Bombos teamBy Piet Baudoin · 12 October 2026
AI is fast enough. What holds a business back now is that nobody has agreed who does what and who checks it. A Rotterdam print company found that out in three months, Albert Heijn solved it in a back room, and Microsoft is now building it into Windows. Then we look ahead.
What happens when everyone in your business starts building with AI?
HelloPrint sells print online from Rotterdam, with about a hundred employees. This spring the company decided to rebuild its thirteen-year-old platform, entirely with AI. Everyone got their own piece and built it themselves, including the head of finance and a commercial manager. Dan de Vries, Head of Product & Tech, told ICT Magazine, a Dutch trade publication (translated from Dutch): “Everyone can make code with AI and send it to production, so why wouldn’t they do that themselves?”
After two to three months the company changed course. At one point about 150 changes were waiting for someone to review them. Pages that should have followed the same rules looked completely different, depending on who had built them. De Vries himself no longer knew who was working on what, because he was building too. His explanation (translated from Dutch): “Everyone can now build something and even put it into production, and so everyone also thinks they are capable enough to do that.” And: “Without those guardrails you put the cart before the horse”.
The fixed roles are back, with planning and meetings. The code still comes from AI. People from the business side no longer put anything live themselves, but they do make an example in a few hours that shows what they mean. De Vries (translated from Dutch): “It is still human work”.
You do not have to build a platform to recognize this. The 4 October edition reported that almost half of employees use a private account for work. That is the same thing on a small scale: everyone has their own way of making a quote, an email or an overview with AI, and nobody knows what the others have made or who checks it.
Source: ICT Magazine, Kim Loohuis, 8 October 2026
Should you know what the work is before you put AI on it?
Albert Heijn, the largest Dutch supermarket chain, did it the other way around. At World Summit AI, Noortje van Genugten, VP Product at the chain, said (translated from Dutch): “If we can, we prefer not to use AI”. The hunch was that an AI assistant could help shelf stockers with questions from customers. But the team built only the front of an assistant in the employee app, with no AI behind it, and sat in the back room of a few stores to answer all questions by hand. “Because maybe AI was not the solution at all”, she says (translated from Dutch).
It turned out it was: most questions were about products. Only then did Albert Heijn build the assistant, which according to Van Genugten now handles about 80,000 conversations a week.
ABN AMRO, a Dutch bank, is making the same turn. Hiek van der Scheer, the bank’s Chief Data and AI Officer (translated from Dutch): “Last year my goal was mainly to get a number of applications into production”. Now an AI application only gets funding there if it demonstrably delivers something. He does not find it easy (translated from Dutch): “That measuring is quite a challenge, I must say.”
That order works without a supermarket chain too. First keep track for a week of which questions really come in, and only then decide what an AI agent takes over. If you get the same ten questions every week, you know after that week where to start.
Source: ICT Magazine, Kim Loohuis, 9 October 2026
Who pulls the emergency brake when an AI agent goes too far?
On 7 October, Microsoft made a closed-off space for AI agents generally available. An administrator or software maker sets which files and which internet addresses an agent may reach, and the system holds that boundary itself, whatever the agent decides. The sentence it all turns on: “An agent cannot be its own security authority.” Microsoft gives the example of an agent that has to update a website and decides that changing the server is the fastest way. Reasonable from the agent’s point of view, and still more than it was allowed.
On Saturday, CEO Satya Nadella went a step further, in a long piece on X:
We must assume a model is compromised and contain it from the start. Think of it like an emergency brake. An authorized person should always be able to pause or shut down a model mid-task.
@satyanadella · 10 October 2026
Assume the model has already been fooled, and make sure someone can stop it in the middle of a task. Microsoft sells Windows and the agents that run in it, so this is also advertising. And not everyone is impressed. On Hacker News, amluto, an account with almost 25,000 points, wrote:
Sadly, skimming the docs about how the different backends work makes me think that almost this entire project was done by a recent-gen LLM that interpreted its instructions as “make these things work at all costs” instead of “make a considered design that cleanly and securely fits its use case”.
amluto on Hacker News · 10 October 2026
His impression: the fence around the agent was itself built by a model, in a hurry. If that is right, it is HelloPrint, but at Microsoft. For you it is the same question as should an AI assistant send email on its own: who says yes, and who can stop it.
Source: Microsoft, 7 October 2026
Where are we headed?
Albert Heijn sat in the back room to see which questions really came in. At ABN AMRO an AI application only gets funding if it demonstrably delivers something. HelloPrint calls its months of searching “far from lost time” (translated from Dutch). Three companies, one movement: first know how the work is really done, then show what it delivers.
I think this is the work everything will turn on in the coming years. Soon the AI of your business will talk to the AI of your customer and of your supplier. It can only represent you well if it knows how you work, what you offer and what the customer really asks. The more it knows and the more is connected, the better it does.
In five years an agent will choose your supplier, and an agent does not read advertising, it reads proof. Today whoever has the most information about a subject still wins online. That is an in-between step. AI models will recognize which business knows something of its own and truly contributes, and which business repeats what was already there. Then the winner is not the business with the smartest AI, because everyone has that. The business that has written itself down best and truly adds something gets found. That is good news for everyone who does their work well.
What does this mean for your business?
The three stories say the same thing: AI is fast enough, and the agreements around it are the work. Which work comes off your plate, who checks, and who can stop it. That is how Bombos starts too. We sit down with you, first look at how the work is done by hand today, and start with one workflow, the shared inbox, for example. An AI agent reads what comes in, finds what belongs with it and prepares the reply. You or your colleague approves it, and that approval remains the standard. No email is left waiting, and your people do the work that brings in money. Then your team teaches Bombos the next task itself. Then you finally have time to work on your business, instead of in it.
Sources
Every source was opened on 12 October 2026 and every quote appears in it word for word. Two caveats. Nadella’s sentences are in the article attached to his post on X, not in the post itself. For amluto, only the number of points the account has on Hacker News is visible, not who is behind it. The Dutch quotes were translated for this edition.
ICT Magazine on HelloPrint (8 October 2026) · ICT Magazine on World Summit AI (9 October 2026) · Microsoft on Execution Containers (7 October 2026) · Satya Nadella on X (10 October 2026) · amluto on Hacker News (10 October 2026)