AI drafts, a human signs off: why the approval step is the real value

Is human sign-off just a stopgap?

When AI writes the answers and a person still checks them at the end, it can look like a half-measure. As if you do not quite trust the technology yet, so you bolt a review step in front of it for safety and quietly plan to remove it once the model is good enough. That reading gets it backwards.

At Corresa the human approval is not a crutch you lean on until the AI is ready to answer alone. It is the reason the tool is any use in specialty retail in the first place. The AI does the typing, the first draft, the digging around in your online shop. What it does not do is decide whether that answer is fit to send to your customer. This division is not a setting you switch off later. It is how the thing is built.

The line for it is short: AI drafts, your team signs off. That is not false modesty. It is the more demanding claim.

What does a wrong answer cost in specialty retail?

Work a real case through. A customer asks whether a sleeping bag will be warm enough for a trip to Iceland in September. A fully automated reply says "yes, well suited" and gets it wrong, because the comfort rating was pitched ten degrees too optimistically. The customer is cold, sends the bag back, and leaves a poor review. The return costs you shipping, refurbishment, and the margin on the sale. The review is read by the next person thinking of buying.

Or a customer wants to know whether eight pairs of crampons can be delivered by mid-August, because a trip depends on it. A guessed delivery promise that does not hold costs you no exchange, it costs you trust. Customers remember promises like that.

In advice-heavy retail the answer itself is the product. A wrongly confirmed size, an invented stock figure, a material that does not suit the stated purpose: every one of these errors walks straight into a return or an angry reply. Unlike a question about order status, there is no harmless class of mistake here. Let a machine answer without oversight in this setting and you save a few seconds per email while risking the very sum the transaction was about.

Why review is a quality mark, not a brake

Because it protects the one thing generic systems get wrong: the link between a claim and the evidence for it. Corresa does not draft from memory. For every inquiry it searches your connected online shop and pulls the relevant product data. Anything no shop record can support, such as a binding delivery date, stays open as a placeholder rather than being guessed, and while a placeholder is open the send button stays locked. How that grounding works under the hood is covered at length in the guide to self-hosted AI customer service.

In the review view the person sees both side by side: the customer email on the left, the draft on the right, plus the trail of what the AI searched for in the shop and what it found. That is how they tell "looked up in the shop and backed by data" apart from "a reliable figure is missing here".

So sign-off is not the moment someone dutifully presses a button. It is the moment your team's expertise meets the draft. A member of staff who knows the sleeping bag spots the over-optimistic temperature rating before the customer reads it. That second of expert attention is the whole value. Without it you have a fast system you cannot trust. With it you have a fast system you can stand behind.

Who is responsible for an answer in the end?

A human, and that is deliberate. When your customer gets a commitment, your business stands behind it, not a model and not the software vendor. That responsibility cannot sensibly be handed to a statistic that estimates the next likely word.

So Corresa keeps responsibility where it belongs anyway: with the people who know the customer and the range. The tool takes the writing off your team, not the professional judgment and not the decision when in doubt. It cuts no job. Anyone who advises in specialty retail brings experience no model has, and it is exactly that experience that decides the sign-off.

There is a practical data protection angle too. Because a human decides at the end, there is no automated decision in an individual case that would need separate handling. The reply stays correspondence between people, only prepared faster.

How is this different from an autonomous chatbot?

An autonomous chatbot answers by itself. It takes the question, writes a reply, and sends it without anyone from your business having read it first. For simple, repetitive cases that is convenient and perfectly legitimate. For advice-heavy inquiries it is the wrong shape, because no one catches the one answer that is wrong.

Corresa is explicitly not that kind of chatbot. It does not talk to your customers on its own. It puts forward a suggestion that a human reviews, edits, and takes responsibility for. The difference is not a nicety of operation, it is a question of architecture: with a chatbot the decision falls inside the system, with Corresa it falls to your team. That is also why the tool chooses its customers on purpose. If you want a machine that takes over customer contact completely and automates it away, you are in the wrong place. If you want to answer faster without giving up control of the content, you are in exactly the right one.

And legally: why does this workflow stay outside the disclosure duty?

From 2 August 2026 the transparency duties of the European AI Act apply. They target systems that interact directly and on their own with people, meaning chatbots that answer without human involvement. Such systems have to disclose that the other party is a machine.

Corresa's reviewed workflow does not fall under this on the wording of the regulation, because every answer is approved by a human before it goes out. That is a welcome side effect of the right design, not a trick.

Exactly how Article 50 classifies the case and which documents you need for it is set out in the guide to Article 50 in customer service. The same caveat applies here as everywhere on this topic: this is guidance, not legal advice. How your specific case should be assessed is something to clarify with your data protection officer.

Isn't this the slower route in the end?

The objection is fair: if a human reads every answer anyway, where is the time saved? It sits somewhere other than the question assumes. Your team no longer starts from a blank page. Instead of composing each email by hand, it checks and completes a finished draft that has already gathered the product data. Writing time per answer drops sharply while control stays fully intact. For concrete figures, see the pillar guide.

The time saved does not vanish. It moves to where no language model can go: to the phone, to advising in person, to the customer standing in the shop. In specialty retail that is the part of the work that brings in revenue.

Human sign-off is therefore not a price you pay for the AI. It is the reason the AI is worth having in this business at all. You keep the speed of a machine and the responsibility of a human instead of trading one for the other. If you want to see how that feels in the review view, click through the demo mailbox without signing in.