A margin copilot for restaurants — one demand model that decides what to order today, and what to charge next (roadmap). It earns the right to act on its own under a floor it monitors itself, keeps the judgment when your staff turns over, and gets sharper every week from what actually happened. It reads your POS, invoices, suppliers, weather and events.
Not another ordering app, inventory ledger, or forecaster — the decision layer that moves food-cost % from both sides.
food-cost %
the one number — now from both sides
18–22×
both-sides ROI at steady state (modeled)
earns & pauses
autonomy under a monitored floor
survives turnover
judgment stays when staff leave
The problem
Restaurants run on a few points of margin — and they leak from every side
Food cost is the largest controllable line, in an industry that can't raise prices any further. The leak isn't one thing; it runs along eight structural fault-lines unique to perishable, demand-volatile, thin-margin food service — each answered by a pillar. Turnover and messy data, the engine handles.
+34%
food costs vs pre-pandemic
94%
of operators cite food cost as their #1 concern
3–5%
typical net margin · 42% made no profit in 2025 (NRA)
#1
staff turnover — the top operational challenge
The NRA puts food cost at 28–35% of revenue and labor near 28–32% — food + labor alone eat ~62% of sales before rent, cards or energy. Just 3 points of food-cost control ≈ $15–45K/yr recovered on a $1.5M operation — without touching menu prices or portions.
Volatile local demand
Covers swing by day-of-week, weather, events — static pars miss it.
→ Covers Forecasting
Perishability
Over-order → spoilage; under-order → 86'd item, lost sale.
→ Build-to-Par Ordering
Invoice & price drift
Pack-size, substitution, line-item creep — caught only at the P&L.
→ Receiving & Invoice Integrity
Inventory shrink
Theft, over-portioning, unlogged transfers — the box holds less than the book.
Cashback & buying-group rebates most independents never claim.
→ GPO & Rebate Capture
Thin margin, pricing power
Everything resolves to food-cost %; the menu price is its other half.
→ Menu Pricing (roadmap)
Labor / prime cost
Labor ≈ the other half of prime cost; schedules miss the demand curve.
→ Labor & Schedule-to-Forecast (roadmap)
Industry figures are indicative ranges from Restaurant365 (2026 midyear), TouchBistro and MarginEdge.
What it costs today
The quiet leaks dwarf the one-off disasters
Each fault-line, sized at the $1.5M reference restaurant — the everyday bleed, not the headline blow-up, and all of it from the same missing judgment.
$17–45K
waste & spoilage · Build-to-Par
$10–22K
invoice overcharge · Receiving
$5–18K
inventory shrink · Inventory Truth
$8–20K
recipe variance · Theoretical-vs-Actual
$5–15K
uncaptured rebates · GPO & Rebate
Modeled, $1.5M restaurant (≈$450K food spend). Overlapping exposure, not an additive total — the cost-side model nets ~$25–55K/yr recoverable (see The outcomes). No production customers.
What it is
The decision layer — not the tools, or the platforms, it's mistaken for
Underneath all of them, the difference is one thing nobody else has: a model that learns from your verified outcomes and earns governed autonomy.
Capability coverage across the 2026 field — filled = built-in, half = partial, open = not offered; the bottom three rows are the part no one else crosses.
The product
One demand model — order it, price it, staff it
Every pillar hangs off one demand model learned per kitchen. The cost side runs in the demo today; menu pricing and labor are the roadmap twins off the same model; the engine is why a thin-margin kitchen says yes. live = in the demo today · near = in build · roadmap = next.
The demand model — learned per kitchen
learns your kitchen
Covers Forecasting
Covers & mix by daypart — from weather, local events, reservations and day-of-week. The one model that drives what you order, charge and staff.
live · core
Cost side — what you buy
order right
Build-to-Par Ordering
Order / more / less / skip per category; par moves with the forecast — over-order spoils, under-order 86s.
live
receive right
Receiving & Invoice Integrity
Scores each invoice line against your contract and the learned norm — market-move vs supplier markup, overcharge and short-ship caught, credit claimed. Line extraction is a 3rd-party OCR (integration on the roadmap); the scoring runs in the demo on parsed lines.
live
no shrink
Inventory Truth & Shrink
Perpetual on-hand vs the physical count — the box holds less than the book, and the gap is theft, portioning or transfers.
live
find the leak
Theoretical-vs-Actual Variance
Sold 100 burgers = should've used 50 lb beef; you used 58 — the recipe, portioning and yield leak, item by item.
near
capture rebates
GPO & Rebate Capture
Manufacturer cashback and buying-group rebates most independents never claim; off-contract drift caught.
near
Invoice-line capture is a commoditized layer — we integrate a third-party OCR / AP tool for extraction (roadmap) rather than rebuild it. Restaurant-RGI's value begins after the lines are parsed: contract & rebate compliance and market-vs-markup decomposition.
