AI-Powered P&L Coaching System

The Challenge

A multi-brand restaurant corporation needed a way to coach general managers on financial performance — consistently, every week, across 35+ locations. The existing process was manual and inconsistent: district leaders would review P&L reports when they had time, and the coaching that happened was often generic rather than tailored to each store’s specific situation.

The critical gap was connecting financial data to actionable coaching. Managers had access to dashboards, but interpreting trends, spotting problems early, and knowing what to do about them required expertise that wasn’t available at every location every week.

The Approach

I built an AI coaching system that generates personalized weekly performance reports for every location, delivered by email every Tuesday morning.

The system pulls data from multiple sources via custom MCP servers:

• Financial metrics — period-to-date and weekly sales, labor cost %, food cost %, guest counts, and comp percentages, all compared against budget and district/brand benchmarks
• Customer satisfaction — OSAT survey scores and verbatim customer comments, including specific employee mentions
• Goal tracking — prior week’s goals are automatically evaluated against actual performance

Each week, the AI coach analyzes the data and produces an 8-section meeting outline for the general manager:

1. Period-to-date summary with interpretation
2. Last week’s performance with local market context (weather, events, promotions)
3. Cost breakdown (labor, food, packaging) with status indicators
4. Customer satisfaction trends with highlighted comments
5. Bonus qualification status — exactly where they stand and what’s at stake
6. Evaluation of prior week’s goals (pass/fail based on actual metric movement)
7. Five suggested action items — each with implementation steps, measurement criteria, and expected impact
8. Goal selection template for the coming week

Managers reply to select which goals they’ll commit to, and those selections are tracked and evaluated the following week — creating a closed-loop coaching cycle.

Key technical decisions:

• Multi-model architecture — Claude for analysis and writing (with prompt caching for cost efficiency), GPT-4 for structured goal extraction and reply parsing, Tavily for local market context
• Prefect orchestration — the weekly flow processes all locations concurrently, with a separate flow polling for goal replies every 4 hours
• MCP-based data access — financial metrics and configuration data are accessed through FastMCP servers deployed on Vercel, keeping the coaching system decoupled from the data sources

The Results

The system runs autonomously every Tuesday, generating and delivering personalized coaching reports to every location without human intervention. Each report is tailored to the store’s specific financial position, customer feedback themes, local market conditions, and progress on prior commitments.

The closed-loop goal tracking — where last week’s commitments are evaluated against actual metrics — creates accountability that didn’t exist in the previous manual process. Managers know their goals will be measured, and district leaders can see which locations are executing and which need support.