that same autonomy is also its main weakness.
Technically, a bot runs on one of two patterns. The first is scripted: a set of predefined branches — "if the customer says X, reply Y" — sometimes with buttons instead of free text. The second is generative: a language model composes an answer on the fly, based on a knowledge base or a product description. A scripted bot is predictable but breaks on any question outside its script. A generative one is more flexible, and for that same reason riskier — the model can confidently invent a detail that exists in no document at all.
Either way, the customer gets a final answer instantly, with no pause for review. If the answer is correct, that's a real speed win. If it isn't, the fix happens after the fact, once the customer is already annoyed and possibly already distrustful.
The core risk of auto-replies
The problem isn't that a bot sometimes makes mistakes — everyone does. The problem is that a bot makes mistakes confidently and straight to the customer, with no human to catch it. One confident wrong answer about billing or a refund costs more than an hour of silence: the customer loses trust instantly, and it's hard to win back. That's exactly why auto-replies are dangerous where the cost of a mistake is high.
Picture a typical case: a customer asks whether they can get a refund for a subscription they cancelled three days ago. The bot, without the exact details of this specific case, answers with a generic phrase pulled from the knowledge base — and gets the refund terms wrong. The customer sees a confident, official-looking answer and acts on it. When the discrepancy surfaces, it's a human who has to rebuild the trust, and the conversation ends up harder than if the bot had never replied at all. On talking to a customer who's already upset, see the article on handling angry customers.
How an AI draft works
An AI draft flips the logic: the model prepares an answer but doesn't send it. Instead the draft appears in front of the operator, who decides in a second — accept, edit or write their own. The customer gets the speed of AI, but every word passes through a human. That's "human in the loop": automation prepares, the human decides. More on the mechanics — in the article on AI drafts.
Technically, it works like this: once, in the settings, the team fills in a product description — what it does, what plans or features it has, the kinds of questions operators usually answer. When a new message comes in — whether from the Telegram bot, the chat widget on the site, or the public API — the model takes that description together with the history of that specific conversation and builds a draft reply right in the field the operator was already about to type into.
The operator has one move to make: hit Tab to accept the draft as is, edit a few words, or ignore it entirely and write their own reply. Nothing goes to the customer until the operator confirms it. The draft is prepared in the customer's own language, too — which matters a lot when a team is running support across several languages at once. More on this — in the article on multilingual support with AI.
Why a draft doesn't slow support down
The main objection to drafts is that they're supposedly slower than auto-replies, since they add a human to the chain. In practice the difference is measured in seconds, not minutes: reading a ready draft and hitting Tab is faster than typing a reply from scratch. The operator saves exactly the time that usually goes into composing — choosing words, recalling details from earlier messages, checking the tone.
This matters most on a team where one or two operators handle the whole flow of incoming messages and burn out quickly on routine, repetitive questions. A draft takes the most tedious part off their plate — writing the first pass of a reply — and leaves them with the part that actually matters: deciding whether this particular answer fits this particular customer.
When a bot does make sense
It would be dishonest to say bots are never useful. They work well on narrow, well-defined scenarios: checking an order status, giving opening hours, linking to a guide. The key is keeping the bot within limits where a mistake is impossible or cheap, and having it hand the conversation to a human the moment the question steps outside those limits.
Before automating a scenario fully, it helps to check it against a short list:
- Is the answer the same for every customer, with no exceptions or nuance?
- Is the worst-case result of a mistake an inconvenience, not lost money or lost trust?
- Can the question be answered with a single fact, with no reasoning or context from the conversation history needed?
- Is it technically easy to detect when the question steps outside the scenario and hand it to a human right away?
If all four are "yes," the scenario is probably safe to automate fully. If even one is "no," it's better for AI to prepare the answer as a draft and for a human to confirm it. On what to automate and what not to, see the article on the limits of automation.
Common mistakes when choosing between a bot and a draft
In practice, teams don't usually go wrong by choosing a bot over a draft — they go wrong in the implementation details. A few recurring mistakes:
- Automating "just in case." A scenario gets scripted for a rare question, when it would be simpler for a human to just answer it in a minute.
- Letting the product description go stale. If it's out of date, the draft (and even more so a bot) leans on outdated facts — plans that have already changed, features that no longer exist.
- Letting a bot handle financial questions. Billing, refunds, subscription cancellations — these are exactly the topics where the cost of a mistake is highest.
- Ignoring the conversation history. A reply that doesn't account for the fact that the customer already wrote about this yesterday reads as indifferent, even if it's technically correct.
Avoiding these is straightforward: keep financial and non-standard questions with a human, automate routine ones carefully, and always keep an easy path to hand the conversation to an operator.
What a small team should choose
For a small team whose reputation is built on every interaction, AI drafts are almost always the safer start. You get most of the speed win without risking trust or spending weeks configuring bot scenarios.
The short version: a chatbot is worth deploying only where the scenario is narrow and a mistake costs nothing. In every other case — which is most real-world messages — it's more reliable to have AI prepare the draft and a human send it.
That's exactly why Duck Float is built on drafts, not auto-replies: the model prepares a context-aware answer, and the operator stays in charge. Every message — from the Telegram bot, the chat widget, or the public API — lands in one inbox as a ticket with history, statuses, and tags, and live updates show new messages without a page reload. Free during the beta; questions — message us through the widget on this site.
