What AI integration is and how it's done

How integration differs from using a general tool

In most companies today, AI gets used like this: someone pastes a text into ChatGPT, tidies it up and copies it back. That works. For drafting an email, shortening a text or gathering ideas, a general tool is enough, and you don't need an integration for it.

The limit is this: a general tool doesn't know your company. It doesn't know your price list, what you quoted a customer last time, or which product is made in which size. You have to explain all of it again every time, and the work never builds up anywhere. On top of that, when you ask about something it hasn't seen, it doesn't say "I don't know"; it produces an answer that looks reasonable.

In an integration, the AI is connected to that information: your product and price data, your accounting software or ERP, your inbox, your shared drive. Nobody has to go to a separate window; the work arrives already prepared.

Where it makes sense

Integration pays off in work that repeats, follows rules and rests on data:

• A draft quote from an incoming request. A request that arrives by email or form is read and matched against your product and price data, and the draft quote lands in front of you ready. You check it, correct it and send it.
• Document entry. Invoices, delivery notes, order forms, data sheets. The system reads them, pulls out the fields and writes them into your software. A line it isn't sure about isn't written; it flags it and asks you.
• Questions your team asks of its own documents. Procedures, contracts, data sheets. The answer comes with the document it rests on.
• A visitor's product question. An advisor that answers from your product catalog, on your site or on WhatsApp.

There are places where it doesn't make sense, too. For a job done a few times a month, the setup cost never comes back. The decision itself isn't automated either: your pricing policy, your discount limit and what each customer gets are yours to decide. The system only applies them.

How it's done, step by step

1. One job is chosen. The one that repeats most and takes the most time. Taking on everything at once is the most common way to finish nothing.
2. Today's state is measured. How many minutes it takes now, how many people it passes through, where the mistakes happen. The only way to say "it got better" later is to know what it was like before.
3. Data and connections are mapped. We work out which software to connect to. If something can't be connected, you're told up front.
4. The system is built, and the approval step stays. No quote or message goes out without a person approving it.
5. It runs side by side for two weeks. The system and the team do the same work together. Differences surface in this round and get fixed.
6. Back to the measurement. If the time dropped, we move on to the second job. If it didn't, the system is removed, because keeping an automation that doesn't work helps nobody.

The most common mistakes

• Taking on every job at once. As the scope grows, nothing gets finished and the team stops trusting the system.
• Starting without measuring. If nobody knows the before, nobody can argue about the after.
• Removing the approval step. Taking the person out of the loop to gain speed brings the whole system to a halt at the first mistake.
• Expecting the model to know what isn't in your documents. A general model can make up information it hasn't seen. That's why where the system answers from has to be designed from the start.

The real question: is the answer right

An integration is worth what its results are worth, not how fast it runs. A draft quote with the wrong price, or an invoice line read wrongly, costs more than not automating the job at all.

The system we build works only from your data and your documents. It records which document each answer rests on. It doesn't produce an answer your documents don't support, and it doesn't write a line it isn't sure of; it asks you.

We tested this approach where a wrong answer is most expensive: law. Our real-estate law advisor, Geolex, was built on 7,500+ legal records, 46 statutes and 172 court rulings and cites the relevant article in every answer. The engine that would run in your company is the same one.

Where your data lives

Your documents live on your own server or on one we run for you. In a cloud setup, only the passages relevant to a question go to the AI provider to produce the answer, and they aren't used to train its model. If you prefer, an open-source model runs on your own server as well, and the data never leaves it. Who gets access to what is decided together during setup.

If you build it with us

1. We start with one job. The second one starts after you've seen the first one work.
2. The before is measured. Before setup, how long the job takes today is put on record in writing.
3. Nothing goes out without approval. We won't build it any other way, even if you ask.
4. The first two weeks run side by side. The system and your team do the same work together.
5. If the number doesn't drop, it comes out. If the time didn't get shorter, we remove the system.
6. Your existing software stays. We don't replace a system that works; we build the bridge to it.
7. Training is part of the setup. Your team is trained, and the new workflow is handed over in writing.

What determines the price

What an integration costs depends on the state of the work today. The main things that shape the quote:

• How many times a month the job is done. For a job done rarely, the setup cost never comes back; if that's your case, we show you the numbers and tell you not to go ahead.
• The state of the data. One tidy list and scattered spreadsheets, old quotes and emails are not the same job.
• How many systems to connect. Accounting, ERP, e-invoicing: each one is a separate bridge.
• How many people will use it. Training and access rules affect the scope.
• Where the model runs. In the cloud or on your own server.

That's why, instead of a fixed price list, we work out the scope with you by looking at how the work runs today.