Four tiers. Yours is the simplest one that solves your problem.
DjbrAutomation does one thing: AI automation systems — workflows, agents, knowledge systems, and web or mobile applications built around automation. Every project starts with a success metric and a written definition of done, agreed before the first line of code.
Audit & essential automations (n8n, API integrations)
Your days go to manual copy-paste: an order re-typed into another tool, a follow-up that slips through, a weekly report assembled by hand. None of these tasks require judgment — they require being executed on time, without errors.
What you receive
A trigger audit: a structured interview of your operations — every predictable event (incoming order, email, deadline, stock threshold) is reviewed as a candidate for automation.
1 to 3 n8n workflows in production, connected to the tools you already use (email, WhatsApp, Telegram, spreadsheets, CRM, Shopify).
Handover documentation: every workflow explained, maintainable by any developer — not just by us.
Simple monitoring: you know it is running, and you are alerted when it stops.
Proven example
n8n workflows in production — RAG support agent, AI newsletter, WhatsApp/Telegram assistant, LinkedIn scraper, Telegram invoicing.
An AI content pipeline demonstrated end to end for a French-speaking advisory firm specialized in company setup in the Middle East: case studies turned into LinkedIn posts, Instagram carousels and avatar videos — Telegram human-approval module delivered; first automated publication on 29/06/2026 (demonstration on test accounts).
DONE = the agreed workflows have been running in production for 2 weeks without intervention, and the metric defined at the audit is being measured.
If your pain is copying and forgetting — not judgment — this tier is enough. An AI agent would only add cost and variance.
Some tasks don't fit a workflow: reading a message, understanding a request, sorting, drafting a reply, deciding a simple case. That is where — and only where — an AI agent is justified.
What you receive
An agent in production on your real channels: email, WhatsApp, Telegram, back office.
Explicit guardrails: a written scope of action, checks before every sensitive action, a log of every decision.
"What stays human", documented: the exact list of cases where the agent stops and hands over.
An interchangeable LLM provider: switching models does not mean rebuilding the system.
Proven example
An autonomous agent operates an e-commerce merchant's Shopify store 24/7: 21 skills in production at the client — 4 of them created by the agent itself, then approved through human review — and 25 in the open-source framework, 4 weekly crons, local semantic memory. Crons active since May 2026: 88 executions without a single error as of 27/05/2026.
Every sensitive action — pricing, publishing, deletion — requires the merchant's explicit approval over Telegram. Never "full autonomy".
DONE = the agent autonomously handles the cases defined at scoping; anything outside that scope is routed to a human; the human-takeover rate is measured over 2 weeks.
If a human still has to approve every output because the task is critical, we say so at scoping — and sometimes the honest answer is to stay at tier 1.
Your data is your asset: client history, quotes, procedures, past conversations. Today that knowledge is buried — finding anything relies on someone's memory, not on a system.
What you receive
Your corpus (documents, history, procedures) structured into a usable base.
A question-answering system at the simplest level that solves your pain — from a well-indexed knowledge base to semantic search. No vector database just because it's fashionable.
Sourced answers: every answer cites the document it comes from.
Optionally, for sensitive data: open-source models and controlled hosting — your documents don't leave for a third-party service.
Corpus updates without manual work.
Proven example
The architecture is proven on our agents in production: a RAG support agent published as open source, and the e-commerce agent's semantic memory runs locally on the client's server — embeddings computed on the machine (36 MB model, SQLite storage), no data sent to any embeddings cloud.
This tier has no dedicated client deployment yet. We would rather say it than dress it up.
DONE = the test questions defined at scoping return correct, sourced answers, and the corpus updates without intervention.
If what you actually need is to find 10 well-organised documents, a well-kept shared folder is enough — and we will tell you so.
Your need goes beyond automating one flow: you need a complete operational tool — inventory, bookings, back office — that generic SaaS doesn't cover, or covers only by stacking subscriptions. And that tool has to automate from day one, not just store data.
What you receive
A web and/or mobile application in production (deployed on Vercel; iOS/Android builds via EAS).
