Enter „how much does an AI agent cost” in Google: you will get answers from PLN 3,000 to PLN 400,000. Each is true at the same time. The discrepancy does not result from the dishonesty of suppliers, but from the fact that an „AI agent” is not one product. It is a family of solutions with a completely different architecture, scope of rights and level of integration. This article breaks down the price into first factors.
AI agent, AI assistant, chatbot , why these three words mean three different budgets
Before we get into the numbers, it is worth distinguishing the concepts, because the market uses them interchangeably, and this is the main source of disappointment in valuations.
| Type | What it does | Access to systems | Approximate CAPEX |
|---|---|---|---|
| Chatbot FAQ | Answers questions from the ready-made knowledge base | Only documents | PLN 5,000-15,000 |
| AI Assistant | Conducts a conversation, summarizes, drafts content | Limited reading from selected sources | PLN 15,000-40,000 |
| AI agent | Carries out tasks independently from start to finish | Full , reads and writes data in CRM, ERP, databases | 40-400 thousand PLN |
A real AI agent differs from a chatbot in that not only responds , takes action. It can send an email, update the offer in CRM, generate a PDF file and pass it on to the customer , without human intervention at every step. It is this autonomy that raises the price: the more systems it can save, the more work on the side of integration, testing and security guardrails.
If you're just getting started, our previous article about 6 tasks of an AI agent in B2B sales shows what specific processes are worth automating in the first place.
How much does an AI agent cost: four levels of architecture and their prices
The following breakdown is not an official industry taxonomy, it is a practical scheme that we use when pricing projects. The L1 level can be implemented in a few weeks. L4 is a multi-month project with its own team on the client's side.
L1: Knowledge base agent (only responds)
Works On rag architecture: takes excerpts from company documents and generates a response. He does not have access to any trading system, he cannot change anything, order anything, save anything. Typical use: internal helpdesk, customer FAQ, employee onboarding assistant.
L2: Agent with read-only integration
It can retrieve data from CRM, ERP or data warehouse, but does not modify it. Typical use: reporting on demand („how many orders from this quarter are over PLN 50,000”), preparing a brief before a conversation with the customer, KPI monitoring. Integration with an external system is the main element raising the price in relation to L1.
L3: Operational agent (reads and writes)
He has full access to at least one transaction system: he can update the status in CRM, create an offer, add a comment to the application, send an e-mail through the marketing automation system. It is the most frequently implemented type in B2B companies. Guardrails (what an agent can and cannot do) is a key element of the architecture , without them the operational agent is too risky to produce.
L4: Multi-agent system (orchestration)
Several specialized agents cooperate under the supervision of the orchestrator. One processes leads, the second generates bids, the third monitors the pipeline and escalates exceptions. Typical use: full automation of the B2B sales process or order handling. It is the project, not the product , that requires the customer to engage their own team throughout the implementation period.
| Level | CAPEX (implementation) | OPEX (monthly) | Implementation time | Typical return |
|---|---|---|---|---|
| L1: Knowledge base | PLN 8-20 thousand | PLN 300-800 | 2-4 weeks. | 6-10 months. |
| L2: Read-only | PLN 20-55 thousand | 700-2,200 PLN | 4 – 8 weeks. | 8-14 months. |
| L3: Operational | PLN 55-150 thousand | PLN 2,000-6,000 | 8-16 weeks. | 10-18 months. |
| L4: Multi-agent | 150-400 thousand PLN | PLN 5,000-18,000 | 16-32 weeks. | 14-28 months. |
What does the price really consist of, the internal distribution of the valuation
The valuation of an AI agent from a reliable supplier consists of several items. We show their indicative participation in the entire L3 project:
Note the position of discovery (15-20%). This is often the first element thrown out during negotiations to lower the valuation. Effect: the agent works in demo, it does not work on production, because not all edge times have been mapped. Or it works, but not the process that generates the most profit in the company.
Four costs that valuation usually does not show
Good suppliers take them directly into account. The others , no. It is worth asking about each of them individually:
1. Work on the client's side
The implementation of an AI agent requires dozens of hours of work of domain experts: process mapping, review of training data, decisions on the scope of guardrails, acceptance tests. In Polish SMEs, it is often 40-80 hours of work of the owner or sales manager. This cost is not visible in the valuation , but it is felt during the project.
