01 · Concept build ·
A faster quote intake for a custom packaging maker
A concept demo that turns a vague packaging request into a checked, rep-ready quote ticket, with a person approving every price.
Problem
A custom packaging business sells boxes made to order online. A customer cannot see a price on the site. Every request, whether it comes by chat, email, phone or the quote form, ends with a sales rep writing back. Competitors already show prices instantly.
Two groups wait. The customer waits for a reply, often while a product launch date gets closer. The rep waits too, because many buyers are new to packaging and do not know the terms. Before a rep can price anything, they have to collect the box style, quantity, size, print and finish, and deadline, usually over several messages.
Public reviews say the same thing in two ways. The specialists are praised by name. The process is called slow.
I did not have access to the company's data, so I could not measure where the time goes. My working assumption is that much of it goes to back-and-forth on the spec and waiting for factory pricing. That is the first thing I would check.
This is a self-initiated concept, built from public information only. I have not worked for this company, and it is not a client.
Decision and why AI
The goal was not to replace the specialists customers like. It was to hand them a complete, checked request on arrival, so their time goes to judgment and pricing instead of chasing details.
When a team wants to use AI, the useful question is not what they could do with it. It is why it has to be AI. Here the answer is narrow. Customers describe what they need in their own words: "nice boxes for my candles", "a couple thousand", "launch in three weeks". Turning that into a clean spec is reading, and reading loose language is the part AI does well. Rules can check minimum order sizes and lead times. They cannot reliably understand how people talk.
So I split the work:
- AI reads and asks. It pulls the spec out of the first message and asks only for what is still missing, one question at a time.
- Rules check. Minimum order size, lead time against the deadline, and jobs that need a specialist are flagged in plain words.
- People decide. A rep approves, edits or escalates every ticket. No price reaches a customer without a person.
- No guessing in silence. If the customer is unsure and the system suggests something, it is labelled "assumed" and the rep is told to confirm it.
- Independence is earned. The system starts by drafting tickets nobody sees. It moves up a level only when its test scores clear a set bar, and moves back down if they drop.
What I built
A working demo on a single page.
- Customer intake. You type as a customer, or pick one of three presets: a skincare brand ordering 300 mailers in three weeks, a coffee roaster with a complete 5,000-carton order, and a vague request for candle boxes. The page shows the spec filling in as you answer, and the flags it raises.
- Rep review. The finished request becomes a structured ticket. The rep can Approve, Edit spec or Escalate. Approve stays locked until the spec is complete. Editing recalculates the flags.
- Test set. Ten labelled cases, including rush orders, below-minimum orders, an unsure customer and an unusual wooden box. You can run it on the page and see each result.
- Autonomy ladder. Four levels, from L0 (drafts tickets in the background) to L3 (instant quotes for standard products). Each level is unlocked only by the test score.
The demo runs on simple keyword rules in the browser. No AI model or company system is connected, and every price shown is a made-up example. It shows how the workflow and the safeguards behave, not how well a real model reads requests.
How I know it works
The demo includes ten test requests with the correct answers written down, and the page checks the system against each one on six points: did it get the spec right, ask only what was missing, flag minimum-order and deadline problems, send hard jobs to a person, and never show a price or hide a guess.
It passed 57 of 60 checks. All safety checks passed.
One case failed. A customer wrote "a couple thousand" instead of a number. The system did not understand it, asked for the quantity again, and sent the request to a person. That is the right fallback, but it is still a miss.
The failure matters for two reasons. First, the gate worked: because extraction scored 90%, below the 95% bar for sending replies without a rep, the system stayed at the level where a rep approves everything. Second, it is the clearest argument for why this needs AI. Keyword rules break on ordinary phrasing like this. A language model is the right tool for that step, and it would need to pass the same tests before it moved up a level.
Ten cases is a small set. It shows how the checks work. It does not prove the system is ready for real customers.
Next
- Measure today first. Time a sample of recent quotes to see where the days actually go: spec back-and-forth, factory pricing or the rep queue.
- Build a real test set. A few hundred past requests, each labelled by reps with the right spec, the questions they would ask, and whether it needed a specialist.
- Run it in the background. Reps work as usual but see the system's draft ticket beside each request and rate it. Nothing reaches the customer.
- Hand reps the first draft. For standard products only, and only once the scores clear the bar. Track how much reps change each ticket.
- Add real pricing last. Instant quotes for standard products come only after the price ranges are tested against what reps actually quote.
The numbers I would watch: time to first quote, rep minutes per quote, and how often a quote becomes an order.
