What is an AI agent for marketing in simple words?

An AI agent is a program that makes decisions and takes actions on its own. Don’t just write a letter, but check the lead → send a letter → add to CRM → write to the manager.

n8n or Make.com - which is better?

n8n - free self-hosted, more control. Make.com - easier, from $9/month. For the first agent - Make. When you hit the limits - n8n.

How much does it cost to run an AI agent?

n8n self-hosted on VPS 500-1000 ₽/month + Claude API $3-20/month. Total 1500-3000 ₽/month.

Which AI is better for an agent - Claude or GPT?

Claude - for solutions and analysis. GPT - for text generation. Best combination: Claude for decision making, GPT for content.

What tasks can be realistically automated?

Lead generation (parsing, qualification), content (collection, drafts), analytics (metrics, alerts).

Which trigger to choose?

According to schedule - for reports. By event - for leads. For 80% of tasks - scheduled once a day + events for critical actions.

Will an AI agent replace a marketer?

No. The agent saves 10-20 hours on routine work. I do the strategy, creative and negotiations.

How not to make an infuriating agent?

Don't write to people without asking. One task per agent. Always log. Golden rule: the agent must show what he has done and wait for confirmation.

All articles

AIMay 26, 202610 min

How to build an AI agent for marketing in 2026: step-by-step instructions with architectures

Step-by-step instructions for creating an AI agent for marketing: from choosing a platform (n8n, Make, Custom) to connecting LLM and setting up triggers. 3 ready-made architectures: lead generation with BANT qualification, content plan agent, analytics agent with alerts. Time-to-first-agent: 4-8 hours.

AI agents in marketing 2026 - not about writing a letter via ChatGPT. This is about a program that decides what image to generate and who to send it to. The difference between a prompt and an agent is like between a manual machine and a conveyor belt.

I have been collecting AI agents on n8n since the summer of 2024. The current one is a content plan agent: collects news, analyzes via Claude, generates posts, sends them to TG for approval. There are currently 8 agents on 6 projects.

AI agent is a set-it-and-forget-it thing. I set it up, the first week I repaired it every day, the second - every two days, on the fourth I forgot that it existed and it fell.

1. What is an AI agent

Script: if CPL is above 500 → alert. AI agent: CPL increased to 700 → Claude analyzes why → makes a decision: reduce the rate by 15% and send an alert with the analysis. The difference is flexibility.

2. Platform selection

n8n — free self-hosted. 400+ integrations, full control. On my projects, 6 out of 8 agents are on n8n.

Make.com — cloud, from $9/month. Easier, faster. There are fewer integrations.

Custom - your code. Full control, development time 5-10 times higher.

Parameter	n8n (self-hosted)	Make.com	Custom
Start cost	500-1000 ₽/month	$9-29/month	0 ₽
Time until first agent	2-4 hours	1-2 hours	2-5 days
Integrations	400+	200+	Any
For whom	Experienced marketer	No code	Developer

3. What LLM is under the hood?

Claude - my main choice. On 8 agents, ~94% decision accuracy versus ~87% for GPT-4o.

GPT-4o - better for text generation and creative tasks.

Local LLM — for tasks where data cannot be sent to the cloud. For marketing, Qwen 2.5 32B is usually enough.

4. Lead generation agent architecture

New lead → Claude checks BANT → if passed → in amoCRM + alert to the manager in TG → if not - in the ripening list. In B2B, SaaS increased conversion to qualified leads from 35% to 58%.

5. Content plan agent architecture

Collects daily news from RSS TG feeds → Claude selects topics → generates drafts → in Notion for approval. The agent does not publish it himself - he only prepares it.

6. Analytics agent architecture

Once an hour, pulls data from the API → Claude compares it with the plan → if the deviation is greater than the threshold → alert in TG. On EZ KATKA I caught CPL drawdowns 2-4 hours before the manager.

7. Five mistakes

1. One agent for everything. One task - one agent.

2. Access to actions without moderation. On the critical path - human-in-the-loop.

3. No timeout. Agent in an infinite loop is a classic.

4. No logs. The agent is a black box without logs.

5. Weak LLM for complex solutions. For qualification - only Claude Sonnet or GPT-4o.

8. Conclusion

AI agents are not hype, but a necessity for teams that want to work faster. n8n + Claude API is a working package for 90% of tasks. Don't try to do it perfectly - build an MVP, launch it, fix it.

If you want to analyze your specific situation, write to Telegram or through form. Starting consultation - 0 ₽.