services · 03
AI integrations that actually cut costs.
Claude, GPT or open-source models plugged in where the team loses the most hours. Automations, agents, RAG, voice — with results visible in P&L, not in a LinkedIn post.
What you get
Specifics, not promises.
Every project is a measurable business outcome. No "social media just because" or "sites that just look fine."
/01
Audit before any rollout
First 5-7 days are a process audit. How many hours per week does the team lose to manual work? Which of those tasks can be safely handed to AI? What pays back in 90 days? You get a map, not a promise.
/02
Numbers, not a demo
Every rollout has a quantitative goal: -30% ticket time, -50% content cost, +20% NPS. 30 days after launch you get a report — what actually shipped.
/03
Security and data privacy
PII flagging, log redaction, EU residency for Claude and GPT, prompt-injection guardrails, per-user limits. Built to AI Act and GDPR, not to a marketing one-pager.
/04
Stack picked for your needs
I'm not married to one provider. Claude Sonnet and Haiku, GPT-4, on-prem Llama, Whisper, ElevenLabs. Choice based on cost-quality analysis on your data, not on what LinkedIn is hyping today.
Process
From brief to launch.
- 01
Audit and ROI map
Five to seven days: team interviews, list of 5-10 candidate use cases, ranked by return and risk. You walk out with a document you can act on, even if you don't roll out further with me.
wk 1 - 02
Prototype on one use case
We take what pays back fastest and I build a working prototype. If it doesn't work on your data — we kill it without production rollout costs.
wk 2-3 - 03
Production and tracking
Integration with your system: CRM, Slack, panel, app. Logging, evals, alerts when quality starts dropping. No 'worked on demo, prod is unknown'.
wk 3-5 - 04
Iteration and scaling
Monthly: metric review, prompt tuning, adding more use cases (if ROI justifies). We stop if it stops returning — no project for project's sake.
retainer
Pricing
Clear ranges, no "it depends".
Final quote always after a 60-minute brief. Regardless of plan, I invoice in stages, never 100% upfront.
01
Audit + PoC
3-8k PLN
2-3 wks
- Audit of 5-10 processes
- ROI map
- 1 use case PoC
- Stack recommendations
- No production commitment
02
Rollout
8-25k PLN
3-6 wks
- Everything in Audit
- Full integration
- Logging + evals
- KPI tracking
- Team onboarding
- 30 days of stabilisation
03
AI Retainer
4-12k/mo
monthly
- 10-30 hrs of work
- Existing rollout iteration
- New use cases
- Quality monitoring
- API cost optimisation
Related work
Here's what it looks like with clients.
FAQ
The questions I get most.
My team is afraid AI will replace them. What about that?
Communication from day one: AI takes the manual work, people keep the things that need judgment and context. I show this concretely on the data — after rollout the team has more 'human-worth' tasks, not fewer. That's why I work mostly through audits, not 'let me build AI'.
Will our data leak to OpenAI or Anthropic?
No, if configured correctly. Claude and GPT in API/enterprise mode don't use your data to train the model — that's in the terms of service. For the most sensitive cases we deploy Llama on-prem or Mistral in EU. Risk-audit decision, not gut feel.
How much does it actually cost monthly after rollout?
API costs alone for an SMB are usually 100-2000 PLN/month. Haiku and GPT-4o-mini are cheap enough that ROI usually breaks in the first quarter. At larger scale (10k+ requests per day) we calculate ROI to the złoty, not approximately.
Do you have experience with RAG, agents and MCP?
Yes. Formified.ai (my own product) has the entire AI Coach built on RAG , vector search in Supabase pgvector. I have ready patterns for Claude Tool Use agents, MCP servers, function calling and structured output (JSON Schema). Not learning this for the first time on your project.
What about hallucinations?
Three layers. One: RAG with source citations so the model has something to anchor to. Two: structured output and schema validation so the format does not drift. Three: human-in-the-loop for sensitive decisions. A production hallucination is a bug — measured and fixed like any other.
Next step
60 min brief. Free.
Hear the 5 questions I ask on the first call , a lot becomes obvious right there, before any project starts.