AI goes into the software you already use.

There is no platform to adopt and no migration to run. The ERP, the CRM and the portal stay where they are: the model comes to them, with the access to data that you decide.

The fastest way to waste an AI budget is to buy a new tool and ask people to go and work in it. The fastest way to see a return is the opposite: leave people where they are and bring the model to them.

An LLM on its own can talk but knows nothing about you. It becomes useful when it can read an order, look up a customer, open a ticket — that is, when you give it tools. That is where most of our work ends up.

How we do it

Your stack is left alone

Rails, Python, Go, Node, PHP. The model is added as one more service, with the same authentication and the same logs as everything else.

Tool-calling against your data

The model does not guess: it calls your functions. Reading an order, finding an item, writing a note are tools we define together, with the permissions of whoever is using them.

Without the data ever leaving

If data cannot leave the company, we use open-weight models on your own machine. Quality is below the frontier models, but on many tasks the difference does not show.

Costs that stay predictable

Caching on repeated prompts, small models where small models are enough, a spend ceiling per environment. The bill at the end of the month should not be a surprise.

Use cases

02 · LLM integration

Quotes and estimates

Specifications, drawings, three-line requests: for anyone working to order, every quote is a job in itself. AI finds similar past jobs and proposes line items, times and costs to validate.

10 · LLM integration

User behaviour

Application logs and databases already show how people use them, but nobody has time to read them. AI reads them, writes a report and suggests what to improve.

Questions we get first

How do you integrate AI into an ERP that already exists?

From the outside, almost always. The ERP already exposes an API or a database: we build a layer on top that the model can call, without touching the core. If the API is missing we add one; if the ERP is closed, we work on what it exports.

Does our data end up training somebody's model?

On the business plans of OpenAI and Anthropic your data is not used for training, and that is contractual. If that is not enough, we go on-premise with open-weight models, and then the data never leaves your network.

Which model do you use?

Whichever one the use case needs. We write the code so the model is replaceable, because over the life of a project something better or cheaper almost always comes out.

How long before we see something that works?

A prototype on your real data takes two to three weeks. Its job is to let you decide with numbers in front of you rather than with an impression.

Tell us the process. We will tell you if it makes sense.

A free 30-minute call, with no commercial follow-up unless you want it.

Request a call