Handwritten orders from a trade-show booth had to go through several manual steps back at the office — re-typing, sorting by ship date, importing in batches. Wrapping up a big show could drag on for days, and every pair of hands the order passed through was another chance for an error. As PM, I didn't just run with ops' proposal ("a chatbot plus computer-use automation"); I asked one level up: why does the import step exist at all? The answer: customers had no way to order themselves. So I proposed the opposite — customers scan a QR and order, staff import in one click — one form, and the whole problem is gone. AI stayed only in the single step of the flow where it actually fits.
CHALLENGE 01
The company runs booths at physical trade shows, and customers at the booth have always ordered by hand — not because anyone likes handwriting, but because there was no self-service way to order on site, the crowd comes in waves, and staff just take orders the fastest way they can.
You might ask: why not just have customers open the website and order there? That was the first thing I checked when scoping, and the conclusion was that the site genuinely couldn't take it, for three reasons:
So my solution was to give the show its own separate entry point: event-specific prices and bundles, no loyalty or annual-spend accrual, and a ship date that follows the customer.
CHALLENGE 02
After the booth closes, handwritten orders get re-typed into Excel. You might ask again: can't Excel go straight in through the back office's import feature? It can — but there's a manual step stuck in front of it: the ship date. The old import feature set "ship-from date" to the moment of import, so ops had to sort the ship dates by hand, split the whole show's orders into groups, work out the dates and import each group on its own day.
One order: written once, re-typed once, sorted once, batch-imported once. Post-show wrap-up regularly ate several days, and every pair of hands was a chance for error. The new design pulls that manual step out at the root: customers pick a delivery date (or a shipping window) when they order, and import writes it straight into the real ship-date field — ops no longer has to use "when do I import" to engineer a ship date, and the whole batch goes in at once.
CHALLENGE 03
The mood at the company was: the boss wanted everyone to think about how AI could solve their own workflows. Nothing wrong with the direction — plenty of processes really were rescued by AI. So ops' first proposal (they wrote a proper planning doc) looked like this: let customers shop conversationally inside LINE, collect the conversation into a form to replace paper; then use computer use to have AI read the Excel and click through the back office like a human, creating the orders automatically.
In fairness, it was an imaginative proposal and it answered the boss's brief exactly. But looking at the flow diagram I noticed something: every segment of the pipeline was moving data we already had.
As PM, the question I asked wasn't "is this AI plan feasible" but one level up: why does the import step exist? It exists because customers have no way to order themselves, so the data can only travel by paper, by Excel, by human hands. Give the customer an entry point and the whole relay chain disappears — no chatbot needed to collect a form (the form is the form), no AI needed to read Excel (there's no Excel), and certainly no computer use to click the back office (the system just creates the order).
Two more layers down:
The most expensive mistake of the AI era is using AI to automate a step that shouldn't exist. First ask whether the step should exist at all, then ask whether to use AI.
My alternative fits in one sentence: if the endpoint is always an import, build a quick-order feature — customers fill it in themselves, staff import in one click, no AI required. The boss heard it and picked this route.
DESIGN FEATURES
Reuse the existing order-creation code, don't rewrite it — the system already knows how to create orders; we're just giving it a new door. The whole feature is three screens:
01 Customer order page — thirty seconds and you're done.
Mobile first: scan the booth QR, enter the event code, pick a bundle, choose a delivery date (or shipping window) directly. Temperature-zone guardrails live at the source: a bundle that mixes temperature zones simply can't be listed (checked once when the product is entered, again on listing) — instead of discovering at shipping time that the order has to be split.
02 Event settings page — one show, one event.
Bundles, show-only prices, and event start/end dates that directly switch the order page on and off; one shared URL plus a dedicated QR per event.
03 Event order management — money's status never runs ahead of reality.
What customers submit is a draft order, unpaid by default — the existing corporate-import feature defaults to paid, and this deliberately flips it: the show collects payment on site, and ops only switches an order to paid after the card has actually been swiped. Once confirmed, drafts are batch-imported as real orders. A 28-column export fallback is kept alongside — if the automatic import ever breaks, ops can export a file and upload it manually. Automation needs an exit.
During planning I used AI to build two clickable HTML prototypes (the customer order page and the back-office settings page). The meeting wasn't a slide deck — we walked the flow live. Every open question was ticked off against the prototype, and the requirements converged in a single session.
WHERE AI FITS
This project isn't "anti-AI" — the proof is that there's one step where I actively chose AI: orders for FamilyMart frozen pickup. FamilyMart store selection is time-sensitive and has no batch API, so there is no deterministic automation path for this step. For those orders we taught ops to use computer use, letting AI select the store and create each order the way a human would — the only stretch of the whole flow that only AI can automate.
The test was never "AI or not". It's this: deterministic processes go to code, probabilistic judgement goes to AI. This project has both, each in its place.
RESULTS
Ops' feedback was "import time dropped dramatically" — and I didn't want to leave it at an adjective, so I pulled the orders from both shows out of the production database by timestamp and compared them.
Old flow, measured (2025 autumn baby-and-maternity expo, a large show): 171 orders, with import actions scattered across 32 different days, dragging from early October to late December — nearly three months. The cause is the mechanism described above: ship date = import date, so every batch had to wait until its ship day and for someone to remember to import it. The data matches the pattern exactly: each import date's corresponding ship date is almost always the same or next day. And that excludes the post-show work of getting handwritten orders into Excel — ops estimated about two days.
New flow, first outing (2026 World Trade Center vegetarian expo, first time exhibiting, smaller volume): 43 orders, zero handwriting, zero re-typing. The post-show main batch of 36 went in during one afternoon in about an hour, peaking at 25 orders in 5 minutes; orders with different ship dates in the same batch (including one shipping six days out) all went in together; stragglers during the show were imported the same day and shipped the next. Nothing anyone had to "remember".
The two shows differ in scale, so this isn't yet a like-for-like comparison — that comes in October: the 2026 autumn baby-and-maternity expo will be the new flow's first large show, a direct comparison at the 171-order level, and the results will be added here.
(Definitions: both shows' numbers come from production database timestamps; the "about two days" of post-show sorting is ops' estimate.)
REFLECTION
For me the biggest value of this project isn't the feature, it's that it recalibrated something for the team: when the boss asks "how do we use AI", the PM's job isn't to rush an answer, it's to first check whether the problem deserves to exist. Ops later started saying it themselves — "this part just needs a form, that part is where AI belongs" — and that judgement is worth more than any single feature.
STACK
Existing order-creation service reused (no rewrite), automatic temperature-zone detection and order splitting, event draft table, 28-column export fallback, computer use (FamilyMart store selection only). Prototypes were two clickable HTML pages.
I owned flow design, requirement scoping, prototyping and UAT; the three screens were built by the engineering team.