Private AI pipeline: how SMBs can process data securely

June 1, 2026 · Bartłomiej Spasowski, Founder, modulla · EN

What is AI Brain and why are small businesses turning to it?

AI Brain is a private AI infrastructure that combines local language models, workflow orchestration tools, and a company knowledge base into a single integrated system. Data stays inside the organization, operational costs decrease, and repetitive tasks stop consuming leaders' time. The solution scales for companies with 5 to 20 people.

Experience shows that the first association with your own AI infrastructure is server room corridors and a dedicated IT department. Meanwhile, the economics of this approach have changed radically in the past two years, and the entry threshold has dropped to a level accessible to small operational businesses.

The problem: SaaS fragmentation and Shadow AI are eating your company's time and data

A typical 10-person company pays between 2,000 and 5,000 PLN per month for subscriptions to various AI tools. ChatGPT Teams costs around $30 per person per month, nearly 1,200 PLN for 10 people. On top of that comes a separate CRM, automation tools, and possible assistants for individual employees. Each of these tools operates in its own silo. Data from one doesn't reach another without manual copying. Automations break down.

And worse, employees solve this problem themselves through Shadow AI: they paste client proposals, source code, HR data, and contracts into public AI models to speed up their work. Poland's UODO surveyed companies and found that nearly 60% of them don't connect AI use with personal data processing at all. This isn't the employees' fault, it's a structural problem: they have no safe alternative.

AI Brain addresses this problem not as yet another tool alongside the rest, but as a shared backbone for the entire company's operational process.

Private AI infrastructure architecture: three system layers

AI infrastructure for SMB companies rests on a three-layer architecture. Each layer can be deployed gradually, without a revolution in the existing technology stack.

Layer 1: Orchestration and integrations

The first layer is the nerve center of the entire system. The workflow orchestration platform connects CRM, ERP, email, and external APIs into a single coherent data flow. It handles queues, retries, and routing between agents. This is where you define which trigger fires which process and where in the flow a human decision is required.

Layer 2: Company knowledge and RAG

The second layer is where AI starts to understand a specific company, because it receives its own knowledge base. RAG (Retrieval-Augmented Generation) is the mechanism that makes a language model "know" about a given organization. Tools like Dify allow you to build editable knowledge bases connected to language models. AI doesn't generate responses from general memory but retrieves current data from company documents, contract databases, HR policies, or commercial offers. A vector database built on PostgreSQL with the pgvector extension delivers high accuracy at large query volumes and eliminates the need to manage a separate specialized database.

Layer 3: Local model and inference

This is where Ollama comes in: the ability to run open-source language models directly on company hardware, with an API compatible with OpenAI. Data never leaves the infrastructure. For companies that need the power of larger cloud models while maintaining GDPR compliance, a two-layer architecture is used: a local model anonymizes sensitive data, and only the cleaned data is sent to the cloud.

AI Brain vs. traditional SaaS stack: a comparison

Area	Traditional SaaS	AI Brain (private infrastructure)
Data security	Data in the vendor's cloud	100% sovereignty, data stays inside the company
Monthly cost (10 people)	2,000–5,000 PLN/month	1,500–2,000 PLN/month (TCO: hardware + energy + specialist + API)
Integration between tools	Manual or via external middleware	Native data flow orchestration
Customization to the company	Limited by the tool's features	Full control over process logic
Shadow AI	No control	Safe alternative eliminates the phenomenon
Investment break-even	Fixed cost with no return	Return on investment in 6–12 months
GDPR and EU AI Act compliance	Depends on vendor's policy	Full control over data flow

How to deploy private AI infrastructure without burning out your team?

The biggest mistake when implementing AI is trying to automate everything at once. The outcome is predictable: technical chaos, team resistance, and a project shut down after three months. Companies that achieve lasting success with AI spread deployment across phases and validate each phase with measurable results before moving on.

Phase 1: Diagnosis and process mapping

A well-planned deployment starts with mapping existing processes. Which tasks consume the most time? Where does data flow break down between tools? Where are employees reaching for unauthorized AI tools? This is the phase for identifying real bottlenecks, not hypothetical use cases. The output is a list of priority processes and a map of sensitive data flows within the company.

Phase 2: Process architecture design

Based on the diagnosis, the target architecture is designed. A technology stack matched to the company's profile is selected, integrations with existing tools are planned, and human-in-the-loop checkpoints required by the EU AI Act are defined. This is the phase for key decisions: what runs locally, what can go to the cloud, and under what conditions.

Phase 3: Build in steps, not in a leap

Workflows are built incrementally. In the first two months, low-risk automations are launched: text editing, meeting scheduling, first RAG agents for internal knowledge. In months 3–4, customer support and CRM automation are deployed. In months 5–6, the full environment is integrated with ERP and operational monitoring. Each phase delivers measurable time savings before moving to the next.

Phase 4: Scaling and optimization

Once the infrastructure has stabilized, it can be scaled. Additional modules are added, the number of agents is expanded, and cost per transaction is monitored. And this is where the core benefit emerges: leaders start reclaiming hours that previously went to operations instead of strategy.

Practical AI Brain applications in a 5–20 person company

Private AI infrastructure is not an abstraction. Below are concrete examples from deployments that are working today.

Recruitment and HR

A 5-person recruitment agency deployed the local Qwen3 8B model through a workflow orchestration platform, integrating it directly with the CRM. The agent automatically parses meeting notes, formats CVs to a consistent standard, and tags data while complying with GDPR requirements. The team redirected time previously spent on administration toward building relationships with candidates and clients.

