Fooodo Insights

Fooodo Insights is a separate product track from the Fooodo ordering platform. The website surfaces it because it is the strategic direction for the company; the working product is in active development and not yet shipped — this page covers what Insights is, what it explicitly isn't, and the architecture that sits behind both.

Status: Phase 0 — internal validation against Čili-shaped data is the current focus. Design-partner chains are being onboarded through 2026. The MVP scope and timeline are at the bottom of this page; if you run 5–200 locations and want to shape the rollout, write to hello@fooodo.com.

The one-line definition

Fooodo Insights is an AI-native decision intelligence layer for multi-location restaurant chains, designed to connect operational actions to EBIT impact and to keep humans firmly in the loop on every decision that affects employees.

It is not "another dashboard." A dashboard tells you what happened. Insights is a system of decision that:

1. Collects fragmented restaurant data.
2. Normalises it into a canonical restaurant operating model.
3. Explains what happened.
4. Estimates why it happened.
5. Predicts what will happen next.
6. Recommends actions.
7. Requires human approval before high-impact actions are executed.
8. Measures whether approved actions actually improved EBIT.

Who it's for

Multi-location chains in the 5–200 location range. The early reference deployment is a Čili Pizza-shaped chain, but the architecture is POS-agnostic, geography-flexible, and adaptable to different formats (QSR, casual dining, ghost kitchens, hybrids).

The personas Insights is built for are executive, not operator:

The Restaurant Manager and CHRO personas are why Article 22 GDPR matters structurally — they are the audience for any recommendation that touches a person, and the approval queue is built so a human is always in the loop on that path.

The principle that makes this different

Every insight, recommendation, forecast, and scenario in Fooodo Insights connects to a measurable EBIT driver:

• Revenue
• Gross margin
• Food cost
• Labor cost
• Marketing ROI
• Table turnover
• Average check
• Order volume
• Waste
• Discounts
• Delivery platform economics

If a recommendation cannot be tied to one of these drivers, it does not enter the action queue. This is the rule that separates Insights from generic "AI optimises everything" copy.

Architecture (six layers)

The product is structured as six layers, each with a defined contract — data flows up; recommendations flow down through the approval queue:

Layer 6Application

Executive cockpit · location dashboards · campaign ROI · profit-target planning · scenario simulator · approval queue · alerts · reports · admin · API

Layer 5AI and agents

One AI · four role views (CFO, CMO, COO, CPO) · conversational AI · insight generation · recommendation engine · explanation engine · human-in-the-loop approval queue

Layer 4Metrics and modelling

EBIT model · revenue · COGS · labor · marketing ROI · price elasticity · forecasting · anomaly detection · attribution · scenario engine

Layer 3Canonical data model

Tenant · brand · region · location · menu · order · order item · payment · discount · tax · campaign · labor shift · employee role · supplier cost · recipe · forecast · recommendation · approval · measured outcome

Layer 2Ingestion and importers

Direct API connectors · CSV/Excel import · scheduled imports · validation · mapping UI · data quality checks

Layer 1Data sources

POS · labor scheduling · time tracking · accounting/P&L · delivery platforms · marketing campaigns · loyalty/CRM · weather · holidays · manual CSV · food cost · recipe data · menu data

Two architectural rules matter most:

• You don't need a data warehouse. CSV/Excel upload is available from day one. Mid-market chains rarely have a working warehouse, and waiting on one was the failure mode of the previous generation of restaurant BI.
• POS connectors are pluggable. Each implements the same connector interface (locations, menus, orders, payments). Adding a connector is an isolated module; it does not touch the metric or AI layers.

Role-specialised agents

Insights runs role-specialised AI agents — CFO, CMO, COO, CPO — on a shared model. Each answers in the vocabulary of the role asking, with permissions and approval routing scoped to that role. Each agent monitors a specific domain and generates EBIT-tied recommendations. None of them executes — every recommendation enters the human approval queue, and any employee-affecting recommendation is gated by GDPR Article 22 (see below).

Labor and scheduling touch employees, so they get the strictest routing: any recommendation that would change a person's hours, schedule, or performance flag is gated by GDPR Article 22 — it requires a named human reviewer before it propagates anywhere, regardless of which agent surfaced it (see Article 22 below).

The EBIT impact engine

The core analytical capability. It connects operational events to EBIT movement and explains the connection in plain language.

What it does:

• Breaks down EBIT change by location, time period, menu category, campaign, labor, food cost, delivery channel, and discount.
• Provides variance analysis: actual vs previous period, vs budget, vs forecast, vs reverse P&L target.
• Separates correlation from estimated causation from confirmed measured impact.
• Assigns a confidence level to every explanation.

Example shapes (illustrative, not real-tenant data):

"Location X's EBIT fell by N euros yesterday. M% of the decline is explained by labor hours exceeding target by P%, combined with a Q% drop in average check."

"Campaign C increased revenue by N euros, but after discounts and extra labor, EBIT impact was +M euros."

"A P% price increase on category K likely reduced order volume by V%, but increased contribution margin by N euros."

These are not chatbot guesses — they are computed against the canonical data model with explicit confidence intervals.

