AI agents in marketing: what I have already automated in n8n

“AI agents” is the main buzzword of 2026. Most of the articles on the topic are theoretical: “opportunities”, “horizons”, “paradigms”. I'll tell you about 3 live agents in n8n, which are working on my projects right now: auto-generation of creatives from the price list, sentiment monitoring of comments, auto-responses in the TG bot via Claude Sonnet 4.6. With configs and savings metrics.

It pisses me off when influencers write “how AI agents will change marketing by 2030” - it’s a useless article. By 2030 we will see this. And I need right now, in April 2026, for me to have working automation. Here they are.

The AI agent is not a “revolution.” This is n8n + Claude API + 3 webhooks. Costs up to 2,000 ₽/month for API, saves 15–25 hours of work monthly. If you are interested in the overall picture of a marketer’s AI stack - overview of 9 tools in one summary.

Why n8n and not Make/Zapier

I tried three popular marketing automation platforms. I chose n8n for three reasons:

• Self-hosted. I keep n8n on my own VPS for 400 RUB/month. This gives complete control over the data (important for clients with NDA) and there is no ceiling on transactions.
• Direct integration with Claude API. Via HTTP Request node - without third-party plugins. One POST to the endpoint, the response is parsed by the JSON node.
• Visual workflow + code when needed. 80% of the work is drag-and-drop. When complex processing is needed, I write JavaScript inside a Code node, also through Claude Code.

Make costs $9–29/month and reduces operations (3,000 ops in the pro plan). It's too tight for my size. Zapier is even more limited. n8n is a compromise: it takes a little longer to set up, but then there’s no ceiling.

Agent 1. Price list → 50 creatives in Nano Banana Pro

Problem

A client at e-com electronics adds a new price list to Google Sheets once a week. You need to: for each product, generate 5 creatives of different formats (Instagram stories 9:16, VK carousel 1:1, Telegram banner 16:9, landing hero, remarketing), upload them to S3, notify the team in Slack that it’s ready.

Manual work: ~30 minutes per product × 10 new products per week = 5 hours weekly for the designer.

n8n-flow architecture

1. Trigger — Google Sheets watcher for changing the “New Products” sheet.
2. Loop on new lines — for each line we parse: name, price, key characteristics, link to the original photo.
3. Claude Sonnet (via HTTP Request) - generates a prompt for Nano Banana Pro, taking into account the client’s brand style (the prompt template is stored in Notion, n8n pulls it up).
4. 5 parallel calls Nano Banana Pro via kie.ai API - each with its own aspect ratio.
5. Webhook → S3 upload — all 5 PNGs are poured into the client’s bucket.
6. Slack notification — message in #marketing with preview and links.

Metrics

• Time for 1 product: 30 minutes → 90 seconds
• API costs: ~70 ₽ per product (5 × 14 ₽ for Nano Banana 1K + Claude prompt)
• Monthly savings: 20 hours of designer work = ~80,000 ₽ at a rate of 4K ₽/hour
• API costs monthly: ~3,000 ₽ (40 products × 70 ₽)
• Net savings: ~77,000 ₽/month per client

Agent 2. Sentiment monitoring of comments in VK + TG

Problem

Client in MedTech (dental brand) - once a day you need to collect new comments under posts in the VK community and TG channel, classify them (positive / neutral / negative / question-needs-an answer), add them to Notion as a feedback base, negative ones - instantly escalate them to Telegram support teams.

Manual work: The SMM specialist spent 1 hour a day on manual review and classification. On a 30-day cycle - 30 hours.

Architecture

1. Cron-trigger — every day at 09:00 starts flow.
2. VK API + Telegram Bot API — take away all comments for the last 24 hours from connected communities.
3. Claude Sonnet 4.6 — classifies each comment according to a 4-level scheme (positive / neutral / negative / question). Also extracts key topics (price, quality, support, brand).
4. Switch node — splits flow by type:
   • Negative → Telegram alert in support chat + entry in Notion with the tag “escalated”
   • Question → Telegram alert in SMM chat “Answer needed”
   • Positive/Neutral → entry into the Notion table “feedback log”
5. Daily-summary — at the end of the day, a separate flow counts the aggregates (how many are positive/negative, which topics predominate) and sends a Slack-summary to the CMO.