2. Maintaining connectors after API changes
Pipedrive, HubSpot, Salesforce , each of these systems changes the API several times a year. The connector (integration) written in week 1 of the project may stop working in week 52 without any warning. The real OPEX should include the maintenance of connectors, not only the costs of the LLM model. Ask directly: „what will happen if the CRM changes the endpoint in 6 months?”
3. The cost of tokens at an unexpected scale
GPT-4o in the cloud version costs per token. With 50 queries a day , irrelevant. With 5,000 queries per day , a few thousand zlotys per month only on the API. The project valuation should include a TCO (Total Cost of Ownership) estimation at the expected volume plus a 5x growth scenario.
4. GDPR and AI Act Compliance
If the agent processes personal data of customers or employees, you need a DPA (Data Processing Agreement) with the model provider (OpenAI, Anthropic, Google). Since 2025, the EU AI Act has been in force, an agent classified as „high risk” (e.g. in recruitment or lending) has additional documentation requirements. In B2B projects with company data, the risk is usually low, but analysis should be part of the discovery.
Three specific scenarios from the Polish B2B market
Below are three cases that describe the types of projects we implement. The numbers are anonymized, but based on real valuations.
Scenario A: Lead qualification agent (B2B service company, ~60 leads/month)
What it does: downloads the application from the form → checks in the CRM whether the company was a customer before, → estimates the value of the transaction based on the industry and the size of the company, → I draft a brief for the salesperson with a priority recommendation by e-mail.
- Level: L2 (read-only with CRM)
- Budget: PLN 25-45 thousand implementation, PLN 700-1,400 per month
- Released time: 4-6 hours per week salesperson (pre-selection and prioritization)
- Return on investment: with 1 additional transaction closed per month (result of faster response) , 8-12 months
Scenario B: Bid Generation Agent (manufacturer to order, 150-300 enquiries/month)
What it does: receives a request for quotation (e-mail or form) → parses the technical specification, → queries the price list and inventory through the ERP API, → generates a PDF of the quotation with the calculation, → saves it in CRM and sends it to the customer with a request for confirmation.
- Level: L3 (write to CRM + mail + PDF)
- Budget: PLN 65-110 thousand implementation, PLN 2,200-4,500 per month
- Change Process offer preparation time from 2-3 hours → 15-25 minutes (agent + salesperson review)
- Return on investment: 10-16 months with 10% increase in the volume of offers handled by the same team
Scenario C: Customer onboarding agent (SaaS company/ software house, ~30 new customers/month)
What it does: after signing the contract collects data of a new customer → creates a workspace in the internal system → assigns a supervisor → sends a personalized welcome sequence (5 emails in 10 days) → monitors activation for 30 days and escalates to Customer Success if the customer has not completed key steps.
- Level: L3 (write to several systems)
- Budget: PLN 45-80 thousand implementation, PLN 1,500-3,000 per month
- Change Process elimination of manual onboarding steps (estimation: 6-8 hours per week CS team)
- Return on investment: 9-14 months, assuming an improvement in retention in the first quarter by several percentage points
When an AI agent makes financial sense and when it's better to wait
This is a question that we put on every discovery , and we have sometimes revived the project, because the conditions have not yet matured.
An AI agent makes sense when:
- Repeatable process consumes at least 15-20 hours per week in the company , with a smaller scale, the cost of implementation will not be reimbursed within a reasonable time
- Company systems have a documented API: no API lengthens the project 2-3x and dramatically raises integration costs
- You have someone internally who understands the process and can participate in the discovery , an agent built without domain knowledge requires expensive iterations
- The input data is reasonably structured , the agent on Excel data with manually typed values is a project, not an implementation
It is better to wait or choose an alternative when:
- Are you looking for a FAQ chatbot: here rag for 8-15 thousand PLN is enough, the agent is overkill
- The process changes quarterly , the cost of rewriting guardrails and logic can consume the entire budget saved
- Data lies in silos without API: first the data platform, then the agent
- Budget below PLN 20,000 for the L3 range: this is a signal that someone is omitting something in the valuation
We write more about how we assess the company's readiness for AI before implementation in article on the costs of implementing AI in a B2B company.
