AI-Assisted Data Mapping
for Distribution Sales
A tool that uses AI to find errors in partners' sales data and suggest fixes. The user reviews each fix and decides.
The project in brief
Partners send sales and inventory files full of inconsistent data. This feature uses AI to compare each file with headquarters (HQ) records and suggest fixes. The user accepts or dismisses each one.
My job was to design how people review those suggestions. The rule: the AI suggests, the user decides, and nothing changes without their approval.
The result: one way of reviewing suggestions, used for both product and partner data. A card handles quick decisions, and a detail panel shows the full reasoning.
Partner files don't match
the company's records
A distribution company works with hundreds of partners. Each one uploads its own sales and inventory files (Excel sheets, CSVs), and the formats change over time. Importing the files is easy. Matching them to HQ records is the hard part.
Naming is inconsistent
One partner calls it 'Lager 330ml,' another 'Lager Bottle Small.' Same product, different file.
Delivery routes are out of date
A partner still delivers to a depot that has closed, or misses a new outlet it should serve.
Cleanup is manual
Teams check every row by hand. That works until a file has 5,000 lines.
One way to review suggestions,
for products and partners
It starts with the upload
The AI reads the uploaded Excel file, finds the records in it, and shows how many columns it matched to HQ fields before review starts.
Suggestions sit on the row they change
Each card is labelled by type ('Rename product', 'Recategorize'), so the user knows the action at a glance.
Nothing extra on correct rows
Rows that already match HQ show 'Matches HQ records' and no card. Attention goes only where a decision is needed.
The change is shown directly
A route suggestion shows the old value crossed out and the new one next to it, so the user sees exactly what would change.
Apply or dismiss on the card
Both buttons are on the card. Accepting an obvious fix takes one click.
Status and suggestion stay separate
A partner can be 'Flagged' and still have a suggestion, or 'Active' with none. The record's status and the suggested change are shown separately.
A quick card for most decisions,
a detail panel for the rest
Most decisions are quick, but some need a closer look. So each suggestion has two layers: a card with enough to decide on the spot, and a detail panel with the full reasoning.
The suggestion card: type of change, old value, new value, confidence and two buttons.
Before and after
The old value is crossed out, with the new value below it. The user sees what would change before reading anything else.
Confidence as HIGH or MED
Each card is tagged HIGH or MED, so the user knows where to look closely and where to move quickly. No percentages to interpret.
Apply or dismiss on the card
Both buttons are on the card. There's nothing else to open.
A detail panel, if needed
The panel shows the uploaded value, the HQ standard, the evidence and the source row. It's optional.
The panel shows the value from the upload, the HQ standard it should match, how confident the AI is and why, and the exact source row. The user can apply or dismiss from here too, but never has to open it.
How I designed it,
step by step
The table was the easy part. The hard question was how much the AI should do on its own, and how to design for the decisions it shouldn't make.
Made the AI suggest, not decide
The original scope was 'automatic data cleanup'. I pushed back, because some calls need a person: is this a new product, or a renamed one? So the AI suggests how to read the data, and the user confirms. This shaped every screen.
The rule I set early: the AI suggests, the user decides. Every screen had to make that clear.
Put suggestions on the rows they affect
A separate 'AI review' screen would split attention between two places. So each suggestion sits in the table on its own row, and the user sees the data and the fix together.
Split each suggestion into a card and a detail panel
Each suggestion answers three questions: what changes, what it becomes, and why. The card shows the first two plus a confidence level, enough for a quick decision. The detail panel shows the why: the evidence, the HQ rule and the source row.
Why two layers: most decisions are quick. The full reasoning is one click away for the rest.
Used confidence levels to speed up review
Each suggestion is tagged HIGH or MED. This shows the user where to focus, and allows a bulk action: 'Apply 4 high-confidence' accepts the safe ones in one click, leaving time for the ones that need a person.
Showed nothing extra on correct rows
If the AI flags everything, people stop reading it. Rows that match HQ only show 'Matches HQ records', so the suggestion column stays mostly empty and gets noticed. 'Needs review' and 'Aligned' tabs let the user filter to what's left.
Reviewed it with stakeholders in rounds
I took the flow through feedback rounds with business and IT stakeholders, the same way the rest of the platform was built. Each round checked one thing: would users trust the suggestions, or just click accept? The feedback shaped the explanation wording and how prominent the confidence level is.
What changed for
the person doing the work
- User checks every row against HQ records by hand
- Errors look the same as correct data, so they're hard to spot
- Results depend on who cleans the file
- Every new partner adds hours of work
- Missed errors end up in reports unnoticed
- The AI checks the file and lists the mismatches
- Each flagged row shows a suggested fix, a confidence level and the reason
- HQ standards are applied the same way for everyone
- Correct rows need no action, so work only grows with errors
- The user makes the final call on every change
“Users no longer search for errors. They review the ones the AI has already found, and make the final call.”
I delivered the design as a clickable prototype, reviewed with business and IT stakeholders and ready for developer handover. The numbers below describe what the design does, not measured results.
What I learned
from this project
The hard part was trust, not layout
I already knew how to lay out a data table. The real work was deciding what the AI could do on its own, and making that clear in the interface.
Two layers worked better than one busy screen
A quick card plus an optional detail panel meant one component worked for both fast approvals and careful reviews.
I'm happy to walk through the decisions behind this project.
Let's talk.
Get in touchSee also the Fondsfinder case study, where I designed from another team's research.