The Problem
Running a crêpe restaurant sounds simple until you’re watching your staff take phone orders by hand, read them back wrong, and then spend the next five minutes on the phone correcting a ticket — all while the lunch rush is hitting. We were losing 15–20 minutes of productive labor per shift to phone order errors alone. Labor was already our second-biggest expense. Every mistake was costing us food, time, and customer trust.
Beyond the front-of-house issues, I had no clean view into the business. Revenue lived in Toast. Labor hours lived in a spreadsheet. Food cost was a guess I made at the end of the week when I remembered to pull invoices. I was running a $500K+ operation off of vibes and catch-up math.
The Approach
I didn’t want to throw a generic AI tool at this and call it a day. The phone ordering problem needed something that could handle real conversations — customers who change their minds mid-sentence, ask about allergens, or want to add an item they forgot. And the dashboard problem needed to pull from sources that don’t naturally talk to each other: Toast POS, our payroll platform, and vendor invoices.
I mapped the full data flow before writing a single line of code. Where does an order originate? Where does it break? What does a manager actually need to see at 9pm to make a smart decision about tomorrow’s prep?
The Build
The voice AI sits in front of our phone line. When a customer calls to order, the system handles the full conversation — takes the order, confirms items and modifications, reads back the total, and injects a structured ticket directly into Toast. Staff never touches the phone for a standard order.
For the dashboard, I built a unified view that pulls Toast sales in real time, calculates labor cost as a percentage of revenue using actual clock-in data, and flags food cost variance when our weekly invoices deviate from historical averages. It runs on a lightweight backend with a browser-accessible interface — no app install, no training required for managers.
I also automated the payroll submission workflow. Bi-weekly runs that used to take 40 minutes of manual data entry now take a review-and-confirm click.
The Outcome
Order errors dropped 35% in the first month. That number came from comparing refund tickets and re-fire incidents before and after deployment — not a marketing estimate. Phone time per shift dropped from roughly 18 minutes to under 4. Managers started actually using the dashboard within the first week because it loaded fast and showed them what they needed without digging.
The bigger win was getting a real-time view of the business. I caught a food cost spike in week two that I would have missed for three weeks under the old system.
What I’d Do Differently
I’d deploy the voice AI during a slower week, not at the start of our busy season. The first few days had edge cases I hadn’t anticipated — customers asking about catering, complex modification requests, callers who didn’t speak English as a first language. We handled them, but it added pressure. Building in a two-week soft launch with easy fallback to staff handoff would have made that smoother. I’d also integrate vendor invoices from day one rather than adding them in week three — the food cost dashboard is only as good as the data feeding it.