XXL ran a 12-week pilot with Riff to see whether AI agents could hold up against the kind of manual work no finance team wants to do by hand. The pilot covered invoice verification first, but it also tested something harder to measure: whether builders across IT, e-commerce, and marketing could ship real software on their own, fast.
The invoice problem nobody had eyes on
XXL processes roughly 40,000 cost invoices a year across six entities. Cost control sat with about twenty approvers, some seeing over 100 invoices a day, with no real capacity to check each one against its contract. Contracts went missing or out of date, indexation errors slipped through, and the growing queue forced approvers to skip checks rather than do them.
An agent that reads the invoice before the approver does
Working with Riff, the finance team built an Invoice & Contract Matching agent that reads each invoice, checks it against the contract database, and posts a one-line verdict directly into the approval workflow before the approver opens it: contract verified, possible discrepancy, or no contract found. The agent informs; it never blocks. Payment flows and human decisions stay intact.
"We go from business problem to production software in weeks, not months, at a fraction of the cost."
What one day of testing surfaced
Run in a single day against one week of historical invoices, the agent processed over 400 invoices. Roughly 9% came back as outright contract breaches, totalling close to 1 million SEK in exposure, and over half had no contract on file at all — immediately surfacing a visibility gap finance didn't know it had.
The invoice agent was only one track
Over 12 weeks, builders across IT, e-commerce, and marketing went from onboarding to shipping production software. One built an AI image-editing tool that replaced an external vendor at a fraction of the cost and turnaround time. Another built an AI ROI hub giving leadership a live dashboard of money and hours saved across every app and agent in the company, with a built-in pipeline for any department to submit and prioritize new automation ideas.
Combined, the registry now tracks 15 apps and agents, 5 already deployed, with a predicted £71,600 in annual savings and over 11,000 hours saved a year.
The harder lesson
Building the software turned out to be the easy part. The harder work was getting clear business ownership and stakeholder alignment in place before a build starts, not after — something that's now shaping how the next cohort is scoped.