AI and data
Custom AI agents that work with your data
An AI agent is not a chatbot with canned phrases. It finds the answer in your documents, performs an agreed step in your systems (creates a record, drafts a reply, sends an alert) and, when unsure, hands the case to a person together with the context.
Who it is for: For companies where the team answers the same questions daily, searches internal material or copies data between systems by hand. Typically e-shops, service firms and B2B companies from a handful to a few hundred people.
- From
- CZK 25,000
- Typical timeline
- 6–8 weeks
- You get
- A live agent, documentation, a quality report
- Who does the work
- Me directly, no middlemen
01
This is probably what hurts.
- Support or sales keeps answering the same twenty questions.
- Knowledge is scattered across emails, PDFs, Notion and two or three people’s heads.
- New hires take months to get up to speed because there is nowhere to ask.
- You tried a generic chatbot and it answered things you never said.
- Questions arrive in the evening and at weekends and wait until morning.
If you nodded at least twice, it is worth a conversation.
02
What we actually solve.
Typical cases clients bring. Yours may look different; we always start from what you need.
Customer support with cited sources
The agent answers from your website, price list and documentation and shows where each answer came from. When unsure, it hands the question to a person with the context filled in, so the customer never repeats themselves.
Internal knowledge assistant
Staff ask about policies, procedures or product specs. The agent answers from current documents and respects who may see what.
Enquiry qualification
The agent collects the brief, asks for missing details, writes the enquiry into your CRM and proposes a call. Sales gets a complete brief, not a one-line email.
Draft replies and proposals
The agent drafts an email reply, a call summary or the groundwork for a quote. Nothing is sent until a person approves it.
Operations assistant over your systems
Ask “how many orders await approval?” and the agent queries the system, answers and, after your confirmation, performs the follow-up step.
03
How it works, step by step.
For each step you see what I do, what I need from you and what you hold afterwards. No black box.
- 01
Brief and use-case selection
Week 1I do
I walk through where the team loses time, choose one bounded case with the best benefit-to-risk ratio and propose a measurable goal.
You do
Describe the current process and send 10–20 real examples of questions or tasks.
You get
A one-page brief: use case, boundaries, success criteria.
- 02
Data and permissions
Week 1–2I do
I map the sources, clean and index them, and set access so the agent only sees what it may.
You do
Provide the documents and flag what is sensitive.
You get
A knowledge base and a list of sources with permissions.
- 03
Prototype
Week 2–4I do
I build a first working version, connect the model to sources and tools and run it in a shared test environment.
You do
Try it with your own questions and mark what is off.
You get
A working prototype you can try yourself.
- 04
Evaluation
Week 4–6I do
I build a test set from sample questions and measure correctness, admitted uncertainty and hand-offs. I tune retrieval and rules.
You do
Approve the boundaries: what the agent may do alone and what it may only suggest.
You get
A test report and a list of known limitations.
- 05
Deployment
Week 6–8I do
I deploy the agent where the team uses it (website, Slack, email, internal tool), set up logging and alerts and hand over documentation.
You do
Name an owner of the agent on your side.
You get
A live agent and operating documentation.
- 06
Operation and tuning
After launchI do
For the first month I review real conversations and fix what production reveals. After that: ongoing care by agreement, or handover to your team.
You do
Feedback from daily use.
You get
A monthly quality overview and improvement proposals.
04
What it can look like in practice.
Model situation: a technical-parts wholesaler
A small wholesaler gets dozens of questions a day about availability, compatibility and delivery times. Two people answer, each time checking three different sources.
Before
- Answers take hours; evening questions wait until morning.
- Information differs depending on who replies.
- Salespeople spend time searching instead of selling.
Process
- The agent reads the catalogue, stock data and common questions.
- It answers and cites the source of every fact.
- It hands an uncertain question to a person with context.
- After approval it creates the enquiry in the CRM.
After
- Routine questions are answered instantly.
- Every answer has a traceable source.
- People handle exceptions and sales.
An illustrative scenario for orientation, not a specific client and not a promised result. We measure the real effect on your data in the Evaluation phase.
05
What I build with, and why.
- Language models via API
- E.g. Claude or GPT. I pick the model by task, cost and data requirements and am not tied to a single vendor.
- Vector search (RAG)
- Answers come from your documents, not from the model’s memory.
- TypeScript / Python and APIs
- Integration with your CRM, helpdesk, e-shop, Slack or email.
- Make / n8n
- Where a ready connector is enough, I do not write needless code.
- Logging and a test set
- Every answer is traceable and quality is measured, not guessed.
What you get at the end.
- Brief and success criteria
- A knowledge base over your sources
- An agent connected to the agreed tools
- A test set and quality report
- Deployment with logging and alerts
- Documentation, source code and team handover
06
What it costs and how long it takes.
from CZK 25,000
Typical timeline: 6–8 weeks
Running costs (model API, hosting) are paid directly to the vendors. I estimate the monthly cost up front from your expected volume.
Full price listWhat moves the price
- Number of data sources and their quality
- Number of tools the agent writes to
- Security requirements and data location
- Number of channels and languages (web, email, Slack)
When it does not make sense.
I would rather say so now than have you pay needlessly.
- You have no material and nobody who knows how things are done. An agent cannot invent a process that does not exist.
- You need 100% accuracy with no human review for decisions with legal or financial consequences.
- There are only a few questions a week. Answering by hand is cheaper and a good FAQ is enough.
07
What people ask most.
How is an AI agent different from a chatbot?
A classic chatbot follows fixed branches or answers from the model’s general knowledge. An agent has access to your data, can use tools and complete a task, for example create a record. Crucially, it has rules for when to stop and hand over to a person.
Can the agent work with internal or sensitive data?
Yes, but first we define where data is stored, who can access it and which services process it. The agent gets only the permissions its task needs. Sensitive material is not sent to external services without agreement.
What if the agent answers wrongly?
That is why there is an evaluation phase: I test on real questions, measure errors and set what the agent must never do alone. In uncertain cases it should admit uncertainty and hand over. Every answer has a source, so you can trace and fix a mistake.
What does it cost to run after launch?
Mainly the language-model API and hosting. For a typical small-company workload that is usually hundreds to low thousands of crowns per month. You get a firmer estimate before we start, based on expected volume.
Who owns the finished agent and code?
After completion and payment you receive the source code, access and documentation. Licences of external services the agent uses are listed in the proposal.
How soon will I see a first result?
A working prototype usually within 2–4 weeks of kickoff. You can try it with your own questions before it is deployed.
Does the agent speak Czech?
Yes, current models handle Czech very well. I still verify quality on your real questions, including specialist terminology.
What should I prepare before the first call?
Just know what slows the team down today and have a few real examples of questions or documents at hand. We cover the rest on the call.
08
Let’s talk it through.
Write a few lines about what you are dealing with. I will reply within 24 hours with a view on whether and how it makes sense, before you commit to anything.