AI Implementation
AI Implementation: adopting AI in your company, with local models if you need them.
From the question of which use case actually carries its weight to the running system. On your own infrastructure where required, so that your data never leaves the building.
Scope
What does AI implementation cover?
We introduce AI where it measurably improves a concrete process, and leave it out where it only creates work. The scope follows your use case:
- Choosing the use case. The uncomfortable question first: which process economically supports an AI solution, and which one is better served by a rule or a script? We sort your ideas by impact and effort before anything gets built.
- The model decision: hosted or local. Hosted models are quicker to start with; local models keep your data in the building and carry no cost per request. We work through both routes for your case instead of offering a matter-of-principle answer.
- Local operation. Setting up open models on your own hardware or your own server, through the established runtimes Ollama, LM Studio, llama.cpp, and vLLM. We operate exactly this connection in our own extension SEOMURAI, which is published in the Chrome Web Store.
- Connecting your own documents. So the AI answers from your manuals, quotes, and archives rather than from the internet. Including the question of which content belongs in there at all and who is allowed to see which answers.
- Automating processes. Classifying, extracting, pre-structuring, preparing reports. Connected to the tools you already use, not as another island beside them.
- The legal frame from the start. Data protection impact assessment, processing agreements, and AI literacy under Article 4 of the EU AI Act are considered up front, not filed afterwards. Covered in depth by the EU AI Act compliance audit.
- Handover and walkthrough. Your team receives the system explained, documented, and in their own hands. If you intend to run the topic yourself afterwards, that is precisely the goal.
Fit
When does an AI implementation fit?
The implementation fits when:
- You have a concrete, recurring process that costs time today. One use case with a clear edge beats an AI strategy without an object.
- Your data must not leave the building, because it concerns personal data, client files, engineering data, or contracts. That is the case local models were built for.
- You already have AI tools in use, but nobody can say which data flows where and who needs to be trained.
- You intend to run the topic yourself in the long term and need someone at the start who accompanies the first steps and hands the knowledge over.
What you receive
What you receive. The same in every project.
Requirements conversation
We clarify the use case, your data situation, and the constraints before a single line is written. If it turns out that AI is the wrong route, we say so.
A basis for the decision
Hosted or local, which model, which hardware, which costs per route. With figures for your case, so the decision belongs to you and not to a vendor.
The running system
Set up, connected, and proven in operation. Not a demonstration build that belongs to nobody after sign-off.
Documentation and walkthrough
Your team gets the system explained and documented, including how to tell when something has stopped working. Followed by 30 days of questions.
Scope
What determines the scope.
The shape follows your starting position, not a choice of sizes. Four questions determine the effort. Tell us where you stand and you receive an offer with a fixed scope and clear terms.
How many processes you want to adopt
One clearly bounded use case, or several one after another. This is the single biggest lever on the effort involved; we put one into operation before the next begins.
Hosted or on your own hardware
A hosted service is usually ready within a few days and pays off where requests are rare. Running it yourself requires hardware selection, setup, and an operating concept, but keeps the data in the building and costs nothing per request.
Where the answers should come from
A model on its own, or one that answers from your manuals, quotes, and archives. The second brings along the questions of data quality and access rights.
How many people work with it
A handful of specialists or the whole company. Beyond many concurrent users, dedicated server hardware becomes the question; below that, an existing machine is usually enough.
Process and terms
Four steps to a running system.
The path to an offer is the same for every engagement. After the first conversation you know where you stand, and you have the price in writing before anything begins.
Requirements conversation
One session in which we walk through the use case, your data situation, and the constraints. Free of charge and without obligation.
Offer with a fixed scope
You receive in writing what gets built, what it costs, and what acceptance depends on. Priced per project or by effort, depending on how clearly the scope can be defined at that point.
Delivery in bounded steps
We start with one use case and put it into operation before the second begins. That keeps the effort manageable and shows early whether the direction is right.
Handover
Documentation, walkthrough, and 30 days of follow-up questions. After that you decide whether you continue alone or we stay within reach.
Frequently asked
What you should know before requesting.
Legal status: 29 August 2026. These notes are an editorial summary and not legal advice; they do not replace a lawyer’s review of your particular case.
What does an AI implementation cost?
That depends on the use case, which is why no figure appears here. We clarify the scope in a requirements conversation and then present an offer with a fixed scope of work and clear terms. One bounded first use case is considerably cheaper than a programme spanning several departments, and we expressly recommend starting with the first.
Is a local model worth it compared with a hosted service?
It comes down to three quantities: how sensitive the data is, how many requests occur, and how good the result has to be. With sensitive data or high request volume, local operation often pays off quickly, because there is no cost per request. For infrequent requests on uncritical data, a hosted service is usually the faster route. We work through both variants for your case before you commit.
With a local model, does our data really stay in the building?
Yes, that is the entire point: a locally operated model processes your input on your own hardware, there is no transfer to a provider and no use of your content for anyone else’s training. What still needs settling are the access rights inside the building, because a system answering from your archives can also surface content that not everyone should see. That is part of the scope.
What hardware do we need for a local model?
Less than most people expect. For many tasks such as classifying, summarising, and extracting, a machine with sufficient memory and a common graphics card is enough; modern workstations with shared memory manage it as well. Only with very large models or many concurrent users does dedicated server hardware become the subject. What you concretely need follows from the use case, not the other way round.
Do you only advise, or do you use this yourselves?
We use it ourselves. minoka publishes its own browser extensions in the Chrome Web Store, and since version 1.5 SEOMURAI connects open models there through Ollama, LM Studio, llama.cpp, and vLLM, plus a model running directly in the browser. What we introduce at your company, we operate in one form or another ourselves.
Can we continue the topic internally afterwards?
That is the normal case and expressly intended. Handover, documentation, and a walkthrough are part of the scope, and where your team wants to go deeper, the matching trainings are available at the minoka Academy. If you only need someone to review your work occasionally afterwards, the advisory retainer is the fitting arrangement.
What does an AI implementation have to do with the EU AI Act?
As soon as your staff use AI systems, the obligation under Article 4 applies, regardless of whether you developed the AI yourself or bought it in. The Digital Omnibus, Regulation (EU) 2026/1744, has been in force since 27 July 2026 and turned this into a duty of effort: you have to act, but you do not have to guarantee a particular level of competence for any individual. We take that into account during adoption, so you do not put a system into operation and then chase the obligations. How extensive the obligations are in your case is settled by the EU AI Act compliance audit.
Request
Ready for your AI implementation project?
Scope and terms follow from the use case, which is why this page names no price. Describe briefly which process occupies you and how sensitive the data is. You receive an answer and a proposed date within 24 hours.
Also relevant
Also from the workshop
- Custom Tooling: Browser extensions, AI automations, and WordPress tooling built to fit.
If the direction is still open
When the direction is not settled yet, an audit clarifies it at a fixed price:
- GEO Audit: Audit at a fixed price.
- SEO Audit: Audit at a fixed price.
- EU AI Act Compliance: Audit at a fixed price.
- Overview of all minoka services: audits, implementation, training, and advisory retainer at a glance.
- minoka Academy: trainings on SEO, GEO, and AI literacy under Article 4 of the EU AI Act, as online module course or in-house workshop.