# Own the heart: local-first AI operations

> Models, runtimes and tools will keep changing. Your business knowledge should not change hands with them.

- Category: AI in practice
- Published: 2026-09-29
- Author: Asaf Eyzenkot (Suf Zen), founder of Realization
- Page: https://realization.world/insights/own-the-heart-local-first-ai-operations

## The short answer

An AI operation has one durable asset: the business's own knowledge, history and identity. Everything else changes fast, including the models, agent runtimes, channels and even the interface. So keep the durable part in plain files the business owns, and treat everything else as a swappable adapter. We call the durable part the heart.

## Chatbots are not an operation

Most small and mid-size businesses adopted AI as a clever text box. Every chat starts from zero, so voice, clients and past decisions must be re-explained each time. The answers are one-shot, so nothing is followed through. And the knowledge ends up living inside a vendor, with no audit trail, no approval gates and no ownership.

## Six problems stacked on each other

An "AI employee" is six hard problems, each depending on the one below: business knowledge that is structured and portable; memory that decides what to recall, when and how cheaply; model economics across several providers; orchestration of multi-step missions with handoffs and retries; security against prompt injection and leaked secrets; and governance over who may do what, with approval and audit.

Get knowledge wrong and memory is noise. Get memory wrong and orchestration drifts. Skip governance and nobody can trust the output.

## What the heart contains

In RealizeOS, the heart is three things. A knowledge base written in plain markdown that people can edit, organised into six layers: foundations, agents, domain knowledge, routines, insights and creations. An append-only event log of what acted, when and why. And identity files that give each agent a stable role and voice.

Because it is plain text, the business can move it to a different model, runtime or vendor without losing anything.

## Route by task, not by habit

Once knowledge is portable, model choice becomes an economic decision. Classify each task, then send it to the cheapest model that does it well: a fast model for lookups and formatting, a strong writing model for content, the most capable model for strategy, and a local model for sensitive data that must not leave the server.

Cost is tracked per step, so the routing policy can be checked against real invoices.

## Approval before anything consequential

Agents should act freely on bounded, reversible work and stop for approval on anything that moves money, sends email or changes records. Even the system's own background learning cycles propose updates into an inbox, and a person approves them before they become "truth".

## Where to start

Start with the knowledge, not the agents. An orderly, owned knowledge base makes every later model better, and it survives every change of vendor. The RealizeOS case study shows the architecture we run our own operations on.