Every month on payday, any active seller on Wildberries or Ozon looks at their bank statement with a slight chill. Your operational team's payroll is a massive fixed vacuum cleaner that sucks out net profit regardless of how the trading week went, whether a truck got stuck at the Koledino warehouse gate, or whether the platform's algorithm tanked your listing's search ranking.
If you add up the minimum roster of specialists without whom a store doing 5 to 10 million RUB in monthly turnover turns into chaos, the numbers are steep:
- A full-time merchandise analyst costs the market around 60,000 RUB.
- An outsourced CFO or management-accounting specialist takes 80,000 to 120,000 RUB.
- An assistant handling routine review moderation and customer Q&A asks for another 15,000 RUB.
- An internal-ads traffic manager holds steady at around 65,000 RUB.
- Add subscriptions to external analytics platforms and spreadsheet trackers, and you are looking at roughly 270,000 RUB in monthly overhead.
Entrepreneur Leo Shevchenko tested a bold hypothesis on real store data: can a seller take the core routine functions of all four specialists and hand them over to Anthropic's Claude large language model, whose subscription—even factoring in cross-border payment fees—runs just 3,000–8,000 RUB a month?
I dug into the mechanics, ran real seller-dashboard exports through the model, and I am ready to break this experiment down piece by piece: where AI pays for itself ten times over in a single evening, and where taking humans completely out of the loop will almost certainly land you in a cash-flow gap.
Payroll Economics: What Sellers Actually Pay 270,000 RUB For
Before you rush to fire your team, let's be honest about what an entrepreneur is actually buying for a quarter of a million rubles a month. On marketplaces, you aren't paying for hours spent staring at a monitor—you're paying for reaction speed when your sales funnel slips and for precision in financial decisions.
When a top-selling SKU unexpectedly goes out of stock at the warehouse, the listing drops out of search results, and clawing back that organic ranking burns hundreds of thousands of rubles in ad spend. When an accountant puts together a settlement report by guesswork, the seller spends months moving items with negative unit economics, effectively subsidizing the marketplace out of their own pocket.
The comparative math is stark:
- Traditional operations team: 270,000 RUB per month, human error, sick leave, turnover with lost context, and reporting delays of 3–5 business days.
- AI-powered infrastructure: a Claude Pro subscription ($20) or a Claude Team plan ($25–$30 per seat) with shared Projects knowledge bases, which comes to around 8,000 RUB per month including payment gateway fees.
That is more than a 30x cost difference. We covered a similar budget shift when looking at AI optimization on marketplaces: where neural networks actually cut costs, where automating analytics and ads boosts net store profit without endlessly inflating top-line turnover. But to turn that savings into reality, a seller has to learn how to task the model using structured system role prompts.
The Merchandise Analyst Layer: Funnel, Stock, and Stockout Audits
The main flaw of manual inventory analysis is fragmented data. A seller downloads a stock report from the WB dashboard, sees a couple hundred units sitting in warehouses, and relaxes—completely missing an approaching out-of-stock event.
To hand this job to an AI, you feed it two standard exports: the "Sales Funnel by SKU" and the "Warehouse Stock Report." Respecting the planning horizon is critical: the export must cover three control windows—30, 60, and 90 days—not a single week. If you only feed the model a two-week slice, it will mistake a one-off spike from an external promo for a steady long-term trend and tell you to overorder.
In a practical test for an apparel store doing 10,000,000 RUB in monthly revenue with a matrix of 20 active SKUs, the system prompt for Claude looks like this:
You are a lead merchandise analyst on Wildberries. Audit the product matrix based on two files:
1. "Sales Funnel" for 30, 60, and 90 days.
2. "Warehouse Stock" as of today.
Analysis rules:
- Work strictly from the facts in the tables without guessing.
- Identify Group A items (80% of revenue) and Group C laggards.
- Calculate 30-day Sales Velocity and compare it against current stock.
- Forecast the stockout date assuming a 30-day replenishment cycle.
- Spot bottlenecks: high CTR with low add-to-cart, or high add-to-cart with low buyout rate.
In a minute and a half, the AI finishes a workload that takes an in-house analyst a full day. The model immediately surfaces anomalies: for example, a women's cardigan SKU gets strong clicks from search, yet its add-to-cart rate is three times lower than the store average. Cross-referencing reviews reveals why: buyers are unhappy that the color in the photos doesn't match the actual garment.
The CFO Layer: Settlement Reports and Unit Economics
Wildberries' weekly financial settlement report is a maze of dozens of columns: agreed discount, sale price, marketplace commission rate (KVV), acquiring fees, outbound and return logistics, storage, paid receiving, and penalties.
Infographic blueprint of the Wildberries settlement report control balance formula on warm white paper with dimension arrows and cards for commissions, logistics, and net profitability.