Revenue & labor — the roadmap twins (off the same model)
price right
Menu Pricing & Discounting
Price the menu and time promotions off the same demand model — stars vs plowhorses. Demand-aware pricing lifts margin 2–5% (modeled).
roadmap
staff right
Labor & Schedule-to-Forecast
Staff to the same covers forecast — the other half of prime cost (food + labor ≈ 60% of sales), the #1 operator pain.
roadmap
Further out: third-party delivery margin — platform commissions run 15–30% per order, a separate fault-line from what you buy and sell. A future pillar brings channel mix and commission-aware pricing onto the same demand model.
The engine — why a kitchen says yes
it survives turnover
Continuity / Judgment Memory
The learned judgment behind every pillar persists across every GM, chef and coordinator change — it's in the engine, not a head.
near
it earns trust
Earned Autonomy & Self-Governance
Auto-acts above a floor it monitors; pauses its own authority when decisions drift.
live
it proves itself
Defensible Dollars / Proof Ledger
A ledger that subtracts what you'd have caught anyway — defensible dollars, including the zero-recovery weeks.
live
The outcomes
The value compounds — on both sides of margin
Food-cost % moves from both sides: what you buy — the cost-side pillars (live + near) — and what you charge — menu pricing on the same demand model, roadmap. As the engine learns, the two compound and pull away from a generic tool that decays: 18–22× at steady state (both sides, modeled). Labor — the other half of prime cost — is further roadmap upside off the same model.
Cost side — buy right (live + near)Menu pricing — price right (roadmap)a generic tool (decays)
Modeled, $1.5M restaurant, Group tier ($349/mo). Multiples = both-sides value ÷ annual subscription; steady state = after the maturity ramp. Cost side live; menu pricing roadmap. No production customers.
The moat
What you can do with Restaurant-RGI that you can't do with others
Your ordering system remembers what you bought — we remember what you learned. Static tools run Day-365 logic identical to Day-1; Restaurant-RGI calibrates to your suppliers, your kitchen, your demand — and earns governed autonomy. Quantified where modeled; otherwise a structural gap.
“Autonomy” here is operator-side and governed — the copilot acting under a floor it monitors, on your buying and pricing. Not the guest-side “agentic commerce” (AI agents ordering for diners) of the 2026 headlines.
Auto-order under a floor that revokes itself
Others set fixed rules or only recommend — none earn or pause their own autonomy.
~35–48% auto-decided at a monitored floor (modeled)
Keep the judgment when the GM or chef leaves
Others store order history and cost — not the learned judgment: that the salmon supplier pads 8% every November, that you need 40% more protein the week before Thanksgiving. The best buyers are 10× better — but one person; when they leave, it leaves.
~$28K lost in the 90 days after a 9-yr buyer left — de-trended vs USDA PPI (modeled)
Prove the learning is real, not just more data
Others can't isolate it from everything else changing.
frozen-twin: waste 6.2% → 3.8% vs 5.9% (simulated)
Market move vs supplier markup — checked against your contract price
Others track what you paid — not why, nor against your contract at the decision.
overpayment + rebate leakage caught
Surface the trust trap in your supplier signals
Others can't — it only shows up by learning from your own verified outcomes.
signal reliability on your own orders
Defend every claimed dollar, down to the $0 weeks
Others give a marketing ROI estimate.
incremental-attribution gate
Every verified decision sharpens the next — the gap against static systems widens, never narrows.
Where the market sits — and the corner no one occupies
market toolsRestaurant-RGI
Record-keeping suites and ordering apps sit bottom-left (same logic Day 365 as Day 1); forecasters learn but only recommend; one agent drafts POs but doesn't learn. Even Toast IQ — the POS platform that owns the data — records, answers and acts on request, but by its own 2026 docs tracks no supplier price, forecasts no orders and keeps no judgment. The learns-and-earns-autonomy corner is unoccupied. Across the 2026 landscape surveyed, no tool is described as learning from verified outcomes or keeping institutional knowledge across turnover.
