Alexandr Chibilyaev on AI agents that manage inventory, optimize menus, schedule staff, and analyze customer feedback — powered by POS integrations with iiko, Poster, R-Keeper, Evotor, and the full restaurant tech stack.
A restaurant isn't one business. It's five businesses operating in the same physical space: a kitchen that produces food, a warehouse that manages ingredients, a retail floor that serves customers, a logistics operation that handles suppliers, and a marketing team that drives demand. The general manager is supposed to keep all five running — usually with a clipboard, a spreadsheet, and a prayer.
Most restaurant technology addresses one slice of the problem. iiko manages the POS. Poster handles inventory. R-Keeper runs the kitchen display. Evotor processes payments. But none of them talk to each other in a way that produces autonomous action. The manager is the integration layer, manually pulling reports from each system and making decisions from fragmented data.
AI agents change this. They connect iiko to Poster to R-Keeper to Evotor. They read sales data, monitor inventory levels, analyze menu performance, schedule staff, and surface customer feedback — not as separate dashboards, but as a single intelligent system that acts on insights without waiting for human intervention.
AACFlow connects the entire restaurant and retail ecosystem through our knowledge base sync architecture. The agents don't "check the iiko API." They understand the business — every sale, every ingredient, every review — because the entire operation lives in their working memory.
Before an agent can manage a restaurant, it needs to see the restaurant. The sync engine connects to every system:
iiko — full restaurant management: POS transactions, menu engineering, inventory by ingredient, supplier orders, recipe costing, kitchen production tracking
Each connector normalizes restaurant data into a unified model. A sale in iiko, a stock adjustment in Poster, a shift schedule in Saby — all mapped to the same entities: products, ingredients, employees, transactions. The agent searches the knowledge base and sees the whole restaurant — not four separate systems.
Restaurant inventory is chaos by design. Ingredients arrive from suppliers. Some go straight to the kitchen. Some sit in storage. Some spoil before they're used. Some get wasted during prep. Some are consumed during service. Every day, the gap between "what the system says we have" and "what we actually have" widens. By the end of the week, nobody knows the real inventory count — and the Sunday rush hits with a mystery shortage of chicken breasts.
The Inventory Management Agent maintains a real-time, accurate inventory picture:
Ingest — the sync engine pulls data from:
iiko/Poster: current stock levels, recent write-offs, production consumption
Actual inventory (latest stock count or system-reported levels)
Discrepancies >5% are flagged for physical inventory check
Predict — based on historical sales patterns, upcoming reservations, day of week, weather, and local events (pulled from knowledge base), the agent forecasts demand for each ingredient for the next 7 days
Auto-order — when an ingredient's projected stock drops below its reorder threshold (accounting for supplier lead time), the agent generates a purchase order:
Selects the preferred supplier from the supplier catalog
Calculates optimal order quantity (economic order quantity with demand buffer)
Drafts the order in iiko/Poster for manager approval
Monitor spoilage — ingredients approaching their shelf life are flagged. The agent suggests: increase usage in today's specials, offer a discount to move quickly, or mark for write-off before it spoils in storage.
Alert — critical stockouts (ingredient at zero, no delivery expected within 48 hours) trigger immediate Telegram alert to the kitchen manager and head chef
The result: the restaurant stops running out of ingredients mid-service. The Sunday rush doesn't expose a week of neglected inventory management. And the kitchen team knows what they have, what they need, and what's arriving — without anyone playing detective with spreadsheets.
Every restaurant menu has dead items. Dishes that nobody orders, that cost more to make than they earn, that the kitchen hates preparing, that pull down the average check. Identifying these items is easy in theory. In practice, the data lives in iiko's sales reports (which the manager checks monthly, if at all), Poster's revenue dashboards (which show revenue but not margin), and the kitchen's informal feedback ("nobody orders the octopus, chef").
The Menu Optimization Agent analyzes menu performance continuously:
Analyze sales — the agent pulls transaction data from iiko/Poster/R-Keeper and calculates per-dish metrics:
Revenue contribution: what percentage of total revenue does each dish generate?
Margin contribution: accounting for recipe costs (from iiko's recipe management), what's the actual profit per dish?
Trend: is this dish growing, stable, or declining over the last 4 weeks?