Customer service and bookings

An agent operating via Telegram handles customer inquiries around the clock: it answers product questions using the company's private knowledge base and independently schedules appointments in Google Sheets or CRM. The operational cost comes to around 0.60 PLN per full consultation and booking. With several dozen inquiries per day, that's a marginal amount compared to the cost of an employee doing the same tasks.

Finance and document extraction

A small accounting firm deployed AI for automatic data extraction from invoices and receipts. Client documentation processing time dropped from hours to minutes. The firm now handles 50% more clients without growing its team.

Sales and CRM orchestration

AI agents pull signals from ERP (inventory levels, project stage, payment statuses), structure notes from sales meetings, and automatically create follow-up reminders in the CRM. One B2B deployment cut the time to prepare post-sale notes by over 75% and eliminated manual updating of sales opportunities in the system.

Content and marketing

Agents autonomously scan industry trends, generate content briefs, write article drafts, and create tasks in project tools with full context. The entire flow from topic to finished draft closed in Asana: 40 to 90 seconds. This kind of content pipeline, built on the second brain concept, integrates these processes into a single flow with a quality gate, rather than connecting several separate tools.

AI Brain economics: when does the investment pay off?

According to Gartner analysis, more than two-thirds of companies report measurable benefits from generative AI deployments within a year of launch. Concrete numbers matter more here than industry statistics.

Sample TCO for a 10-person company on a Mac Mini M2: a one-time hardware purchase of around 5,000 PLN, amortized over 3 years, comes to about 140 PLN/month. Add energy and cooling (around 50 PLN/month), the market cost of a specialist at around 5 hours of maintenance per month, and any external API costs. At market rates for AI/IT specialists, total TCO realistically amounts to 1,500–2,000 PLN per month. That's still less than a standard SaaS stack for 10 people, and break-even compared to previous costs is reached in 6 to 12 months.

After that, the infrastructure generates savings, not costs. Each new process increases efficiency without a proportional increase in spending.

For companies requiring higher-parameter models: a Mac Studio M4 Max with 128 GB unified memory runs 70B-class models without limitations. A server with a cluster of enterprise cards (e.g. 4x AMD MI100 from the used market) is an option for more technically advanced teams with a deeper hardware budget.

GDPR and EU AI Act: three principles built in from day one

The EU AI Act is not another regulation to read and set aside. Fines for improper AI deployments reach 35 million EUR or 7% of global turnover, and regulators have adopted a zero-tolerance approach.

High-risk AI: AI systems used in HR, recruitment, and employee assessment are classified as high risk. Full automation of decisions in these areas is prohibited. Human-in-the-loop is a legal requirement, not an architectural option.
Event logs: events from automated AI systems must be logged for a minimum of 6 months for regulatory audit purposes.
GDPR: the principles of data minimization, purpose limitation, and retention must be built into the process architecture from the very start, not bolted on after deployment.

Private AI infrastructure gives you full control over data flow. That is a fundamental advantage over SaaS tools, where data is processed through the vendor's infrastructure according to the vendor's policy, not the client company's policy.

Most common mistakes when deploying private AI infrastructure

Automating everything at once: jumping to full integration without a pilot phase ends in technical chaos and team resistance.
Skipping process mapping before deployment: AI layered on top of a broken process replicates its problems faster and at greater scale.
Neglecting agent security: agents with access to CRM, email, and the file system represent an attack surface. They require isolated environments (Docker with restricted access) and strict permission controls.
Ignoring Shadow AI: the absence of a safe alternative amounts to tacit consent to corporate data leakage through public AI models.

The conclusion is clear: private AI infrastructure is no longer the domain of corporations. Companies with 5 to 20 people that have repetitive operational processes and a tool fragmentation problem can today build their own AI Brain at a TCO many times lower than their current SaaS stack. A process audit is always a good starting point for seeing which areas will deliver the greatest return in the shortest time.

Schedule a free audit

FAQ: AI Brain for SMB companies

Does deploying AI Brain require your own IT department?

No. The stack based on Dify and Ollama is designed so that it can be deployed and managed without a dedicated IT department. The first automation workflows can be launched within two months of the audit. A well-executed deployment leaves operational documentation that any competent project coordinator can handle.

Are local AI models good enough for business tasks?

For most operational tasks in a 5–20 person company, 8B to 70B parameter models running locally (Qwen, Llama, Phi) deliver fully sufficient results. For tasks requiring deeper reasoning, a two-layer architecture is used: the local model anonymizes data, and only cleaned information is sent to the cloud. This is a compromise between computational power and data sovereignty.

How does AI Brain fit the requirements of GDPR and the EU AI Act?

Private infrastructure is the foundation of GDPR compliance: data does not leave the company, and you have full control over retention and processing. The EU AI Act additionally requires human-in-the-loop for high-risk systems (HR, recruitment) and event logging for 6 months. These requirements should be built into the architecture from the very first day of deployment, not added on as an afterthought.

When does the investment in AI Brain pay off?

For a 10-person company, the realistic TCO after deployment is 1,500–2,000 PLN per month (hardware amortization, energy, the market cost of a specialist maintaining the system, any external APIs). Compared to previous SaaS costs, break-even is reached in 6 to 12 months. After that, the infrastructure generates savings rather than costs, and each new process increases efficiency without a proportional rise in spending.