AI agents for business
AI that independently handles repetitive company tasks: replying to customers, creating reports and transferring data. It works 24/7, never tires and never forgets.
Transparent Investment Ranges // Engineering Standard
A fixed price agreed before work begins, based on the audit. No hidden costs or endless hourly billing.
The first tangible production result in your business. I focus on one critical process, eliminating risk and delivering working software in a fraction of the usual time.
- Multitasking autonomous agent
- Integration with your private company knowledge base (RAG)
- API connections (Slack, WhatsApp, CRM, ERP)
- 30-day stability warranty and post-launch support
A comprehensive platform for the entire organization. Advanced decision loops, deep two-way database integrations, complete audit logging, and dedicated engineering support under an SLA.
- Multitasking autonomous agent
- Integration with your private company knowledge base (RAG)
- API connections (Slack, WhatsApp, CRM, ERP)
- Dashboard for agent decisions and logs
- Implementation instructions and a team workshop
- Priority SLA, 24/7 monitoring, and architect support
AI Agents & Autonomous Systems in practice
Engineering approach and system architecture
I solve business problems by starting with the process. No rigid ready-made templates: I build software around your procedures.01 // MODULE
Three types of agents that save you time
01
An agent for emails, orders, and reports
Handles your inbox, processes orders, and generates reports.
Eliminates manual, repetitive office tasks.
02
Sales agent
For example
for a B2B company: the agent checks quote requests, qualifies leads according to your criteria, and assigns them to a salesperson with a ready brief. It works around the clock, so no lead is lost over the weekend.
03
Customer support agent
Tickets and helpdesk are a separate service.
See customer support agents. AI agents in practice
02 // MODULE
How I design email and order agents that do the work rather than just talk
An AI agent is a program that carries out successive process steps: reads an email, retrieves system data, prepares a result, and stops when a decision should be made by a person. I start with one specific process, such as the order inbox or weekly reports.
01
An AI agent is a loop of decisions, tools and checks
An AI agent is not just a prompt or a chatbot with a different name.
A production agent receives a goal, context, tools and safety rules. It can retrieve API data, check a CRM record, create a task, draft a reply, save a note, generate a report or ask a person to decide. The difference is that it works within a process rather than simply generating text. At Michał Świtalski, I design agents around the smallest controlled scope. First, the agent acts as an assistant: it proposes a result and a person approves it. Then I automate low-risk categories. Only after logs and tests show consistent quality do I give the agent more autonomy.
02
Agent tools must be narrowly scoped and clearly described
An agent with access to everything is a risk.
I design narrowly scoped tools: retrieve customer data, draft a response, check an order status, add a CRM note, generate a report, or notify the team. Each tool has an input, output, validation, and a defined scope of responsibility. This makes testing easier. I can identify when an agent misread the intent, called the wrong tool, or produced a result that needs approval.
03 // MODULE
An AI agent is not a replacement for every automation.
An agent is best suited to a process requiring several decisions, natural-language work, and interaction with tools. If the task is simply moving data from point A to point B, a standard API integration will be cheaper and more stable.
When a process requires understanding an email, choosing a route, retrieving data and preparing an action for approval, an agent starts to have an advantage. That is why I distinguish three things from the outset: a chatbot as a website conversation, automation as a repeatable workflow, and an agent as a program that chooses its own next step and uses tools.
If you need customer support, I have a separate page about helpdesk agents. If the agent needs to write data to a CRM or ERP, I deliver that as a separate integration project.
The agent works through the task itself and stops where it needs a person to decide. Working model
04 // MODULE
An AI agent needs a task, tools, memory, and boundaries.
Task. I describe an agent's task as an operational process, not a wish addressed to a model. The agent must know when to begin, what data it may retrieve, what result to prepare, and when to stop.
Otherwise, it will generate answers that sound good but cannot fit the company's workflow. Tools Each agent tool has a narrow responsibility: retrieve a customer record, check a status, draft an email, add a note or generate a report.
I do not grant one unrestricted access to the system. This allows errors to be tested and blocked at the level of a specific action.
Memory. Agent memory does not mean retaining everything without control. I design it around the process: case history, customer preferences, previous decisions, project context, or a knowledge base.
Data needs a defined scope, retention policy and a clear source. Boundaries Boundaries matter more than the model's intelligence alone.
The agent should know what it cannot do: promise a price, change financial data, approve a complaint, send a message without approval or use data outside its scope. These boundaries are what make it safe to increase autonomy.
Tests I test the agent on real cases, including difficult and incomplete ones. I check whether it recognizes intent, chooses the right tool, can say “I don't know” and leaves high-risk decisions to a person.
Without these checks, the agent should not go into production. After deployment, it needs ongoing care: monitoring costs, response quality and integrations.
What You Receive with the Implementation:
- Multitasking autonomous agent
- Integration with your private company knowledge base (RAG)
- API connections (Slack, WhatsApp, CRM, ERP)
- Dashboard for agent decisions and logs
- Implementation instructions and a team workshop
Technologies and production standards:
I choose technologies for stability, speed, and reliability. No technical debt.
AI Agents & Autonomous Systems
FAQ
Frequently asked questions about AI agents and autonomous systems: architecture, implementation, data security, costs and technologies.
01 How does an AI agent differ from a chatbot? +
A chatbot is primarily a conversation channel: a website widget, knowledge base, leads and human handoff. An AI agent can use chat as its interface, but its value lies in planning next steps, calling tools, working with data and completing a process.
02 What types of AI agents do you build? +
An operations agent for emails, orders, and reports. A sales agent that qualifies leads and assigns them to a salesperson with a prepared brief. I describe customer support and ticket agents separately on the AI customer support agents page.
03 How much does an AI agent implementation cost? +
Automating a single process, such as email classification or a simple follow-up, costs 5–15 thousand PLN and takes a few business days to implement. A system of cooperating agents connected to an ERP or CRM, with contextual memory, costs 20–80 thousand PLN. Pricing always follows a free conversation and an initial needs assessment. A full implementation audit is a separate paid service, which I begin only when we start working together.
04 How do you maintain security with an autonomous agent? +
I assume the model can make mistakes, so I design logs, limits, roles, example-based tests, and human approval for risky decisions. Permissions are narrow: an agent drafting an email has a different scope from an agent updating the CRM. The system records who approved a decision and which data informed the response.
05 When does an AI agent not make sense? +
If a task is one-off, simple, and based on fixed rules, an agent may be excessive: conventional automation or a simple API integration will be cheaper and more stable. An agent becomes useful when a process requires understanding an email, choosing a path, retrieving data, and preparing an action for approval.
06 What stack do you use to build AI agents? +
Claude Opus 5, GPT-5.6, OpenAI Agents SDK 0.11, Responses API, LangGraph 1.3, and n8n MCP. Tool selection depends on the process; part of an agent project still consists of conventional automation.
Start building.
First conversation free.
A free 30-minute technical consultation. After the call, I will provide an architecture proposal, timeline and transparent quote within 24 hours.