Online shop process audit: 3 automation points, ROI 4.2× over 12 mo
Online home-goods shop: 12 SKU categories, 280 orders/day, team of 14. Revenue lagged behind headcount growth. A 2-week audit of six roles found 3 automation points — freed up 1.1 FTE and cut inventory drift 25×.
The team grew faster than revenue — everyone was firefighting
Shop in home goods, two years of steady growth — 280 orders/day on average, 12 SKU categories, sales on own site, Wildberries, and Ozon simultaneously. The team grew to 14 people in 8 months, but revenue lagged the hiring pace. The owner hired on the "a few more people and we'll catch up" principle.
During the auditor's shadow sessions it turned out: all six key roles lived in "firefighting" mode. Stock on the site, in WB, Ozon, and 1C diverged constantly. The manager manually reconciled stocks three times a day, and still oversells and cash gaps happened.
Support answered the same questions 80-120 times per week: "where's my order", "bed dimensions", "is the color in stock". The accountant spent 9 working days on month-end close — assembling closing documents from three sources: 1C, marketplace dashboards, and email.
Audit → 3 automation points with highest ROI
Two-week audit: shadow sessions of six key roles, log exports for 90 days, time-tracking of top 20 operations. Output — prioritized list of 11 potential automation points and an ROI estimate for each. Only the three with the shortest payback went to implementation.
Site ↔ WB ↔ Ozon ↔ 1C, single source of truth. n8n polls every 15 minutes, merges by SKU, source of truth is 1C.
GigaChat answers the top 40 FAQs + RAG over product cards. Escalation to manager on triggers — tone, keywords, retries.
Generation of UPD / invoice / act directly in 1C + email delivery with tracking and confirmation.
Audit as a standalone product
Two weeks is two full phases. First: shadow sessions of six roles at 90 minutes each, log exports from 1C, RetailCRM, and WB/Ozon dashboards for 90 days, time-tracking of the top 20 operations with micro-measurements. Second: formulating 11 automation points, ROI estimate for each with assumptions, presenting the solution to the owner with prioritization.
Stock sync — the most painful spot
On different platforms the same SKUs had different IDs, different units, different stock registers. Solution via n8n: unified SKU mapping table and polling all sources every 15 minutes. Source of truth — 1C. After each sync n8n recomputes site stock and updates cards in WB/Ozon via Seller API.
AI support with escalation
GigaChat 2 Max, answering strictly from the product cards and the FAQ table (retrieval-augmented generation, RAG). The bot handles 78% of incoming messages automatically — "where's the order", "size", "stock", "delivery time". Escalation to manager triggers on: negative tone, keywords ("return", "defect", "lawyer"), confidence < 0.7, or second retry on the same question.
Closing documents as a pipeline
1C (REST + OData) generates UPD, invoice, and act on order shipment, n8n bundles them into a package, attaches a PDF instruction, and sends to the customer with tracking. The accountant moves from "document collector" to "reconciliation controller".
Stack tailored for the Russian perimeter
Orchestrator of all syncs and pipelines on the client's own VPS
Stock source of truth and closing-document generator
Unified order window from site and marketplaces, source for AI support
Stock read/write, card updates, order statuses
Same as WB — two-way stock and order integration
AI first-line support + RAG over product cards and FAQ
FAQ base + process and SOP documentation for the team
n8n and middleware hosting — entirely in RF perimeter, no foreign vendor
What changed in numbers
~1.1 FTE = ₽89k/mo of estimated payroll savings
converged 25× — oversells practically eliminated
audit ₽380k + implementation ₽220k = ₽600k
including escalations, 78% fully automated
accountant controls reconciliation, doesn't assemble
The main beyond-numbers result — the owner stopped non-stop hiring. 1.1 freed FTE means redistribution of people from operations to growth projects: turnover, procurement optimization, new categories.
ROI 4.2× by month 6 after implementation is conservative: calculated by payroll savings and reduced oversell penalties. Not counted: marketing wins from faster support and improved turnover (procurement based on accurate demand).
When an audit beats a "big project"
Universal pattern — "operational debt accumulates faster than hiring". Audit as a product works where the owner already knows "something needs automation", but doesn't know where the highest ROI is:
- → Multi-channel retail and e-commerce selling on their own site plus 2-3 marketplaces
- → Service companies 10-50 people on manual operations: wholesale, B2B services, agencies
- → Manufacturing-trading with a warehouse, suppliers, and own site in parallel
- → Service networks (salons, clinics, auto repair) with multiple locations and shared accounting
- → Any team where the owner feels "we're firefighting" and hires faster than revenue grows
- Shadow sessions of six key roles at 90 minutes each — observing real work as it happens
- Log exports of all systems for 90 days — operations statistics, not "we think"
- Time-tracking of the top 20 operations — real numbers on hours and repeat counts
- Prioritized list of automation points with ROI estimates and assumptions
If the team grows faster than revenue — time to look at processes
Audit as a standalone product: 2 weeks, fixed cost, output is a prioritized list with ROI estimates per point. Implementation — separately, by the owner's decision after the audit.
Deep dive:Inventory sync between 1C, Wildberries and Ozon: why the numbers never match
Related cases
AI chatbot security audit: prompt injection + PII leak + FZ-152
2.5-week red-team per OWASP LLM Top 10. Findings: 8 critical + 11 high + 14 medium. Grounds for an FZ-152…
Event forecasting with an LLM ensemble: calibration and Brier scoring
Five LLM agents with web search over open sources, a supervisor, calibration, every forecast written to a…
Local LLM Trading Engine: Ollama + RAG on M3 Mac
Multi-model decision engine 100% on-device: qwen2.5:14b primary + llama3.1:8b fallback + LLaVA:7b vision +…
The 9,900 ₽ audit — with a concrete report and quote
I'll tell you what to deploy in your business first, what the payback looks like, and whether you need AI for the task at all (sometimes you don't).
Or just send your question — I reply within 2 hours