Day/time patterns: does this dish sell only on weekends? Only at lunch?
Identify problems:
Profit killers: high-revenue, low-margin items. The dish sells but loses money.
Zombie dishes: ordered less than X times in the last 30 days. Taking menu space, generating waste.
Underpriced items: dishes with high demand but below-market pricing compared to similar items at competitor restaurants.
Prep bottlenecks: dishes that consume disproportionate kitchen time during peak hours.
Recommend — the agent produces a weekly menu optimization report with specific, actionable suggestions:
"The Grilled Salmon costs 480₽ to produce at current supplier prices and sells for 890₽. Margin is 46%. Competitor pricing for similar dishes is 1,200-1,400₽. Consider raising to 1,250₽. Estimated impact: +112,000₽ monthly revenue."
"The Octopus Carpaccio has been ordered 3 times in the last 30 days. Monthly ingredient waste: 18,000₽. Recommend removing from menu and replacing with seasonal ceviche, which tested well in focus group and has 62% projected margin."
"Lunch special 'Business Lunch Premium' sells 80% of its volume between 12:00-13:30. Kitchen prep time is 18 minutes per dish, causing bottleneck. Recommend prepping components by 11:00 to reduce ticket time."
Seasonal analysis — the agent tracks ingredient price fluctuations from supplier data and adjusts menu recommendations seasonally. When tomato prices spike in winter, the agent flags dishes heavy on tomatoes and suggests seasonal alternatives.
Competitor intelligence — if competitor menu data is available (via web scraping connectors), the agent compares your menu against local competition: price positioning, category gaps, trending items you don't offer.
The menu becomes a living document, continuously optimized by data — not a static PDF that gets updated once a year when the chef feels inspired.
Restaurant staffing is a three-dimensional puzzle: labor laws, employee availability, demand forecasting. Understaff on Friday night, and service collapses. Overstaff on Tuesday afternoon, and labor costs eat your margin. Get the schedule out late, and employees miss shifts because they made other plans. Get it wrong on hours, and you're violating labor regulations.
The Staff Scheduling Agent solves the scheduling puzzle:
Forecast demand — the agent predicts hourly customer volume based on:
Historical transaction data from iiko/Poster (previous 12 weeks, same day of week)
Upcoming reservations (from reservation system connector)
Local events (concerts, sports, conferences — from web data connectors)
Calculate staffing needs — demand forecast is converted to role-specific staffing requirements per hour:
Servers: 1 per X covers
Kitchen: 1 cook per Y orders/hour
Bar: 1 bartender per Z seats
Support: bussers, hosts, dishwashers based on volume
Match employees to shifts — the agent reads employee data from Saby/Poster:
Availability preferences and constraints
Legal limits (maximum hours per week, minimum rest between shifts)
Skill requirements (some shifts need a senior cook, some need a bartender with cocktail experience)
Labor cost rates (senior staff cost more — optimize the mix)
Generate schedule — the agent produces a weekly schedule that:
Meets forecast demand with appropriate staffing levels
Respects all employee constraints and legal requirements
Minimizes labor cost while maintaining service quality
Is ready by Thursday for the following week
Publish — the schedule is posted to the staff Telegram channel. Employees confirm their shifts. Conflicts are flagged for manager resolution.
Adjust in real-time — if actual demand deviates from forecast (an unexpected rush, or a dead shift), the agent suggests adjustments: call in on-call staff, release volunteers early, rebalance sections.
One restaurant chain we work with runs 15 locations on AACFlow. Before, their operations manager spent 12 hours every week building schedules across all locations. Now the agent generates the first draft automatically. The manager spends 2 hours reviewing, adjusting for special cases, and approving. The schedule is more accurate because the agent processes more data than any human could.
Every restaurant receives feedback from multiple channels: Yandex.Maps reviews, 2GIS ratings, Google Maps reviews, social media comments, direct feedback forms, waiter-reported complaints. The volume is manageable when you have 5 reviews per week. At 50 reviews per week across platforms, the signal gets lost. A pattern of complaints about slow service on Friday nights goes unnoticed for months — until the rating drops and management finally investigates.
