Upload a file of support conversations. UpstreamW protects personal data, sorts every case into a customer pain category and its root cause, and summarizes what customers are telling us. People read the results and write the recommendation.
Start
Upload a file of support conversations
A JSON file with a list of cases. Each case needs a text field; id, market, channel and date are used when present.
No file chosen
Step 1
Protect personal data before anything leaves this computer
Patterns find data with a known shape: emails, phone numbers, card numbers and card endings, account numbers, tax IDs (RFC, NIT, CPF, CNPJ) and amounts.
A small AI model running on this computer finds names and addresses, which have no fixed shape.
Each item is replaced by a label such as [NAME_1]. The real values stay in a local file.
A final check looks for anything left. If it finds something, that case is not sent and goes to a person.
Before and after, every case (highlighted: what was found and hidden)
Step 2
Sort each case into a customer pain category and a root cause
An AI model reads only the protected text from step 1.
It picks one customer pain category from a fixed list. It cannot invent new ones: anything new is marked Emerging for a person to name.
It scores the root cause across the 4Ps, shown here as shares of 100%: People (how the agent handled it), Process (internal steps and handoffs), Policy (rules and limits), Product (the app or platform).
Sensitive cases (fraud, locked accounts, legal) and unsure answers go to a person. The rules live in a settings file that can be changed during the session.
The fixed list of categories
Next step for each case
Auto-tag confident, counted as is. Review a person checks the label. Emerging a person names a new category. Human-led a person handles it.
Every case (click a row to see why)
Step 3
What these customers are telling us
Summary at a glance
Key findings (written from the numbers by fixed rules, not by the AI model)
Summary by category (largest first; click a category name to jump to its detail)
Cases that fit no category (a person reads each one and decides whether it needs a new category)
Root cause across all cases (average share of the 4Ps)
Cases by market (highlighted where a category is much more common in that market than overall)
The deeper analysis
Root cause by category (average share of the 4Ps; the largest in each row is highlighted)