Lazy Ants

Two production AI workflows, both named, both written up with more operational honesty than this register usually sees.

For Eurekantine, weekly menu planning runs through retrieval before generation: dishes carry structured properties and a vector representation, recently used ones are filtered out, an intermediate step extracts keywords from the manager's request, and only the narrowed candidate set reaches the model. Planning time fell by up to 65%. The write-up states plainly that infrastructure cost was never the tracked metric, despite the case being titled around cost control.

For Get Energy, supplier price sheets in inconsistent formats are parsed into a live pricing workflow behind a verify-before-commit gate. The first release covered one format and one path; more were added only after each held in production. Manual prep per import dropped by up to 75%, and nothing reaches live pricing without human confirmation.

Both cases name their highest-risk failure modes, including "structurally wrong output that still looked plausible", and both publish ownership boundaries down to the roles on each side. A third case is the agency's own internal agent, labeled as its own client rather than dressed up as an engagement.

The weaker fit is a buyer wanting a model built rather than a workflow shipped—this is product engineering that reaches for AI when the workflow needs it, alongside Web3 and e-commerce practices.