Who signs the quote? Why a person should press send
An AI that writes a price into an email is a time saver. An AI that sends it unchecked is a liability. The skill is knowing which actions need a human and making that approval almost effortless.
One wrong number
A distributor sets up an AI to draft replies to customer enquiries. It works beautifully for a month: faster responses, consistent tone, happy sales team. Then one afternoon the assistant answers a customer asking for a bulk price. It reads a discount table, misreads a column, and writes 'Unit price: 18,000 VND' for a product that sells at 180,000.
The email is sent automatically. The customer, delighted, places an order for two thousand units and forwards the quote to their finance department. The company can honour the price and lose money, or refuse it and lose a client, and either way it will spend a week untangling who is responsible.
An AI that writes a price into a draft saves time. An AI that sends it without a check is a liability.
The lesson is not that AI is dangerous. It is that the cost of a wrong output depends on what happens next, and the design should reflect that.
Not every action needs the same care
A sensible way to decide where people belong is to sort actions by how much damage a mistake could do and how easily it can be undone. Summarising a meeting, suggesting a subject line or tagging a ticket is cheap to get wrong and easy to fix. Sending a legal notice, issuing a refund, changing a price or deleting a record is the opposite.
- Low risk, easy to reverse: let the AI act and review samples afterwards.
- Medium risk: let the AI prepare, and require one click of approval.
- High risk or irreversible: the AI proposes, a named person decides, and the decision is recorded.
Setting these levels in advance, with the business owner rather than only the technical team, avoids both extremes: an agent that can do anything, and one that must ask permission to breathe.
Make approving easier than doing it yourself
Human review fails in a predictable way. If checking an AI's work takes as long as doing it from scratch, people stop checking. They click approve on everything, and the safeguard becomes decoration.
Good review screens do the opposite. They show the proposed action next to the reason for it, highlight what is unusual, and make the common case a single click. In the pricing example, the screen would show the quote beside the standard price range, with the line 'unit price is 90% below list' in red. A salesperson would catch the error in two seconds.
This is also where the AI can help check itself. Simple rules, such as 'flag any discount above 30%' or 'flag any address outside our delivery zones', cost almost nothing and let reviewers focus their attention on the few cases that need it.
Keep the trail
Every action an AI takes on behalf of the business should leave a record: what it was asked, what it looked at, what it proposed, who approved it, and what actually happened. This is not bureaucracy. It is what lets you answer the questions that always come after an incident.
A good trail turns a vague 'the AI made a mistake' into a precise finding: the discount table had changed format on the 3rd, and the parser assumed the old layout. That finding becomes a fix, and a new test. Without the trail, the same mistake simply happens again in a different form.
Trust is earned in steps
The right way to extend an AI's autonomy is gradually, and with evidence. Start with every action approved by a person. Track how often the approver changes or rejects the AI's proposal. When a category of action has run for weeks with almost no corrections, relax the check for that category, and only that one.
Over time, the humans do less routine checking and more of what only they can do: handling the unusual case, improving the rules, deciding what the system should be allowed to do next. A person pressing send is not a sign of a weak AI. It is how a strong one is deployed responsibly.
Keep reading
Why AI pilots stall between the demo and production
The proof of concept was impressive. Six months later it is still a proof of concept. The gap in between is not technical brilliance. It is a set of unglamorous questions nobody asked early enough.
Read articleYour chatbot answers questions. It has never finished a single job.
It is 9:40 pm and a customer wants a quote. The chatbot says 'our team will contact you tomorrow'. Here is why that sentence costs more than most companies realise.
Read article