Most entrepreneurs only glance at the final payout hitting their bank account. Meanwhile, "parasite SKUs" are hiding inside that report—items that technically generate millions in revenue, yet because of a low buyout rate (say, 25–30% in tricky footwear categories) and constant warehouse-to-warehouse joyrides, they eat up the entire margin of neighboring SKUs.
For a financial audit, the system prompt is anchored around the control-balance rule:
You are an e-commerce CFO. Audit the Wildberries settlement report and the cost-of-goods table:
1. Reconcile the control balance: Payout Amount = Revenue - KVV Commission - Agent Services - Logistics - Storage - Penalties. State the exact discrepancy in rubles if it does not balance.
2. Calculate unit economics per SKU: commission share, logistics cost per bought-out unit, storage per unit in stock, COGS, and Net Margin %.
3. Flag loss-making "parasite SKUs" that show positive top-line revenue.
4. Propose 3 management actions to cut costs.
Here lies a crucial technical detail: language models are text generators by nature. If you force Claude to add up long columns of numbers in a plain chat window, it will inevitably round decimals and botch the totals. To prevent that, calculations must run through code execution in the Analysis Tool: the AI writes a Python script, executes it, and returns a verified, exact result.
We saw a similar engineering approach to cost control in classifieds when breaking down promotion economics on classifieds: hidden business costs and fair specialist rates, where exact fee math protects an entrepreneur from the illusion of easy money.
The Customer Support Layer: Turning Complaints into Technical Specs
Generic auto-replies like "Thank you for your purchase, we value your choice" only irritate buyers—especially when someone receives torn packaging or a defective item.
In Leo Shevchenko's experiment, Claude acts as an analytical complaint filter, sorting incoming reviews and turning them into concrete action items for different departments:
- SKU 462 ("Hair Bows"): negative-review analysis flagged fraying ribbon edges. The AI drafted a strict spec for production: seal cut edges with a hot knife and ban assembling bows from small fabric scraps.
- SKU 1006 ("Women's Bodysuit"): buyers left 2–3 stars because there was no shaping effect. Action item for content: immediately strip "heavy-control shapewear" from the title and infographics, since the knit is comfortable but doesn't sculpt the figure.
- SKU "Athletic Biker Shorts": complaints about fabric going sheer when stretched. Action item for procurement: raise fabric density from 180 to 240 gsm.
- SKU 123 ("Summer Bucket Hat"): the system spotted an expiring EAEU declaration of conformity and warned the seller about listing-block risk 20 days before the deadline.
Instead of tedious copy-paste replies, customer support becomes a continuous quality-control audit that stops waves of returns before they happen.
The Traffic Manager Layer: Cleaning Up Clusters and Auto-Campaigns
Automatic ad campaigns (ARC) bundle placements across marketplace search and catalog under a single bid. The catch is that the platform's algorithm glues search queries into clusters. As a result, a women's winter puffer jacket priced at 6,500 RUB starts burning expensive clicks on queries like "men's mid-season jacket," "kids' puffer," or competitor brand names.
An in-house traffic manager spends hours manually weeding through search phrases. Claude handles this in batches: the seller exports ad-cabinet stats, and the AI sorts queries into four buckets—high-converting target terms, broad adjacent terms, irrelevant junk, and zero-result phrases that burned over 300 RUB with zero clicks.
Using a target ad cost share (ACS capped at 10%) and a listing's 4.5% click-to-order conversion rate, the model instantly calculates the maximum allowable CPM bid above which ads turn unprofitable. Programmatic workflows across hundreds of listings have already proven their worth: look at how an AI Avito specialist on Claude API automates managing 600 listings without manual edits, freeing the owner from tweaking prices and copy by hand.
Limits and Safety Rules
On the surface, it looks ideal: swap out a team for a script and pocket super-profits. As a practicing seller, though, I have to warn you about the traps. AI models come with fundamental limits.
The first limit is context-window size. If you try to dump a 50,000-row spreadsheet tracking thousands of SKUs across warehouses, Claude physically cannot hold every number in working memory and will start aggregating data with errors. The rule is simple: segment your assortment by category and feed the prompt pre-grouped summary tables.
The second limit is trade secrets. On standard consumer plans, conversation data may be used by providers for model training. If you are working with real factory purchase prices, Chinese supplier contacts, and unit-economics formulas, use Claude Team or Enterprise plans with Zero Data Retention enabled, or run your pipeline directly via API.
An AI model is a brilliant analyst and mathematician, but it has no bank account and no legal liability. It prepares decision options, cleans funnels, and spots budget leaks in minutes. Yet hitting the final button—changing a retail price, signing a factory supply contract, or writing off inventory—always remains the entrepreneur's sole responsibility.
The bottom line is pragmatic: AI doesn't replace people entirely, but it gives one competent operator the output of an entire office. Once you digitize your store's daily routine with sharp prompts, your unit economics gain rock-solid resilience.

