There are two ways to bring artificial intelligence into a company. The first is to embed it at the core of the systems: connectors into the customer database, synchronisation with business tools, migration of historical data, proprietary formats. The second is to place it alongside — an external harness, clipped onto what already exists, and detachable. KOAEE chose the second, and this article explains why that choice commits the supplier more than it commits the client.
What "forward engineering" means here
The term comes from software engineering: building forward from what is specified, rather than taking the existing apart in order to reproduce it. Applied to an AI deployment, it changes the order of operations. You do not begin by opening the information system to let a model in; you begin with the process as it runs today, and you build outward.
The consequence is concrete: nothing that already exists is rebuilt to accommodate the tool. The website stays the website, the booking system stays the booking system, the procedures stay the procedures. What is added is a conversational layer placed on top, which reads what it is given to read and writes nowhere else.
Lock-in is not malice, it is a slope
Deep integration is not a trap that someone sets. It is the natural result of a commercial movement: the more a system is wired into others, the more useful it becomes, and the more useful it becomes, the more expensive it is to leave. Every connector added is an immediate functional gain and an additional dependency later on. The exit bill is not paid at installation; it is paid the day you want to change your mind.
That slope always produces the same symptoms: data reshaped into a proprietary model, business rules rewritten inside the supplier's tool rather than documented elsewhere, teams trained on one interface. None of it is reversible in a day. This is not an accusation, it is a mechanism — and it is precisely the one an external harness avoids.
What an external harness looks like
The same event — wanting to change tools — does not carry the same cost depending on the architecture chosen. On one side a chain of operations, each link of which has to be negotiated; on the other, a single gesture. It is the difference between a load-bearing wall and a piece of furniture.
"We lock no one in": three checks, not a promise
A promise of reversibility is worth only what you can test. Here are three questions worth asking any conversational AI supplier, including us.
First, how do you leave? With KOAEE, the assistant is installed with one line of script on the pages where it should appear, and removed by deleting that line. There is nothing to uninstall elsewhere, because nothing was installed elsewhere. The next day, the website is exactly the one you had before.
Second, where does the data go? The question is not rhetorical and the honest answer is rarely a simple "with us". Our servers are in Germany, in Frankfurt. Language processing runs on Mistral, and for French and English so do listening and voice: nothing leaves Europe on those two languages. On the others, the voice still relies on engines outside Europe. We publish that rather than rounding it off, on our transparency page.
Third, who owns the rules? The instructions that govern the assistant — what it should say, what it must not promise, how it steers a conversation — are the client's asset, not the supplier's secret. They are written in readable text, not encoded into an interface. A client who leaves takes them along.
What this choice costs us
It would be dishonest to present the external harness as superior on every count. It is not. A deeply integrated AI sees more: it can cross-reference a purchase history, trigger an action in a business tool, write into a record. An external harness sees only what it is shown, and can act only where it has been explicitly connected. That is a real limitation.
We accept it because the trade-off seems more valuable to a decision-maker: the ability to change your mind. A tool you can remove in one line is a tool you keep by choice, not by exit cost. And for the supplier it is a discipline: you have to be useful every month, since nothing holds the client in place.
The regulatory framework points the same way
The European regulation on artificial intelligence requires, for conversational systems, that a person knows they are addressing a machine. That transparency obligation is easier to meet when the assistant is an identifiable layer placed on a website than when it is scattered across a dozen internal systems where no one knows exactly what is answering. Technical reversibility and regulatory legibility are two effects of the same architectural decision.
It is also what lets us name our engines plainly, and above all change them: because nothing was melted into the client's systems, swapping one model for another asks nothing of them. The map of engines above is not a permanent commitment, it is a state at a given date — and that is exactly what an external harness makes possible. Those choices may change. The important point is that they can change without the client having to redo anything — because the client was never wired to the model, only to the harness.
Source: Eurostat, “Artificial intelligence by size class of enterprise” (isoc_eb_ai) — enterprises with 10–249 employees, France. Retrieved via API on 30 August 2026. The survey publishes neither 2022 nor 2026: the series stops at the latest available year.
In one sentence
We do not try to become indispensable through entanglement. An assistant that installs in one line and is removed in one line forces whoever sells it to stay useful — and leaves whoever buys it the one thing no contract can grant: the freedom to walk away.