QA automation: routine that does itself

Not a fourth service but a way to do the first three faster: scheduled data checks, requests that land in the tracker by themselves, reports assembled without a human. Built on Python and n8n, connected to Telegram, CRMs, spreadsheets and language models.

What exactly we automate

We take processes that repeat every day or every week and consist of clear steps: collect, reconcile, move, notify. Where a judgement call is needed, we add a language model, always with a guard rail — complex cases go to a human.

• Data reconciliation: stock, prices, order statuses, campaign metrics — on a schedule, with alerts on discrepancies.
• Collecting requests from forms, email and messengers into one tracker tagged with source, topic and owner.
• Reports assembled from several sources and sent to the lead at a set time.
• AI assistants on the product's knowledge base: answer routine questions, hand the rest to a manager.
• Integrations between services with no ready connector: CRM, Google Sheets, ad accounts, marketplaces, trackers.
• Pre-release checks: field completeness, link validity, version currency.

Stack and principles

Python for logic and integrations, n8n for orchestration and visual control of workflows, aiogram for Telegram bots, Claude API for tasks that need text understanding, Whisper for voice messages and calls, a RAG approach for assistants over the client's documents. Hosting on your side or on our server with access handed over.

• Every workflow logs what it did and posts to a chat if something goes wrong.
• No black boxes: you get the process diagram and can change parameters without us.
• Access to your systems through separate tokens with minimal rights, stored in your vault.
• The language model never makes financial decisions and never writes to customers without a template.

How a project runs

1. Process walkthrough on a call: who does what by hand and when, where errors occur.
2. Workflow diagram and estimate: steps, data sources, points where a human is needed.
3. Build and a trial run in parallel with the manual process — one or two weeks of comparing results.
4. Switch-over, documentation, team training. Workflow support by agreement.

Small workflows — lead intake or a daily reconciliation — are ready in a few days. Knowledge-base assistants and multi-step integrations take two to four weeks including the trial period.

Examples from practice

• An auto parts seller on a marketplace: marketplace-to-ERP sync by warehouse, XML feeds, an LLM assistant for client chats, bank reconciliation.
• A performance marketing agency: spend-to-top-up reconciliation to the cent, around 200 automated checks of internal tools, scheduled jobs with chat alerts.
• A QA process: support requests become tickets with reproduction steps, regression runs start on a schedule with a report to the chat.