Most companies do not have an AI problem. They have an AI pilot problem. A team runs a promising experiment, it works in the demo, and then nothing compounds. Six months later the pilot is a browser tab nobody opens.

I run AI as a system, not a pilot. It drafts my emails, prepares my meetings, monitors its own health, and researches companies while I sleep. None of that is exotic. It is the result of five layers assembled in the right order. Skip a layer, or build them out of sequence, and the whole thing stays a demo.

Here are the five layers, what each one does, and why the order is the part that matters.

Five layers, in build order. Start with governance. Each layer makes the next one possible. 1 Governance What is allowed, which data, who signs 2 Context, the harness Your files and rules, handed to the model every run. The moat. 3 Agentic workflows Defined steps with a verification step and a human sign-off 4 Agentic OS Orchestration: agents run, hand off work, report health 5 Agents overnight On a schedule, on a server. Briefed at night, ready by morning
The order is the strategy.

1. Governance comes first, not last

The first question is not which model to use. It is what the system is allowed to do. Which data it can read. Which actions need a human signature. Where the data is hosted, and under whose jurisdiction.

I wrote this down before I automated anything. A one-page brief stating what the agents may touch, what they may never do without me, and where everything runs: my own server, EU-hosted, no third party training on my data. It reads like a constraint. It is what makes the rest safe to build. A private equity client will not let an agent near a data room until that page exists. Neither should you.

One rule on that page does most of the work: agents never send anything externally on their own. They draft. I send. The system can write fifty replies overnight and not one of them leaves without me reading it.

2. Context is the moat

A model on its own is a clever stranger. It knows everything in general and nothing about you. The work that makes it useful is not prompting. It is context: the files, the past decisions, the house style, the rules, assembled and handed to the model every time it runs. The industry calls this the harness, or context engineering.

This is where the real advantage sits, and it is the part a competitor cannot copy. Anyone can license the same model. Nobody else has your decade of proposals, your pricing history, your client notes, your way of phrasing things. Feed that in, structured, and a commodity model produces work that sounds like you and reasons from your evidence. Starve it of context and you get fluent, confident, generic output that helps no one.

In practice I keep a structured library of my own material, my writing rules, and the live state of each project. Every agent reads the relevant slice before it does anything. The model is rented. The context is owned. The owned part is the business.

3. People decide what is worth automating

There are two failure modes. The first is automating the wrong thing, a task that was cheap anyway, and feeling busy. The second is trying to automate judgement, the one thing that should stay human.

The rule I use: AI does the gathering and the drafting, people own the decision and the sign-off. My system drafts every reply. I send none without reading. It prepares the meeting brief. I form the view. The moment a task involves weighing conflicting evidence and making a call someone is accountable for, a human stays in the loop, by design, not by accident.

There is an organisational reaction worth naming. When work starts disappearing, people get nervous. Handle it openly. The honest framing is that the boring half of the job goes and the half that needs a person gets more room. Pretend otherwise and you lose the team before the system pays off.

Talking to a model is not the same as a workflow. A CHAT TOOL You prompt Model replies you re-prompt, and re-prompt Output depends on whoever is at the keyboard. Nothing repeats. Nothing runs without a person typing. AN AGENTIC WORKFLOW Trigger Gather Draft Verify Sign-off HUMAN Deliver Runs the same way every time. It checks its own work, and escalates the judgement call to a person. Stenfert Kroese Consulting
The verification step is what separates a workflow from a confident error generator.

4. Workflows, not chat

Most people use AI by talking to it. You ask, it answers, you ask again. The output depends entirely on who is at the keyboard and what they remember to type. That does not scale and it does not repeat.

An agentic workflow is different. It is a defined sequence: a trigger, a step to gather the right context, a step to do the work, a step to check the work, and a point where a human signs off before anything leaves the building. Same steps every time.

The checking step matters most. An AI workflow that does not verify its own output is a confident error generator. The prose is fluent, the structure is logical, and the mistake is invisible from the output. Build the verification in, or you will ship the errors at speed.

5. The agentic OS

One workflow is useful. The shift happens when several of them run together and maintain themselves. That is what I mean by an agentic OS: an orchestration layer that runs each agent on a schedule, keeps shared state in plain files anyone can read, and lets one agent hand work to the next.

Mine writes a health report every thirty minutes. If an agent fails, the report turns red and I see it. There is a kill switch, one command, that stops everything. That sounds like overhead. It is the difference between a system you trust to run unattended and a pile of scripts you have to babysit.

The work happens while you sleep. Brief the system in the evening. Close the laptop. The agents run on a server, not on your machine. VM RUNS UNATTENDED, EU-HOSTED 18:00 You brief the system Inbox sync Draft replies Monitor health Meeting prep Research brief 06:00 Finished work waiting A single server costs less than one monthly software subscription.
The unlock is not the hardware. It is moving the work onto a schedule.

Agents that work overnight

The part that surprises people: the agents do most of their work while I am asleep. They run on a small server, not my laptop, on a schedule. I brief the system in the evening and close the laptop. Overnight it syncs my inbox, drafts the replies that can be drafted, prepares the next day's meetings, monitors itself, and runs any research I have queued. By six in the morning the work is waiting, reviewed and ready for me to judge.

Nothing here needs a large team or a big budget. The server costs less than a single monthly software subscription. The unlock is not the hardware. It is moving the work off the human's hands and onto a schedule, so the system compounds whether or not anyone is watching.

The Order Is the Strategy

Build the layers bottom to top: governance, then context, then the human rules, then one workflow, then orchestration, then the schedule. Most failed AI efforts start in the middle, with a clever workflow that has no governance under it and no context inside it. It demos well and dies quietly. Get the order right and the rest follows.

How to start

You do not need all five layers on day one. You need them in the right order. Start with the page that says what AI is allowed to do in your business. Then pick one expensive, repetitive process and give the model your real context for it. Build a single workflow with a verification step and a human sign-off. Get that running on a schedule. Only then add the second.

The companies getting value from AI are not the ones with the best model. Everyone has the same models. They are the ones who built the system around it: the governance, the context, the people, and the patience to do it in order.