Introducing AI in 90 days without a big-bang project

I’m Denis Shokhirev, Agentic AI Systems Architect based in Freiburg im Breisgau, Germany. My stack is Claude, Supabase, n8n, Doppler, and self-hosted Postgres. I build agent systems for B2B clients in logistics, fintech, and industrial automation. Last month, a client lost four lucrative leads over a holiday weekend—not because of tech, but because nobody saw the incoming forms in time. That’s not a demo problem. That’s a business problem.

Why 90 days? Business reality vs. AI hype

In the DACH region, year-long, million-euro AI projects are for corporates. If your business has under 200 employees, a failed AI rollout means real money lost and months wasted. Gartner found in 2023 that 85% of SME AI pilots never reach production ("AI Adoption Survey 2023", Gartner, link): the number one reason is overdesigning for the entire company instead of fixing one urgent process.

What can you actually automate in 90 days?

My approach: start with a single business process that:

• Happens daily (e.g. incoming lead forms, invoice triage);
• Directly affects revenue or risk (delay = lost sales);
• Is measurable (number of leads handled, response time, error rate).

Example: automating lead intake from the website. No need to build a company-wide chatbot. Instead, you can implement an AI system that:

• Collects all web forms into one table (Supabase/Postgres);
• Filters out spam and duplicates (Claude via n8n workflow);
• Notifies the right manager only for relevant requests (email, Slack, Teams);
• Records every action for traceability (audit log in Postgres).

In a recent logistics deployment, this setup handled 75% of inbound emails automatically within three weeks; only complex or unstructured cases stayed manual.

The architecture: simple, stable, and accountable

1. The workflow: n8n + Claude + Supabase

Everything runs on proven, affordable infrastructure:

• n8n — visual workflow builder, self-hosted in EU;
• Claude API — text analysis, spam/duplicate detection;
• Supabase/Postgres — data storage and audit logging;
• Doppler — secure key and secret management.

No proprietary vendor lock-in. Only tools that speed up the process and can be swapped if needed.

// Example n8n workflow (simplified)
{
  "nodes": [
    {
      "name": "Fetch Inbound Forms",
      "type": "httpRequest",
      "parameters": {
        "url": "https://your-form-endpoint"
      }
    },
    {
      "name": "Claude Spam Filter",
      "type": "httpRequest",
      "parameters": {
        "endpoint": "https://api.anthropic.com/v1/complete",
        "body": "{...}"
      }
    },
    {
      "name": "Store in Supabase",
      "type": "httpRequest",
      "parameters": {
        "endpoint": "https://your-supabase-url/rest/v1/leads",
        "body": "{...}"
      }
    }
  ]
}

2. Data stays in the EU

All processing runs in Germany or Austria. Claude (Anthropic) offers an EU endpoint; Supabase and Postgres can be hosted regionally. No data leaves the EU. If you need even stricter isolation, you can use an open-source LLM on your own server (quality will drop, but privacy is absolute).

3. When NOT to build your own system

If you handle fewer than 2,000 leads a month, don’t build custom software. Public SaaS CRMs like HubSpot or Salesforce already automate most workflows—at a higher price (from €50/user/month, checked August 2026), but with less hassle. I only customize AI when standard products can’t deliver your specific automation or integration needs.

Calculating ROI: hard numbers

Here’s how I measure value for owners:

• Manual: 5 minutes per lead × 800 leads per month = 66 hours/month;
• AI: 1 minute for exceptions, 85% handled automatically = 10 hours/month;
• Saved: 56 hours/month (~1/3 of a full-time specialist’s work).

This logic applies to finance, support, logistics. If you have less than 200 repetitive tasks a month, automation will not pay off in under a year—don’t bother.

Comparison table: custom AI vs. standard SaaS

FAQ

What does a 90-day AI deployment cost?

Depends on the number of leads and integrations. Most technical projects take 3–6 weeks, with infrastructure costs from €0–200/month (checked August 2026 for n8n, Supabase, Anthropic). No hidden fees—just your server and licenses.

What happens if the system fails?

Every action is logged in Postgres, so you can recover missed leads. If AI fails, the system alerts your responsible manager. For critical processes, I always keep a manual fallback in place.

Does my data leave the EU?

You can fully isolate the system: all storage in EU AWS or your on-premise server. Claude offers an EU-only endpoint; Supabase can be set to Germany or Austria. For maximum control, use an open-source LLM, but expect higher costs and longer deployment.

Who is liable for AI mistakes?

Business decisions remain your responsibility. I configure AI to only assist, not make final calls without human review. During the initial rollout, all edge cases are double-checked by staff.

When is it NOT worth automating?

If you have few repetitive processes or already use a CRM with built-in AI, a separate project adds no value. If automation won’t save you at least 30–40 staff hours a month, skip it.

Where do your leads get lost most often—in the initial intake, in assignment to staff, or during manual processing? I’d genuinely like to know.

I run a free 30-min stack audit for DACH founders building AI in regulated markets. DM me on LinkedIn or write to @ger_dennis_ai.