An AI agent for business: tasks, architecture and a realistic budget

An AI agent for business is not an "employee in a box" but a program in which a language model decides the next step: reply to a customer, search the knowledge base, create a CRM record, ask a human for clarification. It pays off where a task repeats and can be verified. Below: which tasks fit, what an agent is built from, how to cap spending and what to check before launch.

Which business tasks suit an agent

The criterion is simple: the task repeats, the result can be verified, the cost of a mistake is limited.

• First-line support. Answers from a knowledge base, clarifying questions, handoff to a human when the question is outside the base.
• CRM work. Parsing an incoming request, filling card fields, hinting the manager at the next step. An integration with a specific system is covered in the article on amoCRM↗.
• Documents. Extracting fields from contracts and invoices, reconciliation, short summaries.
• Internal requests. "Find in the base what we promised client X", "prepare a draft reply to a tender".

Where an agent is weak: decisions with a high cost of error (payments, legal obligations, dismissals), tasks without a clear correctness criterion and anything a human cannot check quickly. Here the agent prepares a draft and a human confirms.

What an agent is built from

Any agent consists of five parts. If one is missing, you get a chatbot or, worse, uncontrolled automation.

1. A model. Makes decisions and writes replies. Choose by price, context and strength index, then verify on your own scenarios.
2. Tools. Functions the agent can call: knowledge base search, reading a customer card, creating a task. Each tool is a separate function with permission and argument checks. The basics are shown in the article on an agent in Python from scratch↗.
3. Memory and data. Conversation history, company documents. For knowledge base search RAG is the usual approach: only the relevant document fragments go into the request, not the whole base.
4. Guardrails. A maximum number of steps, a maximum number of tokens, a list of allowed actions, human confirmation for risky operations.
5. A journal. What the agent did and why: without it, debugging mistakes is impossible. Write the journal without personal data: mask phone numbers, addresses and document numbers before writing.

If you do not want to write the agent loop yourself, look at ready agent shells such as Hermes and OpenClaw: guides for connecting them to Zukko are in the Agents section of the documentation. Check the capabilities and limitations of each in their own documentation.

How to cap spending

An agent has a dangerous property: the number of steps inside one task is unknown in advance. A logic error turns into a loop, and a loop into a bill. Protection is built in layers:

• A step limit in code. No more than N model calls per task, then handoff to a human.
• max_tokens on each call. Caps the reply length.
• A separate key for the agent. In the dashboard each key has a name, a daily and monthly dollar limit and a list of allowed models. When the limit is exhausted, a 429 error with the code key_daily_limit_reached or key_monthly_limit_reached is returned, and the agent spends not a cent more. Details are in the key limits and cache documentation↗ and in the article on spending limits for agents↗.
• A model cascade. A cheap model handles routine work, hard cases escalate to a strong one; pick the model for each level by testing on your own requests.
• Monitoring. Spending is visible in the dashboard and via the API; you can build your own threshold alert on top of the API data.

A monthly budget: an example with a formula

Calculated for 18 October 2026. Assumptions for the example: a first-line support agent handles 1,500 conversations a month. Each conversation takes 5 model calls of 3,000 input tokens (system prompt, knowledge base fragments, history) and 1,000 output tokens in total. So per conversation: 15,000 input tokens and 1,000 output. The cache is not counted, so the estimate has a margin.

Formula: (input × input price + output × output price) × number of conversations.

Hence "a budget without illusions": tokens are, as a rule, the cheapest part of the project. The main costs lie elsewhere: integration with your systems, knowledge base preparation, testing, staff time for reviewing answers. And two caveats. First: the real number of calls per conversation depends on your logic, so measure it in a pilot. Second: a cheap model does not guarantee the required quality, and an expensive one does not guarantee the agent will not err; compare models on real conversations.

What to check before launch

A short checklist for an agent:

1. You have a set of 50–100 real requests with reference answers; the agent was run on it and you read the results yourself.
2. The agent has no rights to irreversible actions without human confirmation.
3. There is a step limit, max_tokens and a key with a daily limit.
4. The instruction "if you do not know, hand over to a human" was tested on questions outside the knowledge base.
5. Customer data does not reach the journal in plain form. Before sending personal data to any external service, check legal requirements and your company's policy.
6. It is known who takes the conversation over from the agent if the customer asks for a human.
7. The agent does not promise what the company does not guarantee: prices, deadlines and discounts come only from verified sources.
8. Behaviour on provider errors is planned: retry and a fallback model. For example, on a 503 retry with a pause and then switch to another model from GET /v1/models.

Launch with a limited audience: for example, on one channel or on 10% of requests, and widen after a week of observation.

Where to start

Create a separate key for the pilot in the dashboard↗ with a small daily limit, take one task from the list above and build a minimal version with a step limit. After a week the Usage section shows the real cost of a conversation and you can recalculate the budget with the formula above. Payment is available via SBP, MasterCard or crypto.

FAQ

How is an AI agent different from a chatbot?

A chatbot replies with text. An agent also chooses actions: it calls functions, looks into the knowledge base, creates CRM tasks, and works in a loop until the task is solved or a guardrail trips.

How much does an AI agent cost per month?

It depends on volume and model. In the example above 1,500 conversations cost from $0.35 to $12 in tokens depending on the model. Count development, integration and answer review separately.

Can an agent work with amoCRM?

Yes, usually through the CRM's API or webhooks: the agent receives the request data, and your service checks permissions and writes the result. An example is covered in the article on amoCRM integration. Check exact API methods against the CRM's own documentation.

How do I keep an agent from burning the budget?

A step limit in code, max_tokens, a separate key with a daily and monthly dollar limit and spending monitoring. When a key limit is reached, requests get a 429 error.