The context
The agency runs product trial campaigns for brands, with hundreds of participants across several markets. Two tasks weighed on the team with every campaign:
- translating presentations for clients and for each market, including text inside images;
- preparing, for each market, the product shipping file from participants’ addresses, in the expected format.
What we built
- A presentation translator. The application reads every slide, including text inside images through character recognition, then translates it with language models.
- A shipping file generator. From the address file, the application recognizes the columns even when their headings change, checks the data, then produces the file for each market.
- A web application, accessible from a browser, with a detailed user guide.
Testing with difficult files
Real campaign files rarely look like clean examples. Before go-live, we tested the generator with deliberately tricky files: broken encodings, text in several alphabets, duplicates, cells containing formulas. This unglamorous work is what lets the team trust the tool.
The results
- 50% less manual time on these tasks.
- An application the team uses without technical skills.
- A team trained on the tool and its limits.
The fastest gains often come from tasks nobody notices any more, because they have always been done that way.
What we take away
For a small team, AI is useful when it disappears into a simple tool. Nobody is “doing AI”: the team drops in a file and checks the result.
A task comes back with every project and takes hours? Let’s talk for 30 minutes.
Book a callMore notes