The Customer Feedback Agent turns scattered feedback into actionable intelligence:
Aggregate — the agent pulls reviews from all connected sources: Yandex.Maps, 2GIS, Google Maps, Telegram feedback channels, in-app surveys
Classify — each review is categorized by:
Sentiment: positive, neutral, negative
Topic: food quality, service speed, staff attitude, cleanliness, atmosphere, price, portion size, menu variety
Respond — for positive and neutral reviews, the agent drafts a personalized response in the restaurant's voice, referencing specific dish mentions and staff names if provided. Response is posted to the review platform after manager approval.
Food safety complaints → alert general manager and head chef immediately
Aggressive or inappropriate staff behavior reported → alert general manager
Review mentions a competitor favorably → log for competitive analysis
Detect patterns — the agent continuously analyzes review topics and sentiment over time:
"Service speed complaints increased 40% in the last 2 weeks, concentrated on Friday/Saturday dinner shifts at Location #3. Cross-reference with staffing levels: those shifts are at 85% of recommended staffing."
"Positive mentions of the new seasonal cocktail menu are 3x higher than the previous menu. Consider extending the seasonal menu by 2 weeks."
"Delivery orders have 23% more negative reviews than dine-in. Common complaint: food temperature on arrival. Recommend reviewing packaging and delivery radius."
Weekly report — the agent generates a digest: top 3 things going well, top 3 issues to address, sentiment trend, platform comparison (are Yandex.Maps ratings different from Google ratings? Why?)
The agent doesn't just collect reviews. It reads them, understands them, and connects the patterns to operational data — so the restaurant knows not just that there's a problem, but why and what to do about it.
One of our customers — a restaurant chain with 15 locations across three cities — runs AACFlow to manage operations. Their tech stack was classic: iiko for POS, Poster for some locations, 1C for accounting, Saby for HR, spreadsheets for everything else.
After deploying AACFlow agents:
Inventory: stockouts during peak service dropped 78%. The agent catches low-stock situations 2-3 days before they become problems, and auto-generates supplier orders. Physical inventory now reconciles within 3% of system counts instead of the previous 12-15% variance.
Menu: the Menu Optimization agent identified 11 underperforming dishes across the chain. Replacing them with higher-margin alternatives increased average check by 8.4% and improved kitchen efficiency during peak hours.
Staffing: the Scheduling agent reduced labor costs by 12% while maintaining service levels. The secret was better demand forecasting — the agent allocates staff precisely when they're needed, not flat coverage across the day.
Reviews: average response time to negative reviews went from 3 days to under 4 hours. The chain's Yandex.Maps rating improved from 4.1 to 4.4 across locations in 3 months — driven primarily by faster issue resolution and consistent engagement with reviewers.
Manager time: general managers reported saving 15-20 hours per week on operational tasks. That time is now spent on the floor — training staff, engaging with guests, improving service.
The operations director told us: "We didn't realize how much of our managers' time was consumed by looking at data and making routine decisions. The agents handle the routine. The managers handle the restaurant."
The same agents work for retail. Evotor is the POS. 1C is the ERP. Inventory management, pricing optimization, staff scheduling, and customer feedback analysis — the agent patterns are identical. The products are different (no food spoilage, but seasonal trends and shelf-life for perishable goods), but the architecture is the same.
Retail-specific additions:
Price tag management: the agent monitors competitor prices and suggests adjustments to stay competitive without racing to the bottom
Promotion analysis: the agent measures promotion effectiveness — did the "buy one get one 50% off" actually increase margin, or just cannibalize full-price sales?
Supplier performance: the agent tracks supplier reliability (on-time delivery rate, quality issues, price competitiveness) and recommends supplier mix adjustments
Connect your POS — iiko, Poster, R-Keeper, or Evotor from the connector catalog
Connect your backend — 1C for accounting and inventory, Saby for HR and scheduling
Connect feedback channels — Yandex.Maps, 2GIS, Google Maps for review aggregation
Deploy agents — start with inventory management (biggest operational pain point), then add menu optimization and customer feedback
Expand across locations — the same agents scale to multiple locations without additional configuration
Restaurant and retail are high-touch, high-variance businesses. But the operational layer — the data, the patterns, the routine decisions — is remarkably consistent. AI agents handle the consistency. Humans handle the art.
