Which AI model do you use?

Whichever one fits the job and the rules you operate under, and we expect to change it. Models are the fastest-moving part of this and the easiest to swap. What takes the work is everything around the model: getting your data to it, checking what comes back, and deciding what happens when it is wrong.

Will our data be used to train someone else’s model?

Not under the arrangements we set up. Business tiers of the major providers exclude your content from training by contract, and where that is not enough we run a model on infrastructure you control. Which of the two applies is a question we settle before anything is built, not after.

Is this going to replace people?

In the projects we take on, it takes tasks off people rather than removing the people. The work AI is good at is the copying, the looking-up and the first draft. Judgement, exceptions and anything a customer will remember stay with a person, and the system is built so that person can see and correct what it did.

What is agent orchestration?

Using several narrow agents for one job instead of asking one general agent to do everything, with something coordinating them and checking the result. One reads the documents, one looks up the numbers, one drafts the answer. Each does a small job that can be verified, which is what makes the whole thing trustworthy enough to leave running.

What if the AI gets it wrong?

It will, so the design question is what happens then. Every workflow we build has a defined failure path: low confidence goes to a person, nothing irreversible happens without a confirmation, and every run is logged so you can see what it did and why. A system that cannot be wrong safely is not finished.

We do not have our information in order. Can we still start?

Yes, and almost nobody does. It does shape where we start: with messy information the first win is usually a narrow one, on the one process where the data is already good enough. If the state of the information is the real blocker, that is its own piece of work and we would say so.

AI Integration Consulting | Paraguay, Switzerland, Germany

The demo always works. The rollout is the hard part.

Everyone has now seen an AI do something impressive in a five-minute demonstration. Very few companies have one doing anything useful on a Tuesday afternoon.

The gap is not the model. It is that a demo gets clean input and no consequences, and your business has neither.

Useful AI needs three things a demo skips: your actual data reaching it, somebody checking what comes back, and a defined answer to the question of what happens when it is wrong.

That is most of the work, and it is the part we do.

What we actually do

In the order it tends to happen.

Find the work worth taking

We follow what your people actually do all day and look for the repetitive, high-volume, low-judgement parts. Every business has different ones, and they are rarely the ones the software vendor is selling.

Build it where the work happens

Inside the systems your team already has open, not in one more tab they have to remember. A tool that needs a habit change usually gets one week of enthusiasm and then nothing.

Orchestrate, don’t improvise

Longer jobs get several narrow agents with something coordinating and checking them, rather than one general agent asked to do everything and hoped for. More on that below.

Train the people who own it

Including what it is bad at. A team that knows where the system is unreliable will catch the mistakes; a team that was told it is magic will forward them to a customer.

Three of the four steps are worth automating. The fourth is the one your customer will remember, so it keeps a person.

Agent orchestration

Several small agents beat one clever one.

Ask one general agent to handle a whole job and you get something that works four times out of five and fails in a different way each time. There is no point in the process where you can check it, because it did everything at once.

Split the job and that changes. One agent reads the documents, one looks up the numbers, one drafts the answer.

Each step is small enough to verify, so a wrong answer gets caught where it happened rather than three steps later. When something does go wrong you can see which agent did it.

It is the same reason you would not give one person the whole of accounts payable with no second signature.

The point is not that three agents are cleverer than one. It is that you can check each of them.

How it works

Three steps. You can stop after any of them.

Assessment

One to two weeks, fixed price. We look at how you actually work and come back with the three or four places AI would earn its keep, what each would cost, and which ones we would skip. The list of what not to do is usually the more useful half.

Pilot

Four to eight weeks on one process, with real users and real data. It runs alongside the manual way until the numbers say it is better. If they never do, you have learned that for the price of a pilot rather than a programme.

Rollout and handover

We extend it to the next process, wire up monitoring, and train your team on what it does badly as well as what it does well. You own the prompts, the code and the data.

Who this is for

This is for you if

You run a company of roughly 10–200 people and you make the decisions.
You keep hearing “AI” and cannot tell which part of it is real.
There is a job in your company somebody does the same way fifty times a week.
You would rather be told what not to automate than sold a platform.

Probably not, if

You want a chatbot on your website.
The goal is a headcount reduction you have already decided on.

Questions we get asked

Which AI model do you use?: Whichever one fits the job and the rules you operate under, and we expect to change it. Models are the fastest-moving part of this and the easiest to swap. What takes the work is everything around the model: getting your data to it, checking what comes back, and deciding what happens when it is wrong.
Will our data be used to train someone else’s model?: Not under the arrangements we set up. Business tiers of the major providers exclude your content from training by contract, and where that is not enough we run a model on infrastructure you control. Which of the two applies is a question we settle before anything is built, not after.
Is this going to replace people?: In the projects we take on, it takes tasks off people rather than removing the people. The work AI is good at is the copying, the looking-up and the first draft. Judgement, exceptions and anything a customer will remember stay with a person, and the system is built so that person can see and correct what it did.
What is agent orchestration?: Using several narrow agents for one job instead of asking one general agent to do everything, with something coordinating them and checking the result. One reads the documents, one looks up the numbers, one drafts the answer. Each does a small job that can be verified, which is what makes the whole thing trustworthy enough to leave running.
What if the AI gets it wrong?: It will, so the design question is what happens then. Every workflow we build has a defined failure path: low confidence goes to a person, nothing irreversible happens without a confirmation, and every run is logged so you can see what it did and why. A system that cannot be wrong safely is not finished.
We do not have our information in order. Can we still start?: Yes, and almost nobody does. It does shape where we start: with messy information the first win is usually a narrow one, on the one process where the data is already good enough. If the state of the information is the real blocker, that is its own piece of work and we would say so.