Claude Code AI agents in procurement and cash flow: what changed in 8 weeks (marketplaces and Amazon) (ИИ-агенты Claude Code в закупках и cash flow: что изменилось за 8 недель | #маркетплейсы и #amazon)
B · Useful, limited Agency founder Vladislav Mikhailenko (SellerUp) walks through a client build of five Claude Code agents for replenishment and cash planning with specific before-and-after numbers, but it is a self-reported sales case with NDA-altered amounts and confounders he names himself.
Summary based on the full transcript·Original: about 15 minutes to watch
In short
A 15-minute Russian YouTube case talk (2 November 2026) by Vladislav Mikhailenko, founder of the SellerUp agency. A client importing home goods from China (about 260 SKUs, about 40 shipments in transit, 17 warehouses, own site plus 10+ European marketplaces, not Amazon despite the title) replaced an Excel purchase plan run by one buyer with five Claude Code agents. The agents forecast, size orders by capital payback, track shipments and lay out a 90-day payment calendar. Humans still place every order and payment. Payback fell from 105 to 89 days over eight weeks.
Key numbers
- 105 to 89 dayscapital payback (stock value divided by daily profit) before vs after the agents, over an 8-week window
- $184,000reduction in stock over 8 weeks, against about $2.1 million held in inventory
- 242 signals, 159 reorder prompts, 81% closedalerts sent by the agents in 12 weeks and the share the team acted on
- 3.5x to 4.1xcapital turns per year before vs after
As reported by the source.
Russian·Video·opens youtube.com
What it reports
- Before: one buyer ran purchasing in Excel; when that person was on holiday, purchasing stopped. Factory and carrier payments landed on the same days.
- Why not an off-the-shelf forecasting tool: sea lead times of 28 to 113 days (sometimes 8 months), minimum order quantities, kits stored as several SKUs, one product under three article numbers, seasonality, two-part payments.
- Five agents: data (orders, stock across 17 warehouses, in-transit batches, payments and FX every 4 hours), forecast, purchasing (payback, batch size, supplier comparison), shipments (90-day payment calendar) and a free-text assistant. Alerts go to Telegram.
- Reorder rule: sales velocity times lead time plus a 15% buffer, using last year's same period for off-season items and re-forecasting within 4 hours after a promotion spike.
- One real case: without the agent an order went in with 20 days of stock and a 60-day lead time, leaving 40 days out of stock; with the agent the signal came at 69 days of cover.
- Over 12 weeks the agents sent 242 signals, 159 of them reorder prompts; the team closed 81%.
- Results, late July to late September: payback 105 to 89 days, capital turns 3.5x to 4.1x a year, each $100 in stock earning $34 a month instead of $29, the same capital earning 18% more.
- Build: first agents live in 4 weeks, the full set in 10, then about a month of testing and 240+ fixes. The first week ran in parallel with the manual plan, reconciled daily.
How they got it
A before-and-after comparison over one 8-week period chosen because demand was flat before the season, presented by the agency that built and sells the system. On-screen demo uses substituted data. Amounts are altered under NDA; the speaker says percentages and timelines are real.
Limits
- Self-reported by a vendor; ends with a pitch for consulting and turnkey builds.
- The speaker says price and assortment changes also drove profit in the same period, so the payback gain is not isolated.
- No build cost or Claude usage cost given; the timeline does not reconcile cleanly (8-week result window, 12 weeks of alerts, 10 weeks to build plus a month of testing).
- Source is auto-captions; some figures may be garbled.
Why it matters for companies
The case is a template for inventory-heavy businesses where one planner and a spreadsheet are the single point of failure. The design choices are the useful part: a business metric (payback days) as the ranking key, a parallel run against the manual plan, and humans keeping the order and payment decisions. The payment calendar that shows the heaviest cash day weeks ahead is the kind of output finance teams can use without trusting the forecast blindly. Discount the results until someone measures them with the confounders removed.
Key passages in translation
Russian original
Важный момент, агенты сами ничего не заказывают, не платят сами, решение всегда принимает человек.
An important point: the agents order nothing and pay nothing themselves; a person always makes the decision.
Russian original
Без агента заказ сделали, когда товара оставалось на 20 дней. Это реальный случай. А партия ехала 60.
Without the agent, the order was placed when 20 days of stock remained. This is a real case. And the batch took 60 days to arrive.
Russian original
У нас менялись цены и ассортимент, поэтому это не единственная причина роста прибыли.
Our prices and assortment changed, so this is not the only reason profit grew.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com