What is human-in-the-loop quoting?
Human-in-the-loop quoting is an automation pattern where AI drafts the quote and a person approves it before anything reaches a customer or the ERP. The system reads the RFQ, matches line items against the item file, and applies the customer's price levels; the human reviews the draft in an approval inbox, fixes what is wrong, and sends. Every edit is counted as the Human Edit Rate, so trust becomes a measured number instead of a feeling. As the rate falls, review gets lighter, stepping from Assist to Guarded to Autopilot, with the human in command at every rung.
Why an approval inbox instead of a black box?
A black box reads the RFQ and sends a quote with your name on it. When it misreads a line, you find out from the customer, or from the order file, weeks later. An approval inbox reverses that: the AI does the reading, matching, and pricing, then queues a draft that shows exactly which SKU it matched, which price level it applied, and which lines it could not resolve. The human approves, edits, or rejects. Nothing reaches a customer or the ERP without that decision.
The human stays in command because the ambiguity is real. Distributors we talk to see a request for a 10 watt light map to 50+ catalog variants. A system that guesses silently turns that ambiguity into wrong orders. A system that flags it turns it into a quick call by the person who knows the customer, the stock, and the job.
How does the Human Edit Rate turn trust into a number?
The Human Edit Rate is the share of drafted quotes a person has to correct before sending. Every approval and every fix in the inbox feeds it, so the question of whether to trust the AI stops being an argument and becomes a chart. A falling rate is earned trust, line by line, customer by customer. A spike is a warning you catch in the inbox, not in a returned order.
Estimators we work with who tried AI before say the same thing: a draft you cannot trust is slower than starting clean. The edit rate is the honest answer to that objection. It shows, per desk, whether review is getting lighter or the tool is wasting time. If the number does not fall, you should not automate more. That is the point of measuring it in the open.
How do Assist, Guarded, and Autopilot build the trust ladder?
Assist is the starting rung: the AI drafts, a person reviews every quote before it ships. Guarded is the middle: routine quotes flow while anything flagged, an unmatched SKU, an unresolved price, an unfamiliar customer, holds for review. Autopilot is the top rung, earned per category rather than declared: quotes the system has proven it handles go out under rules you set, and every one still lands in the same inbox with a full audit trail.
The ladder moves when the number says so, not the calendar. You do not schedule Autopilot for week three; you graduate a category when its Human Edit Rate stays low, and one setting drops it back to full review. The human is not being phased out. The review is being right-sized: full attention where the AI is unproven, spot checks where it has earned its way up.
Related questions
Is human-in-the-loop quoting just slower automation?
No. The slow part of a quote was never the decision, it was the assembly. Distributors we talk to average about 30 minutes of desk work per quote; reviewing a fully drafted quote in an inbox takes a fraction of that, and the fraction shrinks as the Human Edit Rate falls.
Does Autopilot remove the human?
No. Autopilot changes review depth, not command. Quotes flow for categories the edit rate has proven, everything still lands in the approval inbox with an audit trail, and one setting returns any category to full review.
What happens when the AI is not sure?
It flags instead of guessing. An unmatched line, an ambiguous item, or a price it cannot resolve surfaces as a question in the inbox, never as a silent default in a sent quote.
See it on your own work
Distributors: two steps and a kickoff call. Estimators: upload a plan on a live trade. Either way, the product proves it or it does not.