AI agents & LLM applications
I'll build you a conversational agent trained exclusively on your company's knowledge. It answers questions from customers and colleagues, performs actions in your systems, and hands complex cases to humans, with full context.
Who it makes sense for
- Your support or helpdesk handles dozens to hundreds of repetitive questions a day
- The answers exist, but they're scattered across documentation, wikis and e-mails
- Public AI tools are off the table due to sensitive data or regulation
What you get
- An agent over your knowledge base, every answer cites its source
- Connected to the channels you already use: web, Slack, e-mail, internal systems
- An escalation protocol: uncertainty goes to a human, the agent never guesses
- Runs in a private cloud, APIs with no data retention
How the agent reaches an answer
Every query takes the same path, and at the end there's always a source citation, not a guess.
What it looks like in production
Sample conversation (illustrative data).
What an agent can and cannot do
An agent is not a chatbot with answers written in advance. The difference is where the answer comes from and what the agent is allowed to do with it.
It answers from your documents
The knowledge base can be a mix of PDFs, wiki pages, emails and spreadsheets. The agent finds the passage it is answering from and shows that passage alongside the answer. When the sources do not contain the answer, it says so and offers a human. That is the whole difference between an evidenced answer and an invented one.
It takes action in your systems
Open a complaint, move an appointment, issue a proforma invoice, look up an order status. Actions always go through your APIs and are bounded by a list of what is permitted. Irreversible steps can be locked so a human has to sign them off.
It hands the case over with context
Escalation is not a transfer back to square one. The person receives the whole conversation, what the agent already established and why it was unsure. That cuts the time they spend on the case even when the agent did not solve it.
When I will talk you out of it
Not every company needs an agent, and it is cheaper to find that out now than after deployment.
- When you have no documentation. An agent can only answer from what exists. If the answers live in people's heads, the first job is writing them down, not AI. Skip that and you get an expensive guess generator.
- Under ten questions a day. Running and maintaining it will not pay for itself. Usually a better form or three paragraphs of help text does the job.
- When the answers change weekly. A knowledge base rewritten faster than it can be indexed turns the agent into a source of stale information.
- When it would decide about money or about people. That belongs to a fixed rule and a human, not to a model. I am happy to tell you where the line sits.
Parameters
FAQ
What models do the agents run on?+
Depends on the nature of your data and budget: commercial APIs with no data retention, or open-source models in your private cloud. I justify the choice in the solution proposal.
What if the agent answers incorrectly?+
The agent answers only from your documentation and backs answers with citations. When unsure, it hands the case to a human, it never guesses. I measure and report answer quality every month.
What does an AI agent cost?+
Deployment starts at CZK 25,000 for one bounded use case; a typical project with system integrations lands higher depending on how many. Add running costs for model calls, hundreds to low thousands of CZK a month at ordinary volumes. The cost breakdown is in How much does an AI agent cost, and full pricing is on a separate page.
How long does deployment take?+
Six to eight weeks from brief to live operation. The first two weeks go on collecting and cleaning the knowledge base, another two on the agent and the integrations, the rest on measuring answer quality and tuning. If your documentation is already in order, it gets shorter.
Will our data leave the company?+
Not without your knowledge. Depending on how sensitive the data is, either an API with no retention or a model running in your own environment. What goes where is written down before work starts, not agreed verbally. I will sign an NDA before looking at any system.
Does the model train on our data?+
No. The agent does not feed your data into training, it retrieves it at question time and attaches it to the answer. Your documents stay yours and can be removed at any point without rebuilding the model. Technically this is the RAG architecture.
Related
- Data, RAG & backend, the layer the agent draws its answers from
- AI automation, when it should process documents rather than talk to people
- Automation & integrations, when connecting systems with rules is enough
- When an AI agent pays off, three questions and the payback maths
- Pricing, prices and lead times in one place
It starts with a 30-minute technical assessment. I'll tell you whether an agent makes sense for you, and how much it saves.
Write to me, I'll reply within hours.