AI agent development for businesses in Kosovo and Albania

An AI agent is software built on a language model that takes actions in your systems through a fixed set of tools, such as creating a booking, filing a document or updating a customer record.

A chatbot answers. An agent answers and then does the work that follows: it reads the request, picks the right tool, fills in the data and, wherever it matters, hands the result to a person to approve. If answers are all you need, a chatbot is the simpler and cheaper choice, and AI agent or chatbot shows where the line sits.

The tools are narrow actions we write for your business, for example: look up a customer, create a draft booking, attach a document to a job. The agent can use those and nothing else. Agents are one part of our AI and automation work.

Five kinds of work where an agent earns its place. Each example shows what the agent does and what stays with a person.

An agent that can act can also act wrongly, so the limits are part of the build, not an add-on. Every agent we deliver has these five:

• A fixed allow-list of tools. The agent can call only the actions written for it, and each action checks its own inputs before anything changes.
• Human approval for anything involving money or anything irreversible: payments, invoices, refunds, cancellations, and any message that commits you to a price or a date.
• An audit log of every action: what came in, what the agent proposed, which tool it called, who approved it and when.
• A fallback when the model fails or is unsure. The case goes to a person with the context attached, and the manual way of working keeps running.
• Testing on real messages in Albanian, English and German before go-live. We run the agent on examples from your own inbox, anonymised where needed, and review the results with you before it touches live work.

Your data and the accounts the agent runs on are in your name. The agent sees only what its tools return, never your whole database.

The language model runs at a provider named in the contract, so you know where every request goes. On the Swiss platform described below, for example, the model runs at a provider in the United States, and the offer says so in writing.

The application and its data can be hosted in Europe where you need it. If the model itself must also run in Europe, we set that up and quote it separately. Where we process personal data for clients in the EU or Switzerland, a processor agreement under Art. 28 GDPR applies, and what it contains is published.

For a field-service company in Switzerland we built a complete operations platform with seven AI features. One rule governs all of them: the model proposes, deterministic code decides. The model never computes money, never writes to the database and never renders a layout, and a person accepts each proposal with one tap.

Two of those features do jobs from the table above. A pasted email or WhatsApp message fills in the job form, the customer is matched by code on phone number, email and name, and nothing is created until someone saves. An assistant answers questions about jobs, equipment and money from the real records, attaches the records it found, and says so when something is not in the data. Rough notes become billable lines from the company's own catalogue: the model cannot invent a price, only pick one. Every feature has a deterministic fallback, so a basement with no signal breaks nothing.

The assistant in this site's header, Pluton AI, follows the same rule at a smaller scale. It answers questions about our services, prices and how a project starts, in Albanian, English or German. The model only writes text: code lifts every link out of its reply and turns only addresses that really exist, our own pages and our clients' live sites, into buttons, while an address the model makes up is dropped. The WhatsApp and call buttons are added by code when a visitor asks about price, timing or getting started, messages that try to rewrite its instructions get a fixed reply without reaching the model, and every question and answer is logged.

Work

Service Company / Switzerland · 2026

Field-Service Platform

Operations platform for a Swiss service company: dispatch, iPad job reports, equipment rental, timesheets, German documents and an AI layer that proposes and never decides.

Pluton Studio is a software studio in Prishtina, Kosovo, founded in 2018 by software engineer Adrian Fusha. You work directly with the engineer who builds your agent, from the first call to go-live:

• A call about one process: where the requests come from, which systems hold the data and what must never happen without a person.
• A fixed written quote within two working days of the call.
• A first working agent on one process in 2 to 4 weeks, tested on your real messages before it goes live.
• Then we extend it: the next process, another channel or another language, each quoted before it starts.
• Half at the start and half at go-live; above 3,000 euros, three milestones. Any defect found within 30 days of go-live is fixed at no cost.

A first AI agent on one process starts at 1,500 euros. More processes, systems and languages are itemised in the quote.

Running costs for the language model are separate. They depend on how many messages or documents the agent handles, and the quote estimates them per month. Larger AI systems, with several agents across several systems, are quoted after a call.

If the work always follows the same fixed steps, process automation is often enough and simpler to run. How an agent's price is built, and how to estimate the model's running costs yourself, is in what an AI agent costs.