The same week — before & after
Five fault-lines, five fixes
The same leaks the fault grid names — each caught by its pillar, with the engine doing the work underneath. The loop under each: ① context · ⑧ gate · ②③ act · ④ memory. Illustrative; modeled at a $1.5M restaurant, no production customers.
The Saturday the Weather Turned
Volatile local demand → Covers Forecasting · live
92° and a street festival meet a spring par. Without: the fixed par over-orders braised proteins and 86s the cold plates by 7. With: the forecast pulls proteins down, cold & shareable up — waste-checked before it adds.
① context → ⑧ gate → ② act → ④ memoryweekend waste 6.2% → 3.8%~$1.8K / weekend
Your 11-year distributor quotes +14%, "driver shortages." Without: it clears as "the market." With: decomposed — market +5% vs markup +9% — checked against the contract ceiling; 9 points over, credit claimed.
Recipe / yield coupling → Theoretical-vs-Actual Variance · near
Food cost is dead-on 30% — the burger is bleeding. Without: the aggregate hides it. With: per-item variance finds the burger at 38% (7.1 oz vs a 6.0 oz spec) — re-spec or re-price.
① context → ⑧ gate → ② act → ④ memoryburger 38% → target~$7K / yr
When Eggs Went to $8
Commodity price volatility → Earned Autonomy & self-governance · near
Bird flu takes eggs from $1.50 to $8. Without: buy-to-par rides it up — every egg dish's food cost blows up, repriced a month late. With: the price break is flagged, it switches to your backup spec, flags the egg dishes to reprice, and stops buying forward — holding margin until prices settle.
Geometry, not rules — and it earns the right to act
Each order cycle, every category becomes a seven-factor vector. The scorer picks the action from learned decision-prototypes — then the gate decides whether it's earned the right to act, or routes to a human with the reasoning attached.
7 factors · read from your systems
Expected demand
Day of week
Weather
Local events
Waste history
Supplier OTIF
Price memory
→
scorer
Nearest prototype5 categories × 4 actions, learned from verified outcomes
→
action + gate
order as planned
order more ← event
order less
skip
auto-approve · earned (≤ floor) — else route to human
5 categories × 4 actions × 7 factors = 140 learned values — the whole judgment, auditable by hand.
Quick wins
What the first 90 days feel like
The pilot proves the cost side before it writes to anything — each win is a fault-line, caught in the order it pays off. Menu pricing and labor follow on the roadmap.
The economics
Priced like a premium inventory tool — self-funding from the first avoided over-order
Per-location monthly subscription, banded to the three buyers in the GTM. Competitor price landscape below.
Independent · single-unit
Operator
$499per location / month
Multi-unit · hotel F&B — beachhead
Group
$349per location / month · 5+ units
Chain · franchise
Enterprise
Customplatform + rollup
At a $1.5M restaurant the subscription is ≈ 0.3–0.4% of monthly revenue (Group to Operator tier) — a rounding error against the food-cost points it moves. Priced on the location, not per-decision, not gain-share.
The beachhead is multi-unit hotel F&B: a $28M operation — restaurant, bar, banquet, room service — where 1–3 points off food cost is ~$84–280K/yr (modeled). One proof there sets the baseline for the whole group.
Where we sit — competitor price landscape ($ / location / month)
market toolsRestaurant-RGI (proposed)
Crunchtime is enterprise chain back-office at $5,000+/mo — off this scale. We land mid-to-upper in the inventory-suite band — a premium tool, not a new tier. Competitor prices are published or estimated list ($/location/mo, Oct 2026).
Buyer ROI is the value trajectory in The outcomes above — 18–22× at steady state (both sides, modeled).
The platform advantage
It doesn't start from zero — the platform teaches it
Restaurant-RGI runs on the same Graph Attention Engine as the Source-to-Pay copilot. The shared tensor warm-starts a new domain from the structure the platform has already learned — so a restaurant copilot inherits structural priors from day one, not a cold start.
One engine, five copilots — cross-domain discovery warm-starts each new copilot (~2–7× faster, modeled).
Restaurant-RGI is the margin copilot on the Compounding Intelligence platform — one engine that learns your kitchen from its own outcomes and earns governed autonomy. The same engine powers copilots across the enterprise; the kitchen is the wedge.
Built by Arindam Banerji, PhD — 20+ years in enterprise AI and data, deep in retail and supply chain. Creator of SOA (Service-Oriented Architecture — the backbone of modern e-commerce) and co-creator of e-Speak at HP Labs (now in the Smithsonian); built and commercialized the Accenture Insights Platform, one of the earliest full-stack AI/ML platforms. 9 US patents.
