AI-assisted ordering: the AI drafts, the sales rep decides
A feature in the field sales app that turns photos, PDFs, voice notes and messages into a draft order to review, cutting order entry time by 30%.
At a glance
Context
A tablet app that sales reps use to take orders from customers, during visits or remotely.
Problem
Orders arrive as photos of notes, PDFs, voice or text messages. Re-typing them item by item takes time and invites mistakes.
My role
Sole designer: I gathered the requirements and designed the feature, from uploading the file to reviewing the items.
Team
Me, two developers, a project manager working with the customer, a product lead.
Timeline
About two weeks of design, then review and testing. In production since March 2026.
Outcome
Sales reps start from a ready-made draft and check it, instead of entering everything by hand.
−30%order entry time
4input types: photo, PDF, audio, text
2weeks of design
Some references, such as names, data and brands, are made up or left out because they’re covered by non-disclosure agreements (NDA).
The problem
Sales reps don’t get orders in a clean format. The customer snaps a photo of their notes, sends a PDF of items to restock, records a voice message or types a list on WhatsApp. Then the rep has to copy everything into the app, item by item.
It’s slow, repetitive work, and every manual step is a chance to get it wrong: the wrong code, a misread quantity, one variant mistaken for another. On an order, a mistake lands on both the rep and the customer.
Constraints
A tablet, so little room to show several things at once.
One file at a time: several files can’t be uploaded together.
Connection required: without internet the AI doesn’t run.
The manual order flow stays as it is: the feature had to add to it, not replace it.
Two weeks of design.
No sales reps in testing: the customer had no time to spare, so testing was done with their IT manager.
How I worked
I gathered requirements and technical constraints with the developers, then built a working prototype of the order flow with Claude Code, with the AI built in. We showed it to the customer: it had a real wow effect and validated several decisions before development.
Pick the filePhoto, PDF, audio or text, one at a time
The AI processes itA loading dialog; no connection, no start
ReviewFile and items side by side, each with a confidence badge
OrderItems move into the existing manual order
The flow, from file to order. Sales reps stay in the flow they already know: the feature adds to it rather than replacing it.
Key decisions
1. The original file next to the list
Decision
After processing, the rep sees the recognised items on one side and the original file on the other: the photo, PDF, text or audio.
Ruled out
Showing the list only. On a tablet there’s little room for two columns, and dropping the comparison was a real option.
Why
Without the file in view, the rep has nothing to check the AI’s work against. I pushed for the side-by-side comparison, and it was the most appreciated choice in the prototype demo.
2. Items in the same order as the file
Decision
The list follows the order in which items appear in the file.
Ruled out
Sorting them in other ways.
Why
The rep scans file and list together, line by line. That makes checking faster.
3. Confidence in words, not numbers
Decision
Every item gets a badge: “Reliable”, “Partly reliable” or “Not recognised”. When the rep replaces an item, it shows “Edited”.
Why
Users want to be able to check everything, yet they pay little attention when they do. The badge draws the eye only to what matters, and says it in words rather than relying on colour alone.
✦AI order[Code] [Company name]
+ Add item✕ Cancel→ Next⋮
‹
6 tomato passata
4 mozzarella
2 flour 00
1 EVO oil 5 l
3 spaghetti no. 5
2 moka coffee
1 c. salt
AI can make mistakes. Check the information entered.
Wireframe of the review screen, with made-up content: the notes and, next to them, the items in the same order, each with sales unit, quantity and confidence badge. At the bottom, how many items the AI detected, proposed and couldn’t find.
The review, step by step
01 · File
The customer sends their notes: a photo, a PDF, a voice message or text copied from WhatsApp.
02 · Reading
The AI reads the file and fills in the items in the same order, so the rep checks file and list together, line by line.
03 · Confidence
Each item has a badge that says in words how far to trust it: “Reliable”, “Partly reliable”, “Edited”.
04 · To check
Whatever the AI doesn’t recognise stays highlighted, to search for or add by hand.
05 · Summary
At the bottom, the summary: how many items the AI detected, proposed and couldn’t find.
06 · Order
The rep confirms, and the items move into the manual order they already know.
4. The AI drafts, the rep confirms
Decision
The AI fills in most of the items. Unrecognised ones are highlighted so the rep can search for them or add them by hand, and an item that wasn’t in the file can always be added on the fly. After review, everything moves into the manual order the rep already knows.
Why
The rep is still responsible for the order. The AI should save them time, not take control away from them.
5. A working prototype instead of a mockup
Decision
I built a prototype of the flow with Claude Code that actually works, with the AI built in.
Ruled out
Static mockups.
Why
With AI, behaviour matters as much as looks: waiting, half-recognised items, corrections. A working prototype shows all of that.
What it meant
In the demo the customer saw the value of the feature straight away, and some choices, like the side-by-side comparison, were validated before development.
The solution
Four input types: photo, PDF, audio and text copied from WhatsApp.
Processing: a loading dialog while the AI works; without a connection the feature doesn’t start.
One file at a time: while a file is being processed, picking another is blocked and a label explains why.
Review: file and list side by side, items in the same order, confidence badges, unrecognised items highlighted, items added on the fly.
Confirm: items move into the existing manual order.
Results
−30% order entry time. We measured it in a test with the customer’s IT manager: the same order, from a photo of an item list, entered first by hand and then with the feature. The time includes AI processing and checking.
In production since March 2026 for the customer, with positive feedback.
On its way to the standard app: the feature is now being built into the version for all customers.
What I learned
With AI, control is a paradox: users always want it, yet exercise it with little attention. So the interface shouldn’t ask them to recheck everything, it should point clearly to where to look.
Next time I want to test with real users, the sales reps in the field. This time the customer couldn’t spare the time, and an IT manager doesn’t work under the same pressure as someone taking an order between two visits.