A named person approves every reply before it sends, and we think that constraint is the most valuable thing we ship.
Most AI support tools are sold on a single promise: fewer humans. Tickets close themselves, replies send themselves, and the headcount line goes down. We built SoftReply the other way around. Our AI drafts every reply, but a named person on your team approves it before it reaches the customer. We call this reviewed AI, and the human veto is the point, not a training-wheels phase we are trying to outgrow.
A wrong draft costs nothing, because nobody sees it. A wrong send costs a refund, a compliance exposure, or a customer who no longer believes a word you write. The economics of support are asymmetric: the downside of one confidently incorrect message dwarfs the time saved across a hundred correct ones. Reviewed AI puts the human exactly where the risk concentrates, at the moment of sending, and nowhere it does not need to be.
This is also why the AI drafting the reply is genuinely useful even though it never sends on its own. It reads the thread, pulls the relevant facts from your knowledge base, and writes the first version. Your agent stops being a typist and becomes an editor. Most of the work, the reading and the recall and the phrasing, is already done. The judgment stays with the person whose name is on the message.
Customers can tell when they are talking to a system that has no accountability behind it. A reply that a real person stood behind reads differently, and increasingly, buyers ask. Procurement questionnaires now include "does a human review AI-generated responses." Being able to answer yes, and to show the record, shortens sales cycles and survives audits. Honesty about what the machine does and does not do is not a marketing position for us. It is the product.
Every sent reply in SoftReply carries its provenance. We record who sent it, and when the AI's draft goes out unedited we keep the knowledge-base fact it cited by number and the model's confidence at draft time. Edit the draft and it is recorded as what it now is — that person's own words. When a customer disputes what they were told, you are not reconstructing a memory. You open the conversation and read the exact fact the answer rested on and the name of the person who agreed it was right. For teams in regulated or trust-sensitive markets, this trail is often the difference between a defensible process and an anecdote.
We do let mailboxes opt into higher autonomy, per mailbox, deliberately. But we treat it as an exception with a narrow shape. It fits a specific class of message: high volume, low variance, low stakes, and well covered by cited facts, such as "where is my receipt" or "how do I reset my password." Even then we suggest starting with review on, watching where the drafts and the approvals agree over weeks, and only then loosening the gate for that narrow lane. Refunds and anything legal or security-related are not a matter of advice: they are locked to a human in the product, and no combination of settings will auto-send them. If a vendor tells you their AI can safely run your whole inbox unsupervised, ask them who signs the refund.
The industry keeps framing the human as a cost to be removed. We think the named human approving the send is the most valuable signal in the entire system, the one thing that tells a customer, and an auditor, that someone is actually accountable for what was said. We are not trying to remove that. We are trying to make it fast.