— FOR IT

AI for the IT team, six architecture decisions.

For CTOs, IT directors and AI engineers. What to choose, what to avoid, where to invest, where to cut. Vendor-agnostic, on-prem-friendly.

— DEFINITION

For IT is the yesfor.ai track for technical teams. We focus on architecture decisions (vendor, RAG vs fine-tuning, on-prem vs cloud), cost management, observability and security. We support internal teams, we do not replace them.

— HEADS UP

Half your team already uses AI. With no company policy.

Research from 2025 shows that most employees in companies regularly use AI tools (ChatGPT, Claude, Copilot) with no company awareness and no policy. This is known as “Shadow AI”.

  • ·Pasting customer data into free chatbots (a GDPR breach)
  • ·Copying fragments of contracts and terms in for a “rewrite”
  • ·Sending source code in for analysis (IP leakage)
  • ·Sharing employee data (salaries, performance reviews)
78%

of employees bring their own AI tools to work (BYOAI)

Microsoft Work Trend Index, 2024

over half

feed company data into public AI tools

Global security research, 2024-2025

2-4 wks

to roll out an AI policy in a company

yesfor.ai

— SOLUTION

A company AI policy, a Shadow AI audit, a team workshop. Cybersecurity & AI Policy, 4,500 to 17,500 PLN.

See the service →

Six decisions that define an AI project.

  1. 01

    Choosing an AI model with no commitment to a vendor

    OpenAI, Anthropic, Google, local models (Mistral, Llama, Gemma 3). For sensitive data we work with LLMs running inside the client's own infrastructure, so the data never leaves the company. Each model has its own strengths: cost, response time, quality, regulatory fit. The choice depends on the specific use case.

  2. 02

    RAG vs fine-tuning architecture

    RAG (Retrieval Augmented Generation) for most enterprise use cases. Fine-tuning is rarely worth it for LLMs. Vector DB, chunking strategy and reranking are 80% of the success.

  3. 03

    On-prem vs cloud

    Cloud (OpenAI, Anthropic, Bedrock) for most. On-prem (Llama, Mistral) for finance, healthcare and legal, where data cannot leave the client's infrastructure.

  4. 04

    Cost management

    LLM costs explode at scale. Caching, batching, model routing (a cheaper LLM for simple tasks, a more expensive one for complex tasks), monitoring per use case, hard budgets per team.

  5. 05

    Observability and evaluation

    Without observability an LLM drifts silently. Prompt logs, quality metrics, A/B model tests, regression tests. Without them the first incident ends the project politically.

  6. 06

    Security and prompt injection

    Prompt injection is the top OWASP risk for LLM apps (2025). Input sanitization, output filtering, sandboxing, rate limiting, audit logs. A traditional security stack does not cover LLMs.

„Most AI projects do not fail on the model, they fail on the integrations with master data.”

yesfor.ai Manifesto

Tech Principle

Four categories of integration.

— Master Data

  • · ERP (SAP, IFS, Comarch)
  • · CRM (Salesforce, HubSpot, custom)
  • · Data warehouses

— Document stores

  • · SharePoint, Confluence
  • · Google Drive, Dropbox
  • · Custom DMS

— Comms

  • · Slack, MS Teams
  • · Email (Outlook, Gmail)
  • · Ticketing (Jira, Linear)

— Custom

  • · Internal APIs
  • · Data lakes (S3, ADLS)
  • · Event streams (Kafka)

— FOR IT TEAMS

Monthly support or a turnkey rollout.

IT support →Turnkey rollout