How figures earn trust

Three mechanisms already built into the product back those claims:

• A five-label confidence taxonomy on every figure — the provenance layer behind the confidence levels above. Each number in an analytical answer is labelled Faktas (actuals), Biudžetas (plan), Patvirtinta (an admin-confirmed organisational fact), Apskaičiuota (machine-computed), or Prielaida (a declared assumption) — the labels shown verbatim, as they appear in the product. Labels are deterministically audited against the actual provenance of the underlying data — the AI cannot self-assert a stronger label than the source supports.
• A Methodology Pack per tenant. Each tenant's admins curate board-grade facts — employer-cost conventions, incremental margins, plan figures — on an admin-only methodology page. Every fact carries a value, unit, named source, valid-from date, and confidence label (capped at 60 facts per tenant; changes apply to new conversations). The AI applies the house methodology instead of generic industry knowledge and cites the source.
• A CFO answer contract. Profit-impact analyses are anchored to plan and gap-to-plan, decomposed into levers with their binding constraints, and every analysis carries a sources line.

Human-in-the-loop approval queue (GDPR Article 22)

This is the load-bearing constraint, not a feature.

GDPR Article 22 restricts solely-automated decisions with legal or similarly significant effects on individuals. In a restaurant context that means anything that affects employees — schedule changes, hour reductions, performance scoring, automated penalties — must be reviewed and approved by a human before execution.

Every recommendation Insights produces enters an approval queue showing:

• The recommendation itself
• The reason
• Expected EBIT impact
• Confidence
• Risk
• Required approver role
• Data sources used
• Suggested implementation steps
• Expiry / urgency
• Approve / reject / modify / defer options

Approved actions are tracked as experiments, so the team can later measure whether the action achieved the expected EBIT result. (Closing that measurement loop fully — feeding outcomes back into ranking automatically — is later-phase work, not the MVP; see below.)

Employee-affecting recommendations specifically — schedules, hours, performance flags, retention measures — never execute autonomously. They sit in the approval queue with a named reviewer, audit logs, and explainability before propagating anywhere. This is a legal requirement under GDPR Article 22, not a product preference.

Marketing ROI engine (statistical, not before/after)

Campaign measurement is one of the places restaurant BI tools fail most visibly. "Revenue went up during the campaign" is not causation.

Insights measures campaigns with proper statistical testing — group comparison with unequal variance, effect-size calculation, and an explicit confidence verdict — combined with a financial materiality breakdown: revenue uplift, gross margin uplift, discount cost, extra labor cost, food cost impact, and net EBIT impact. Every result is tagged either "statistically significant but financially irrelevant" or "financially material but statistically uncertain" so a CFO can read the verdict without parsing the math.

The recommendation can be: scale the campaign, repeat only in specific locations, stop, adjust discount, change timing, change product mix — with the EBIT-tied reasoning attached.

Reverse P&L planning

CEO/CFO inputs a target annual EBIT. The system works backward from it — the basic version in the MVP back-solves the required turnover, by month with seasonality; the fuller decomposition below is the target state:

• Monthly and daily revenue targets per location
• Allowable labor hours and labor cost %
• Theoretical food cost %
• Required average check, order volume, campaign uplift
• Required price/mix changes

Then it compares actual performance against the path and alerts when a location is drifting. Example shape (illustrative):

"Target EBIT N/year. Location A must average X/day revenue, labor below Y%, food cost below Z%. Current forecast misses target by P unless average check increases by some % or labor reduces by some %."

What's in the MVP, what's later

MVP:

• CSV/Excel import
• Canonical data model
• Location dashboard + executive cockpit
• EBIT calculation and variance analysis
• Pre-generated AI insights at import time
• Manual campaign measurement with statistical testing
• Reverse P&L (basic)
• Human approval queue
• AI proposing in CFO and CMO views
• Data quality checks
• Role-based access + audit logs

Deferred to later phases:

• Full POS API integrations beyond the first
• Advanced price elasticity modelling
• Camera-based dish quality scoring
• Complex causal inference
• Fully automated execution (deliberately not on the roadmap; see Article 22)

POS and data sources Insights plugs into

Insights is POS-agnostic by design. R-Keeper is the live reference integration — ingested through a managed data pipeline today, not a generic REST connector; additional POS connectors (Toast, Square, Lightspeed in Western markets; iiko in Eastern Europe; others) are scoped per-customer and quoted on demand. Until a real-time connector is commissioned, the CSV/Excel importer covers any POS the chain runs.

Beyond POS, Insights plugs into delivery platforms (Wolt, Bolt), accounting systems, labor scheduling tools, marketing platforms, weather/holiday data, payment providers, and Fooodo's own ordering and payment system.

Connect via MCP

Every Insights tenant ships a remote Model Context Protocol server. AI clients that speak MCP — ChatGPT, Claude, Copilot, Gemini, Cursor, and others — connect with OAuth 2.1 and Dynamic Client Registration, no API keys or custom integration project required. See Insights MCP server for the tool inventory and per-client setup steps. The feature is dark by default per org; enablement runs through support@fooodo.com.

Status and timeline

The working product is not yet shipped — the marketing surface at /insights describes the vision; phase 0 (internal Čili-shaped data validation) is the current build focus. Expect substantive product news through 2026, not before.

If you are a multi-location chain interested in being a design partner, write to hello@fooodo.com. The fastest way for the team to prioritise a capability is to have a customer with the data and the appetite to validate it.