// services.ai_deploy Module 06 of 06

AI deployment
in your real processes.

Q: How is AI deployment different from AI training?

These are two different services that complement each other. Training teaches your team to use AI on their own: they get ready-made prompt patterns, rules and an organised way of working with AI day to day. Deployment builds a system that works by itself: an assistant on your knowledge base, an automation connecting your customer database, email and calendar, an automation answering the phone on your helpline, automatic transfer of invoice data into your system. The sensible order is training first (the team understands what AI can and cannot do), then deployment built around their daily work.

Q: What about GDPR and sensitive customer data?

We start with the question of what data the AI should even see, because not every deployment needs the same guardrails. For business and operational data (policies, procedures, correspondence, B2B counterparties): Claude API plus a data processing agreement with you (art. 28 GDPR) plus Anthropic's standard DPA, which includes SCC clauses for the transfer to the US. For data subject to special protection (professional secrecy, medical data and other special-category data under art. 9 GDPR, documents under NDA): we move the AI to Europe - Claude via AWS Frankfurt configured to keep data within the EEA, or the Polish Bielik 11B model on your own server or Hetzner PL. The formal side of the deployment is the data processing agreement, a DPIA where the process requires one, an entry in the record of processing activities and a privacy notice for your customers about the AI's role in handling their case. We do compliance before we choose a model, not after deployment.

AI that answers only from your own documents, automations that run your processes, voice handling, automatic data extraction from invoices and contracts. With hard guardrails, a full log of conversations and cost control (API, RAG, agents, voice agents, AI guardrails). No "AI for show".

haski.io / deploy.run

$ deploy --ai

[ consultation ]45 min · free

[ scope ]tailored

[ guardrails ]as standard

[ stack ]claude · rag · n8n · bielik

Initial consultation

Free

Deployment areas

Project scope

tailored

// problem >> solution

The pains we remove

deploy.schema

2 tables · 6 fields

current_state

◆

pilot_that_wont_scale The demo wowed everyone at the start. Two months in it turned out that the AI makes things up on real documents, costs grow in step with the number of queries, and the team has no idea who owns answer quality.

prod-gap

◆

hallucinations_without_audit The AI gives a customer an answer you cannot later defend - no source, no log of the conversation. Legal and finance say the same thing: in this shape it makes no sense.

trust

◆

lock_in_to_one_model The whole thing rests on one model from one provider. Pricing changes, the model gets retired, the rules get tighter - and suddenly you have no plan B. Moving to another solution costs as much as building from scratch.

risk

fix

target_state

◆

rag_with_grounding Every answer has a source. The AI answers only from your materials, it does not make things up. We check accuracy on real questions - before anything reaches production.

grounded

◆

hard_guardrails_logging Control over what the AI may and may not talk about, with hand-off to a human wherever an answer needs a decision. Every conversation logged in full - a history you can show in case of a dispute or an inspection.

compliance

◆

model_router_with_fallback We match the model to the task: cheaper for simple cases, stronger for analysis, the Polish Bielik on your own server for sensitive data. A second provider as plan B. Alerts on cost and response time - the budget stays under control.

resilient

// pain.delta fk → impact

// scope

What an AI deployment includes

Chatbot on your documents

An assistant on your own documents - policies, procedures, knowledge base, customer history. Built on your preferred AI model. Every answer cites its source, and we check accuracy on real questions.

Agents for processes

An automation connects several systems (customer database, email, calendar, your company's main system) and runs tasks end to end: gathering information about a counterparty, a first draft of a contract, reminders, handing difficult cases to a human. With hard guardrails and a full log of every action.

Voice automation

A voice bot that answers calls, meeting transcription with an automatic summary, extraction of action items into your task system. Native Polish. With the option to route to a human.

Document extraction

Invoices, contracts, scans, PDFs, correspondence - data pulled straight into your system. With validation, a confidence score on each recognised field and a manual-review queue wherever confidence is low or liability risk is high.

Guardrails + full conversation log

Rules for every task - what the AI may do, what it must not, when it hands a case to a human. A full log of every conversation with context, GDPR-compliant storage, export of the logs for review.

Monitoring and costs

A dashboard with cost per use case, response time, accuracy and how often the AI hands a case to a human. A budget with alerts. Tests before every release - a drop in quality will not slip by unnoticed.

// stack

[ answer generation ]

Claude API

[ search in your documents ]

ChromaDB + Voyage

[ orchestration ]

n8n + Apps Script

[ on-prem · PL ]

Bielik 11B on-prem

// faq

Frequently asked questions

How is AI deployment different from AI training?

These are two different services that complement each other:

◆Training teaches your team to use AI on their own - they get ready-made prompt patterns, rules and an organised way of working with AI day to day.
◆Deployment builds a system that works by itself - an assistant on your knowledge base, an automation connecting your customer database, email and calendar, an automation answering the phone on your helpline, automatic transfer of invoice data into your system.
◆The sensible order - training first (the team understands what AI can and cannot do), then deployment built around their daily work.

What about hallucinations and how do you prevent them?

AI can state something untrue. It is a known property of every model and nobody sensible will promise it disappears entirely. Four layers of guardrails:

◆AI cites the source - under every answer you see which policy or procedure it came from. If something is not in your materials, AI says 'I don't know' instead of making it up.
◆Uncertainty threshold - when AI is not sure of an answer, it hands the case to your team. We set the threshold together.
◆Decisions always stay with a human - AI does not issue an invoice, send a quote or reply to a complaint without verification. It prepares a draft, you approve it.
◆A full conversation log - if a customer complains or someone asks for details, you can see what the AI said, on what basis and when.

What about GDPR and sensitive customer data?

We start with the question of what data the AI should even see - because not every deployment needs the same guardrails.

◆Business and operational data (policies, procedures, correspondence, B2B counterparties) - Claude API plus a data processing agreement with you (art. 28 GDPR) plus Anthropic's standard DPA with SCC clauses for the transfer to the US.
◆Data subject to special protection (professional secrecy, medical data and other special-category data under art. 9 GDPR, documents under NDA) - we move the AI to Europe: Claude via AWS Frankfurt (configured to keep data within the EEA) or the Polish Bielik 11B model on your own server or Hetzner PL.
◆The formal side - a data processing agreement, a DPIA where the process requires one, an entry in the record of processing activities, a privacy notice for your customers about the AI's role in handling their case.

We do compliance before we choose a model, not after deployment.

How long does deployment take and how much does it cost?

It depends on the scope of the project, which is why a concrete quote only comes after a call. What is worth knowing up front:

◆Every project is different - a document chatbot for an accounting office has one scope, invoice extraction for a pizzeria another, and an agent handling a contract from its creation to archiving yet another.
◆Three typical cost components - one-off deployment (analysis, build, testing), monthly maintenance and the bill for actual AI usage.
◆A concrete quote you get after the consultation and process mapping - a schedule, a cost range and the scope in writing, with no obligation.

// next_step No obligation

Put AI to work
in a real process

First we map your process, point to the first concrete use case for AI and show what genuinely makes sense for you - an assistant on your documents, an automation connecting your systems, voice handling or data extraction from invoices. We recommend, you decide.

Book a call Send a message arrow_forward

AI deployment in your real processes.