A back office with roles (staff, manager, admin).
Automation built in from the start: crons, WhatsApp notifications, invoicing, alerts.
The code delivered, with handover documentation.
Proven example
Inventory management for a restaurant: a web application (Next.js, Supabase) in production on Vercel, plus a native iOS/Android application (Expo) on the same backend — guided evening count, purchase list generated automatically from thresholds, WhatsApp notification, daily safety cron, three role levels. Web application verified in production on 02/07/2026; iOS/Android mobile application built (EAS), same backend — store submission pending.
Two complete hospitality websites live — a guesthouse and a hotel — with an ElevenLabs voice agent, back office, and PDF invoicing.
DONE = the application is in production, the team uses it daily, and the handover documentation is delivered.
If an existing SaaS covers 90% of your need for €30 a month, take the SaaS — we will simply wire it to the rest (tier 1).
The legitimate fear with any contractor is the never-ending project: fuzzy scope, drifting invoices, dependency. The method exists to neutralise it.
Deterministic workflow when possible. An AI agent only when necessary. Sometimes the right answer is: you don't need AI.
01
Discovery: the trigger audit
We don't start with "where can we put AI". We start with a structured interview of your operations: you walk through a typical week, and every predictable event — incoming order, email, deadline, stock threshold — is logged as a candidate trigger. The trigger is what identifies the automation worth building; without this audit, you end up automating the spectacular instead of the useful.
You
1 to 2 hours of interview, and read access to the tools involved.
Us
running the interview and mapping the flows.
Deliverable
a map of your triggers, ranked by estimated gain and by risk.
02
Scoping: one metric and a definition of done
Before the first line of code, every automation is tied to one observable metric — time, errors, delay — and to a definition of done written into the quote. No defensible metric, no project. This is also the step where we tell you, when it's true, that a spreadsheet is enough.
You
approve or amend the scope.
Us
write the scope, the metric, the definition of done.
Deliverable
a fixed quote on a defined scope.
03
Build: the simplest system
A deterministic workflow produces the same result on every run. An AI agent introduces judgment — and therefore variance. We use it only where judgment is required; everywhere else, deterministic wins. The build moves in short iterations, tested on your real cases, not on demo examples.
You
provide real cases and settle the business trade-offs.
Us
build, test, document.
Deliverable
the system in pre-production, proven on your cases.
04
Production, measurement, maintenance
Go-live, then the metric defined at step 2 is measured over 2 weeks of real operation. After that, maintenance mode: monitoring, fixes, and the handover documentation that makes you autonomous — the code is delivered and can be picked up by any developer.
You
use the system and report gaps.
Us
monitor, fix, hand over.
Deliverable
the measured metric, the code, the handover documentation.
What we will also tell you, whenever it's the case
When a three-node n8n workflow is enough. When a spreadsheet is enough. When AI adds nothing to your problem. A call that ends with "you don't need us" is a successful call.
Confidentiality
Your data stays within the agreed boundary: hosting decided at scoping, no client data sent to any non-contracted service, logged access. For sensitive data, an option exists: open-source models running on your infrastructure or on a dedicated server. This is not theoretical: our e-commerce agent's semantic memory computes its embeddings locally on the client's server (36 MB model, SQLite storage) — nothing leaves for an embeddings cloud.
GDPR compliance: data minimisation, contracted processors.
No financial advice. Our algorithmic-trading lab is a technical demonstration — Telegram signals to MT5 orders, pre-trade checks — never a service, never a promise of returns.
No numbers promised before they are measured: the figures published on this site come from real, dated systems, and any future figure will come from yours.
Every system is sized on your actual operations: the price is set after scoping, on a written scope — a fixed quote, not an open-ended retainer. Every quote includes the success metric and the definition of done for the tier involved: you know what you are buying, and what the end looks like.
EVERY QUOTE INCLUDES THE DEFINITION OF DONE.
A scoping call. 30 minutes. Free.
We identify your tier — or we tell you that you don't need one. No sales deck: your operations, your triggers, one recommendation.