How much does an AI agent cost for a business in Kosovo and the region?
An AI agent has two kinds of cost: the build, paid once, and running costs, paid every month. This guide shows both, what moves them, and how to estimate the model's monthly bill yourself.
By Adrian Fusha, Pluton Studio
On this page
The two kinds of agent
An agent for messages answers customers on Instagram, Facebook, WhatsApp, Viber or your website, gives catalogue prices, checks size and stock and takes orders and bookings. An agent inside your other systems handles invoices into accounting, dispatch, answers from your records or leads into a CRM, and is the larger build, because every system it touches needs its own tools, connections and tests.
Both include their tools, a log of every action, approval on the steps you choose and tests on your real messages. Extra processes, systems and languages are itemised in the quote.
What you pay for
One-off build work, then monthly running costs:
| Part | What it covers | When you pay |
|---|---|---|
| Discovery | Mapping the process: where requests come from, which systems hold the data, which steps wait for your approval | The first call and the quote are free; on a larger system, mapping is part of the quoted work |
| Build, per process | The agent's instructions, its tools, the approval steps and the log for one process | Once |
| Integrations | A connection to each system the agent reads from or writes to: inbox, WhatsApp Business, calendar, booking system, accounting, database | Once, per system |
| Languages | Instructions, replies and tests for each language the requests arrive in: Albanian, English, German | Once, per language |
| Testing | Running the agent on your real messages and documents, and fixing what it gets wrong, before go-live | Once, part of the build |
| Model usage | What the language model provider charges for the text the agent reads and writes | Monthly, by volume, estimated in the quote |
| Channel fees | What the messaging platform itself charges. WhatsApp: Meta bills per delivered message, for templates and, since 1 October 2026, for replies beyond 1,000 a month per number, at the utility rate of about 0.6 cents to a Kosovo number, 1.75 cents to Albania and 4.6 cents to Germany (marketing templates cost more). Viber: €115 a month per bot, plus about 1.4 cents for each message the bot starts to Kosovo or Albania. Instagram, Facebook (Messenger) and web chat: no platform fee | Monthly, by volume; you pay WhatsApp's fees to Meta directly |
| Hosting | The server, database and log the agent runs on | Monthly |
| Maintenance | Updating rules, tools and replies as the business changes, reviewing the log, moving to a new model when a provider retires the old one | Monthly or per change, as agreed in the quote |
What raises the cost, and what lowers it
The same agent can cost very different amounts depending on these choices:
| Raises the cost | Lowers the cost |
|---|---|
| Writing into your other systems: accounting, dispatch, a CRM | Working in messages only: answers, orders, bookings |
| Several processes at once | One process first, the next once it works |
| Systems with no way in, which need exports or email parsing | Systems with a documented connection (an API) |
| Requests in several languages, or mixed and in dialect | One main language for the first version |
| Approval rules with several roles and amount limits | One person approving, or the agent confirming routine steps itself |
| Poor scans and handwritten documents | Clean PDFs and typed emails |
| High volume, which raises the monthly model bill | Plain rules for the cases that need no model at all |
How to estimate the model's running costs
Models are billed per million tokens (a token is a piece of a word), usually at one rate for the text the model reads and another for the text it writes.
An agent re-reads its instructions and its tools' data at every step, so one task often uses several thousand tokens, more for multi-step tasks and long documents. This site's assistant, for example, reads several thousand tokens of instructions and page index per question.
The estimate: tokens per task, times tasks per month, times the provider's current rate. 1,000 requests a month at 5,000 tokens each is 5 million tokens; multiply by the rate on the provider's price page. During testing we measure real tokens per task on your messages, so the quote's figure comes from your work, not a guess.
Rates change often, so we print no provider price table. A spending limit at the provider, where offered, and plain rules for cases that need no model keep the bill predictable.
Typical scopes and how long they take
Start with one process, live and measured, before the next. The free written quote fixes the build price for your scope and estimates the monthly costs.
| Scope | What it includes | First version |
|---|---|---|
| Messages | One agent answering your customers' messages: prices from your catalogue, size and stock, orders and bookings, with log and testing | 1 to 2 weeks on your website, 3 to 6 weeks on several channels |
| One process in your systems | One agent for one job inside your systems, such as supplier invoices read and prepared for accounting: one source of requests, the system it writes to, log and testing | 2 to 4 weeks |
| Several processes | Two or more jobs sharing tools and one log, such as orders plus lead follow-up, added one at a time | The first process as in the rows above, then one after another |
| Several systems | Agents working across several systems, such as bookings, accounting, customer records and messaging, often with a review screen for your staff | Planned in phases in the quote |
Where an agent sits among our services
Process automation runs the same fixed steps every time, even when one step uses AI to read a document. An agent decides which of its tools to use for each request. Which one fits:
| What you buy | When it fits |
|---|---|
| AI chatbot or AI agent for messages (the same offer): answers on Instagram, Facebook, WhatsApp, Viber and your website, gives prices, takes orders and bookings | The same questions keep arriving in messages, and orders or bookings can be taken right there |
| Process automation: one complete process, same steps every time | The work never changes shape, only the data does |
| AI agent in your systems: documents into accounting, dispatch, answers from your records, leads into a CRM | Requests vary, and the work has to continue in your other systems |
| Web app or platform | You need a whole system around the agent, with screens for your staff |
| Larger AI systems | Several agents and systems at once, planned in phases |
How to get an exact number
A call about one process, then a free, fixed written quote within two working days: build price, estimated monthly costs, timeline. No hourly rate. Payment is half at the start and half at go-live, and larger projects run in three milestones.
How we build agents, and the safeguards in each, is on AI agent development. Not sure you need an agent? Read AI agent or chatbot first, or see what a chatbot costs.
Frequently asked questions
How much does a simple AI agent cost?
A simple agent, one that answers customer messages and takes orders and bookings, is our smallest AI build and the same offer as our chatbot; one that works inside your other systems costs more. The exact build price comes in a free, fixed written quote within 2 days of a call, with the monthly model and hosting costs estimated alongside.
What are the monthly costs of an AI agent?
Model usage, billed per token and growing with volume; hosting; maintenance as agreed; and on WhatsApp or Viber, the platform's own fees: Meta bills WhatsApp per message, which you pay to Meta directly, and a Viber bot costs €115 a month. The quote estimates all of them for your volume before you sign.
Does an AI agent cost more than a chatbot?
Not for messages: a chatbot and an agent for messages are the same offer at the same price. The price rises when the agent acts inside your other systems, such as accounting, dispatch or a CRM, because each system needs its own tools, connections and testing.
Can we start small?
Yes, and we recommend it: one process, live and measured, then the next. Each extension is quoted before it starts.
Do you charge by the hour?
No. You get a free, fixed written quote within two working days of the first call. Payment is half at the start and half at go-live; larger projects are paid in three milestones.
How do I estimate the model costs myself?
Multiply tokens per task by tasks per month, then by your provider's current rate per token. One agent task often uses several thousand tokens; long documents use more.