The most-asked question of 2026: "Where do we put AI in the business?" The most common mistake: trying to put it everywhere at once.
On KERNEX IT projects with AI integration, we work to three rules. Sharing them so you don't burn €40,000 on a POC that dies in iteration two.
Rule 1: AI solves a specific problem, not a "general" one
We don't start with "let's add AI." We start with "who on your team loses the most time to repetitive work, and what exactly are they doing?"
The answers are always oddly specific:
- A sales manager spends 90 minutes a day pulling threads from WhatsApp and CRM.
- An accountant re-keys 200 PDF invoices a month.
- Marketing writes ad copy variations and burns four hours a week iterating.
Those are the entry points. AI fixes that specific thing, not "internal communication."
Rule 2: Don't change the tools the team already uses
The biggest mistake in AI integration is forcing the team to learn a new platform. If sales lives in WhatsApp and HubSpot, AI shows up in WhatsApp and HubSpot. We don't move them into a separate "AI dashboard."
In practice that's webhooks, API integrations, and background automations. The user sees the same flow — except part of the work now does itself.
Rule 3: Human in the loop where errors cost
AI writes well, but it sometimes writes confident nonsense. For client emails, contracts, financial decisions — we always keep a human in the loop. AI output arrives as a draft, a person approves or edits, then it ships.
Where the cost of an error is low (internal categorization, summarizing notes, meeting minutes) we let it run autonomously.
The stack we use
For anyone asking: Claude / GPT via API, vector store for context (Postgres + pgvector for small data, dedicated for large), a simple orchestrator (Inngest or cron + queue), monitoring (Langfuse). Self-hosted where data demands it.
A few concrete examples
- Retail company, 40 staff. AI reads incoming invoices (PDFs) and pushes them coded into 1C. 200 documents a month at 8 minutes each → 26 hours saved, under one hour of QA.
- Real-estate agency, 12 agents. AI catches missed calls, summarizes them, opens deals in the CRM and sends follow-up. Lead-to-reply rate: from 38% to 91%.
- Video studio, 8 people. AI processes raw footage, flags good takes, generates three rough cuts. The editor sits down to a pre-assembled timeline.
None of the projects asked the team to learn Python. All of them paid back inside four months.
Where to start
If you read this far and thought of a specific problem in your business — call us. We can usually tell you in an hour whether it's feasible and what it would cost.