Our engineers work on site with you, on your real data, alongside the people who will use the tool. We build the interfaces, connect the data, define the interaction experiences, and automate the processes. Our own platforms run at more than 100 clients.
Partner programmes
They work with Kabaun
Four steps, on site, with the people who will use the tool.
What runs the business, as it is.
They observe the real work, alongside the people doing it.
Every project goes through the same acts.
Built on your real data, put into service.
Everything below on this page falls into one of these four boxes. If you do not know which one you are in, that is exactly what the diagnostic is for.
When the tool does not exist
You run your business in a spreadsheet, or in software that speaks a different trade than yours. We build the platform that is missing, and keep it alive in production.
When one precise task takes too long
The question you are asked forty times a day. The agent absorbs it, without ever touching what commits you: a price, an availability or an address stays computed by the server.
When a flow repeats with nobody to hold it
Processes that run without being launched: a brief that becomes a published site, a follow up sent at the right moment, data collected as it comes. Nothing to open, nothing to click.
When your teams must hold the wheel
We pass on what we practise daily: framing a use case, writing instructions that hold, and knowing what you never hand to a model.
What people ask us, word for word
Four sentences we were told, almost word for word. Each one led to a project.
« We have three dashboards, and nobody knows which one is authoritative. »
The tool grew by accretion. Every department got its own screen, and nobody kept the overview. The answer was not a fourth dashboard, but merging them into one.
« The same question comes back forty times a day at the welcome desk. »
Two people tied up for three days pointing to hall B. The knowledge exists, it is in the programme and on the map: it simply is not reachable by the visitor at the moment they ask.
« The quote goes out, then nothing. Who follows up? »
The sales cycle lives in a spreadsheet, an inbox and the memory of two people. Nothing is wrong, everything is fragile, and the unpaid invoice surfaces three months too late.
« We will know our carbon footprint in March. Last year's. »
An annual re-entry campaign produces an estimated figure, arriving too late to steer any decision, and resting on whoever ran it. If knowing takes a year, you are not steering.
Where to start →In all four cases, the problem was not a missing tool.
It is the only question that really matters when you put AI into an organisation. Our answer is structural, not reassuring: the agents orbit the person who decides, they never take their place. Each one absorbs a repetitive task and hands back control the moment judgement is needed.
"AI" covers three very different things: general purpose large models, compact specialised models, and automation with no language model at all. Choosing between the three is a project's first impact decision. Our framework is written and dated.
We write what the system must and must not do before choosing the tool. If rules or a lexical search are enough, no model. If a model is needed, the smallest one that does the job.
The first impact decision, before any optimisationThe agent answers from a closed corpus and cites where the answer comes from. Every exchange is logged and anonymised. Anything that binds the company is computed and checked server side, so it can be replayed.
An answer can be traced backThe agent introduces itself as an agent and says who to turn to: that is an obligation under the European AI regulation. Humans keep advice, sales, arbitration and relationship; AI routes, it does not replace.
Regulation (EU) 2024/1689 and GDPR as the baselineTokens consumed, cost, carbon estimate with the factor used and its date. Short answers by instruction, context limited to what is needed, no model call for what a script does.
Never a saving percentage we have not measuredWe build on these environments and we know how to move from one to another. That is what makes our independence rule workable: the model call is abstracted, a fallback without a model exists, and no project depends on a single vendor.
Our consultants' certificates.
The four Ds framework: delegation, description, discernment, diligence. The human skills that working with a model requires.
The four properties of a generative model: next token prediction, knowledge, working memory, steerability. What stays true as the models improve.
Running long sessions: scoping the work, writing instructions that hold, enforcing the rules that cannot be skipped, handing off to sub agents.
The open protocol that connects a model to an organisation's tools and data. The advanced topics: building your own servers, transports, and controlling what you expose.
Designing, automating and deploying reliable workflows with the OpenAI tools, Codex and the API.
The same craft on the ChatGPT side: designing uses that hold up once they are in service.
The first one costs nothing and commits you to nothing. It serves us as much as you: it says whether the project deserves to exist.
What can be automated, what must not be, the order of magnitude of the gain, and a straight answer if it is too early for you to equip yourself.
Design and go live, technical guardrails set from the start, testing and a security check before delivery. Priced after the diagnostic, never before.
Custom business agents, automation and training over time, with one contact who knows your system.
No. An agent handed a judgement call is a costly mistake waiting to happen. What we absorb are the repetitive tasks with no judgement: the question asked forty times a day, the search through files, the summary to copy out again.
It cannot get anything that commits you wrong, because it is not allowed to produce it. Price, availability, reference, address: computed or checked by the server, picked from a closed list, then re-read before display.
In France for hosting, in Europe for the rest. Our agents' conversations are stored outside the web root, with a short retention.
Most of the time, no. An agent plugs into what already exists, and a new application is only justified if the tool does not exist or speaks a different trade than yours.
It depends on the scope, and we say so after the diagnostic rather than promising a deadline before having looked. The first version arrives sooner than expected and does less than was asked.
We will tell you. Sometimes the right answer is to wait, or to fix an organisational problem first that no software will solve. That is also why the diagnostic is free.
We always start with a diagnostic: what can be automated, what must not be, and what it really costs.
Rather talk it through? Book a demo slot