Knowledge lives in documents
Contracts, procedures and instructions are scattered, and the answer depends on who remembers where to look.
Private AI for businesses
Private AI that knows company documents and leaves decisions to people.
I implement AI assistants for documents, complaints, service and customer support. The solution can run locally, in a European cloud or through a secure API, with access control, cited sources and human approval.
Service and B2B companies · Kraków and Poland. The first step is one process, not buying a server.
Cited answersA source to check, not just plausible text.
Measured choiceA local model is not the goal by itself.
No blind decisionsCritical actions require human approval.
Where companies get stuck
Documents, questions and data in one controlled process.
The value comes from connecting company knowledge to real work. A chatbot without sources, permissions and measurement soon becomes another place to check by hand.
The same questions return
The team spends time reconstructing answers instead of handling work that needs experience.
Data needs care
Permissions, retention, logs and rules of use matter more than the model name.
The choice is unclear
Local model, European cloud or ready API? It depends on data, quality and total task cost.
Concrete use cases
One process that takes people's time today.
A good first project has repeatable input, an existing source of truth and a person who can evaluate the result.
Complaints and warranties
AI collects case data, checks completeness, finds the procedure and prepares a response for approval.
AI document assistant
An employee asks about instructions, contracts, pricing or policies and gets an answer with its source.
Service and helpdesk
The system classifies the problem, proposes diagnostic steps and routes a difficult case to the right person.
B2B sales
The assistant prepares a reply from the offer, customer history and current commercial terms.
Employee onboarding
A new team member finds procedures faster and knows when a manager must decide.
Documents and administration
The model extracts agreed fields, structures the case and starts a deterministic process step.
Working proof of concept
A complaint moves through documents, rules and human approval.
The PoC runs locally on a Mac and uses fictional cases with photos. The model prepares a summary and draft, while a separate rules engine checks the order number, required evidence, a PLN 200 limit and the user's role.
This is a process demonstrator, not a client deployment. A pilot uses an agreed sample of real cases and measures quality, time and cost.
- 01Case with photos↓
- 02RAG procedure retrieval↓
- 03Analysis and reply draft↓
- 04Rules and permission check↓
- 05Human approval↓
- 06Local audit log
Four ways to work together
From a business question to a solution you can maintain.
You can order the audit, the pilot or the full integration. Scope and cost follow the process review.
01 / I do not know where to start
Data and process audit
I review one concrete process, its knowledge sources, systems and constraints. I compare changing the process, buying a tool, using an API and running a model privately.
- Process and risk map
- Pilot scope and measurement plan
- Technology recommendation
The result may be a decision not to build anything.
Ask about this service02 / I want to test one case
Knowledge assistant PoC
I build a small, measurable prototype over legally shared documents. Answers show their sources, respect permissions and hand uncertain cases to a person.
- Working pilot
- Test set from real questions
- Decision on the next direction
A pilot does not pretend to be a finished product or an automatic decision.
Ask about this service03 / The pilot works, now connect it
Integration with company work
I connect the assistant to email, storage, CRM, SharePoint or another system. I add roles, operation logs, error handling and approval for consequential actions.
- Observable, testable process
- Permissions and audit trail
- Handover documentation
Provider dependencies and limits are clear before work starts.
Ask about this service04 / I have a system to maintain
Evaluation and care
I monitor answer quality, availability, cost and changes in the knowledge base. I update rules and models when measurement shows a real need.
- Quality and cost report
- Regression tests on real questions
- Next-improvement plan
Care is a separate service, not a hidden subscription.
Ask about this serviceArchitecture choice
A local model is not always the safest or cheapest option.
I compare quality, data classes, volume, required latency and total operating cost. For many small businesses, a hybrid design works best: documents, permissions and rules stay local while only minimal, redacted context reaches the model.
Ask for an architecture comparisonDatalocal or in a European cloud
RAGsources and citations in answers
Testsquestions, errors, cost and latency
What private means
Privacy is architecture, not a sticker.
Data: Local or in a European cloud, depending on requirements and cost.
Sources: Answers point to the document and passage behind them.
Access: Users see only what they are allowed to see.
Human: Critical actions and uncertain answers go to approval.
Measurement: We test real questions, errors, latency and task cost.
Handover: Code, configuration, logs and instructions stay with you within the agreed scope.
First project
A small pilot, clear criteria and a decision after the result.
Choose one process
We do not start with a model. We start with a question that costs time or causes errors today.
Take a data sample
Using legally shared material, we define what the system can see and what a good result looks like.
Compare solutions
We test a local model, a European cloud and ready APIs. Local does not always mean cheaper.
Make the decision
You get the pilot result, tests and the cost of keeping it running. We can continue, change direction or stop.
Questions before the first implementation
Private AI without shortcuts.
- Does private AI have to run on our server?
- No. It can run locally, in a private cloud or through a business API with suitable retention terms. The choice depends on data, quality, volume and operating cost.
- Is data sent to AI used for model training?
- It depends on the provider, product, agreement and configuration. The audit checks the exact data flow, retention, processing region and features that persist state.
- What is RAG?
- RAG retrieves relevant passages from company documents before the answer is prepared. It makes sources visible and can limit retrieval to knowledge the user is allowed to access.
- Can an AI model run on a Mac?
- Yes. A small or medium model can run on a modern Mac and suit a demonstration or a single user. Multi-user production needs separate performance, reliability and cost measurement.
- How long does a first pilot take?
- A well-scoped pilot for one process can usually be prepared in a few weeks. Timing depends mainly on source availability, integrations and acceptance criteria.
- Can AI approve complaints automatically?
- The model can prepare an analysis and draft. Approval should follow explicit rules and permissions, while financially consequential actions should include an agreed human approval step.
Start with the problem
Do you have documents your team keeps searching for?
Describe where time disappears today. To start, you only need the process, never passwords, API keys or customer data.
- ✓A conversation about one process
- ✓Audit, PoC or integration as a separate service
- ✓Scope, cost and acceptance criteria before work starts