An estimated 75% fewer manager interruptions across 5 locations

Challenge

Managers across five locations were being interrupted all day to answer questions that already had documented answers. How much of this goes in that. What is the prep order for this dish. Which step comes first in the closing procedure. More than 20 of them a day, every day, each one small and each one pulling a manager off whatever they were doing.

Underneath that was a fairness problem. Much of the kitchen staff did not speak English, and the documentation did. So the people who most needed a reference could not use one, and the only way to get an answer was to ask a person, which is exactly the loop the managers were stuck in.

Approach

We built for the person standing in a kitchen, not for the person who owns the documentation. That decision drove three things.

Bilingual from the start, not translated later. Staff ask in English or in Spanish and get an answer in the language they asked in. A Spanish layer bolted onto an English system would have been a second-class experience for the people who needed it most.

The assistant reads the source of truth, not a copy of it. Documentation that has to be manually re-uploaded goes stale, and a stale answer about food prep is worse than no answer. The content had to sync automatically from where the group already maintained it.

Answers, not search results. A cook mid-service is not going to browse. The assistant had to respond to a plain question with a plain answer.

Solution

We built the assistant on Voiceflow, deployed across all five locations, handling recipe and standard-operating-procedure questions in English and Spanish.

Behind it sits an automated sync pipeline in n8n that pulls hundreds of multimedia records from Airtable, the training videos and recipe images the group already maintained, into the assistant’s knowledge base. When the group updates a recipe in Airtable, the assistant knows about it. Nobody re-uploads anything, and there is no separate copy of the documentation to drift out of date.

That combination is the whole design: a conversational front end staff will actually use, fed by a pipeline that keeps it current with zero manual upkeep.

Results

• 20+ staff questions a day answered without a manager.
• Manager interruptions dropped by an estimated 75%. That figure is the client’s own estimate, and we report it as an estimate. Nobody was counting interruptions before, so there is no measured baseline and we will not present one.
• All 5 locations running the same assistant, which is what made operational consistency possible across them rather than per-location habit.
• Non-English-speaking kitchen staff got equal access to the same procedures everyone else was working from.

What made it hard

The content was multimedia, and multimedia is not text. Recipes lived partly in training videos and recipe images, so the sync pipeline had to move hundreds of records of mixed type and keep them addressable by the assistant rather than just present in it.

The harder part was answer quality in two languages against one source of truth. The documentation existed in one language, the questions arrived in two, and a translation that is merely accurate can still be wrong in a kitchen, where terminology is local and specific. Getting that right was iteration on real questions from real staff, not a configuration setting.

The adoption risk was real too. An assistant that gives one unhelpful answer during a busy service gets abandoned, and the fallback (ask a manager) is always available. It had to be faster and clearer than interrupting someone from the first week, or it would not have been used at all.