You
Compare both offers: price, delivery date and included services.
Private AI for business
Your AI.Privately run.
§ 4 Warranty. Spare parts: 24 months from delivery. Wear parts: 12 months.
Chapter 9, downtime: note the error code, call the service line. Callback within 4 hours.
Recurring feeder downtime since commissioning. We request a prompt service visit.
Series 4, base module: €48,900 net. Feeder: €6,250. Lead time Series 4: 6 weeks from order.
114-021 conveyor belt 1,200 mm, stock 2. 114-034 inductive sensor, stock 8. 114-051 drive belt, stock 0.
Series 4 line with feeder and PLC control: €64,950 net. Lead time six weeks from order.
Belt 114-051 replaced, sensor calibrated. Follow-up check agreed in four weeks.
3.2 Approvals. Orders above €9,000 net: approval by the head of purchasing. Above €50,000, management approval as well.
- Complaint 03/26
- Price list 2026
- Quote 4711
- Framework agreement 2023
- Maintenance manual S4
- AA-114 spare parts
- Service report
- Purchasing policy
- Salaries 2026.xlsx
- Employment contracts
§ 4 Warranty. Spare parts: 24 months from delivery. Wear parts: 12 months.
3.2 Approvals. Orders above €9,000 net: approval by the head of purchasing. Above €50,000, management approval as well.
Series 4, base module: €48,900 net. Feeder: €6,250. Lead time Series 4: 6 weeks from order.
Chapter 9, downtime: note the error code, call the service line. Callback within 4 hours.
Training
Your model · trainedKnows your business.
The feeder has stopped. What now?
Note the error code and call the service line. Callback within 4 hours.Maintenance manual S4 · ch. 9
Your environment · privately runChatGPT
Claude
Stay outNothing leaves your environment.
Collect documents.Folders, mail, archives. Only what you approve.
Extract knowledge.Facts, rules and figures from the pages.
Train the model.Your model learns with your knowledge. Not with a provider’s.
Run it privately.In your environment. External AI services stay out.
AI does not have to run at a model provider. We set up suitable models on your own infrastructure or in a dedicated environment operated for you. Your documents can then be processed where you define data flows and access.
Self-hosting means the AI model itself runs in your controlled environment — not just the interface.
01What changes
More control.More possibilities.
If confidential data has been holding back your use of AI, consider how it is operated. That determines who runs the model and which conditions you can shape yourself.
Your scopeExternal AI serviceExternal servicePrivately operated AIPrivately run
Processing
At the model provider
In your environment
Data & access
On the provider’s terms
On your terms
Model & adaptation
Within the provider’s offering
Selectable and adaptable
Usage
Plan and allowances
Your own compute
- GDPR-compliantData flows documented
- No data sharingNot with external AI services
- Role-based accessApplies to answers too
Control needs dependable operations.A private environment needs to be configured, secured and maintained. Your IT team can handle that, or Cognatio can provide the agreed services. Self-hosting alone does not satisfy privacy requirements.
02Your company knowledge
The right answer startsin your documents.
Handbooks, policies and offers: we make approved content searchable. AI receives relevant passages for a question and provides an answer with traceable sources.
Your private AI environment
Prepare documents
Offers01
Handbook02
Policy03
Prepare · assign permissions
KNOWLEDGE BASEKnowledgeSources & access
When someone asks
Who else needs to approve offer B?
From your knowledge base: two relevant passages
The answer follows from both
Offer B needs additional approval from the head of purchasing.
Preparation, retrieval and answer generation are configured for the private environment. Access permissions also apply to the retrieved content.
03Adapted to your task
A smaller model.A specific speciality.
Route requests, recognise details and prepare next steps: a smaller model can be adapted to a well-defined task, using your examples and expert corrections.
What does model training mean here?
Looking up knowledge and training behaviour are different things. We test which adaptation your task actually needs.
Fine-tuning uses suitable examples to adapt a model to tasks and response behaviour. Reinforcement learning (RL) can also help when results can be clearly evaluated. Testing determines whether training is worthwhile and which method fits, not model size alone.
Turn a message into a clear request.
Adapt using your examples
Approved example
“Please change the delivery address.”
Expert correction
General supportOrder service
New request · not the training example
Your specialist modelAdapted to your routing rules
How AI should prepare the request
- Request
- Change delivery address
- Responsible team
- Order service
- Next step
- Find the order and review the change
The team reviews and approves.
04Familiar work
Two offers.One informed follow-up.
Attach two PDFs, compare offers and spot an uncertainty. Your team works in a private chat interface and checks the result themselves.
Private AI
Two offers. One decision.
Your documents
AI model
Attached files can cover individual tasks. Regular access to company knowledge runs through the searchable knowledge base.
05The longer view
Your usage shapesthe economics.
The actual use case determines whether private AI makes economic sense. Three things belong together:
01Usage
How regularly?
Recurring work can make better use of provisioned compute capacity.
02Task
How demanding?
Required quality and response time determine the suitable model and hardware.
03Operations
What is included?
Setup, ongoing infrastructure and dependable support belong in the same calculation.
Which costs are compared?
Two cost structures.
Same task volume · same quality requirement · same periodExternal AI service
- One-off
- Setup and any integration
- Ongoing
- Suitable licences or API usage, plus integration and support
Privately operated AI
- One-off
- Setup, hardware and training where needed
- Ongoing
- Compute, storage, energy or hosting and support
Locally processed requests do not use an external model service allowance. The limits of your compute capacity, model context and software licences still apply.
06Cognatio sets it up
Private AI.For your team.
A usable interface, suitable access and a model for your task. With your knowledge base where needed. We set up the solution and agree who will run it.
Discuss private AIA description is enough to start. Please do not send confidential documents through the contact form.
Another question?
What runs on our infrastructure with this solution?
Company chat and AI model run together where we agree the operating location: on your infrastructure or in an environment operated for you. The chat is not merely a front end for an external model API.
Do we need new hardware?
That depends on team size, usage, the selected model, and your existing infrastructure. The computer illustration on this page explains the principle; new hardware is not automatically part of the proposal.
Is a self-hosted model as good as the well-known market leaders?
Not as a blanket statement. Quality, speed, and suitable tasks differ by model and hardware. We select and test the approach for your actual work instead of promising equivalence.
Does self-hosting automatically make the setup privacy-compliant?
No. Self-hosting gives you more operational control, but it does not replace an assessment of purpose, data, access, retention, and other legal requirements.
Do we own the AI model after setup?
Not automatically. Model usage rights and licences depend on the selected model. We state clearly what you may use and which installation or documentation deliverables are handed over.