Agent AI vs Sat vs SaaS: comparative table
The most common question we get: „isn't it cheaper to hire a man?” The answer depends on the context, but the table gives a starting point for the calculation:
| Criterion | AI Agent (custom L3) | Full-time (junior,mid) | SaaS AI tool |
|---|---|---|---|
| Commissioning cost | PLN 55-150 thousand | PLN 5-15 thousand (recruitment + onboarding) | PLN 0-2 thousand |
| Monthly cost (steady state) | PLN 2,000-6,000 | PLN 8,000-14,000 (gross employer) | PLN 300-3,000 |
| Availability | 24/7, 365 days | ~8h per day, ~22 days/month. | 24/7 (depends on plan) |
| Alignment with the company process | High , built for your process | High , man learns | Low , standard flows |
| Scalability | Virtually unlimited | 1 person = ~1 scope of work | Depends on API plan and limits |
| Boot time | 6-16 weeks | 8-20 weeks (recruitment + implementation) | 1-5 Days |
| The risk of „leaving” | None | High , process knowledge comes from the company | Moderate (vendor lock-in, SaaS churn) |
In practice, the AI agent and the job do not exclude each other, the agent takes over repetitive tasks, and the employee focuses on exceptions and relationships that the agent does not support. Conversations where „full-time vs. agent” is the only option on the table usually end in a bad decision in both directions.
Frequently Asked Questions
Can I first start with a cheaper chatbot and upgrade it to an agent?
In theory, yes, in practice rarely. The chatbot and the AI agent have a completely different architecture , the chatbot is mainly a conversational interface over documents, the agent is a set of tools with decision-making logic and integrations. Migrating a chatbot to an agent is often a rewrite from scratch, not an extension. If you want an agent in the end, it's better to design with this architecture in mind right away.
How long does it take for the agent to start generating value?
At L2 (read-only), the first effects are visible 4-6 weeks after kickoff. At L3, the pilot usually operates after 8-10 weeks, but before the agent handles 100% of the production volume, it usually takes 16-20 weeks (shadow mode phase when the agent operates in parallel with a human and the results are verified before taking over the action).
Which LLM models are most often used in agent projects and how do they affect the price?
B2B projects are dominated by GPT-4o (OpenAI) and Claude Sonnet (Anthropic) , both in the cloud variant. The difference in API cost between them is marginal today for typical B2B volumes. The choice of model has a greater impact on the quality of reasoning than on the price of the project , the supplier should justify the choice architecturally, not budgetary.
Do I need my own IT department to maintain the agent after deployment?
No. An agent designed for a company without an IT department has different maintenance requirements than an enterprise system. In a typical L2,L3 project, client-side maintenance is an hour per month: reviewing exception logs and reporting if something isn't working. The rest is handled by the supplier under the SLA. Requirements grow with L4 and large volumes.
What if the LLM provider (OpenAI, Anthropic) changes the API or raises prices?
This is a real risk. A good agent architecture isolates model calls behind the abstraction layer, changing supplier means replacing one module, not rewriting the whole. Ask the provider: „what does the model migration look like if I change the API provider in a year?” If the answer is „it's a big project,” the architecture is too tightly coupled.
Does an AI agent require signing a DPA with OpenAI or Anthropic?
If the agent processes personal data, yes. OpenAI and Anthropic have standard Data Processing Addendums for API clients: they are signed by the implementing company or the client directly. If you process particularly sensitive data (health, finance, HR), it is worth considering on-premise model , the data then does not leave the company's infrastructure.
How to price an AI agent if I don't know the exact scope yet?
By discovery. A reliable supplier will propose a 2-4-week discovery phase (cost PLN 5-15 thousand), after which you will get a detailed valuation with a breakdown into components. Pricing based on one phone and a general description of the process has too large a margin of error to serve as the basis for an investment decision.
Do the PARP or KPO grants cover the implementation of the AI agent?
Yes, in several ways. Measure 1.1 PARP (SMART Pathway) and regional funds for SME digitization include AI implementations as an eligible cost. Conditions change every round, but indicatively: co-financing up to 50-70% of eligible costs, minimum project from PLN 100,000, implementation period 6-18 months. We write more about this in an article on the costs of implementing AI in a B2B company from the perspective of PARP.
Next step
You already know how much it costs.
You don't know what pays off yet.
During a free call (30-45 min), we will assess your process in terms of readiness for automation: which agent level makes sense, where are the non-obvious risks and whether there is a better alternative than full implementation.
