CRM Import Module

Context & problem

Bulk import is where CRM onboarding fails.

Constraints

Four rules I held to.

I settled these first, and everything downstream followed:

Fixed destination schema

Fifteen CRM fields. Email and Company Name are mandatory, and Continue stays disabled until a settled mapping fills both.

Nothing writes until the end

Every stage writes to a staging layer in the browser. Nothing reaches the CRM until you press Finalise Import, which is what makes going backwards safe.

One column, one field

Two source columns must not claim the same destination without telling you. A collision you cannot see costs more than a decision you have to make.

Dense data, no type identity

A native system font stack rather than a web font. On a grid this dense, loading one costs a reflow you watch happen. People operate this screen.

Solution

Spend the user's attention where the machine is least certain.

The importer runs in four stages: Upload, Map columns, Review data, Resolve & import. A stepper keeps finished stages navigable.

1. 01

   Confidence decides how much attention a column earns

   I score each source header against the schema on exact match, substring coverage and token overlap, which gives it a confidence from 0 to 100. Above 95 a column maps itself and costs zero clicks. Between 50 and 94 it pre-fills and takes one click to confirm. Below 50 it stays empty and waits for a manual choice. Those three bands become the three tabs, Action required, Needs review and Auto-mapped, so the queue sits in the order of how much you are needed. When two columns compete for one field, the strongest keeps it and the others drop to manual.

   

2. 02

   Review the staged records, not the file

   A paginated pass over the staged dataset, showing rows as the CRM will store them rather than as you uploaded them. This is the last point where changing the shape of the import stays cheap.

   

3. 03

   Quarantine is organised by cause, not by row

   Conflicts fall into four buckets: Missing Email, Bad Date Format, Duplicate IDs and Blocked by Parent. Each bucket carries only the bulk actions that suit it, so you generate placeholders, reformat dates, merge duplicates, keep the newest, or convert to leads. One decision clears hundreds of rows, and you can undo any bulk action as a batch. Email is mandatory, so the flow never offers to import without one. Those rows become Leads, which need no email.

   

Impact

What the prototype shows.

How it changes the shape of the work:

Unblocked user flow

Quarantine holds the broken rows, so the valid ones import straight away.

Reduced cognitive load

Confidence scoring clears the exact matches, so you decide only on the columns it could not place.

Protected backend

Validating in the browser keeps malformed CSVs from reaching the server.

The claim is narrow and testable. Group decisions by cause rather than by row, and the number of decisions falls to the count of distinct causes.