Metrics

• SMM time: 1 hour/day → 5 minutes (summary check only)
• Speed of reaction to negativity: from 8–24 hours to 5–15 minutes
• API costs: 0.04 ₽ per comment via Claude Sonnet 4.6 (easily included in the context window). For 200 comments/day - 8 ₽. For a month - 240 ₽.
• Monthly savings: 25 hours SMM = ~50,000 ₽ at a rate of 2K ₽/hour
• Bonus: the quality of classification is more stable than manual one (humans get tired, AI doesn’t)

The most valuable thing about this agent isn’t even the hours it saves. This speed of reaction to negativity. When a client writes “the doctors at the clinic are rude,” you have 15 minutes to respond, not 12 hours. This is a different service quality class.

Agent 3. Auto-responses in TG-bot via Claude Sonnet 4.6

Problem

EdTech client - TG bot for lead generation of online courses. Leads per day – 50–100. Every second lead asks standard questions: “when is the next flow”, “is there an installment plan”, “what is included in the course”, “what is the difference between the tariffs”, “are there any discounts”. Managers responded manually, with an average response time of 4–8 hours.

The goal is to automatically answer 80% of standard questions in 30 seconds, and escalate 20% of complex questions (individual consultation, contract, corporate formats) to a manager.

Architecture

1. Telegram Bot webhook → n8n receives every message
2. Code node — filters system messages, leaving only real text messages
3. Claude Sonnet 4.6 with system prompt:
   • Context: full FAQ + information about courses + prices + current promotions
   • Instruction: answer standard questions, for complex ones - return JSON { escalate: true, reason }
   • Style: client's voice (adult, no emoji, factually)
4. Switch node:
   • If the response is from Claude - send to TG to the user
   • If the response contains escalate: true - forward the message to the manager + reply to the user “the manager will respond within 30 minutes”
5. Logging — every request/response in Notion for quality

Metrics

• Average response time: 4–8 hours → 30 seconds
• Auto-response rate: 78% (75% threshold passed)
• Lead to sale conversion: +18% (quick response keeps interest)
• API costs: 1.2 ₽ per dialogue on average (included in Claude Sonnet 4.6 input/output ratio). For 100 dialogues/day - 120 ₽. For a month - 3,600 ₽.
• Freed up 1 manager for 70% of tasks - he now deals only with complex conversations and sales

Architectural Patterns I Reuse

Pattern 1. Claude as router

Don’t run the entire flow through strict “if/else” rules. Better - Claude receives the input data, returns JSON with classification, and the Switch node in n8n works with this JSON. The flexibility is enormous: you can change the classification rules with one phrase in system-prompt.

Pattern 2. Notion as “persistent memory”

All logs, FAQ, context are stored in Notion databases. n8n reads the context for each request from there and writes the results there. Advantage: the project owner sees and edits data directly in the Notion interface, without touching the code.

Pattern 3. Slack/TG for alerts and summary

Do not send every event - otherwise you will drown in notifications. Only:

• Real escalations (negative comments, complex questions)
• Daily summary at the end of the day
• Weekly digest on Mondays

Pattern 4. Cost monitoring

Each flow is logged into a separate Notion table: the number of requests, tokens, costs. Once a week - summary. Without this, AI costs quickly grow out of control.

What could NOT be automated?

Failure 1. Auto-generation of posts in the TG channel

I tried to make a flow “news from RSS → Claude rewrites in channel style → publication.” It worked on small channels. On channels with a thin voice - a failure: even Claude Sonnet 4.6 does not maintain the characteristic author's voice over a long distance. After 2 weeks, subscribers began to complain “the channel has become faceless.” Rolled it back.

Lesson: voice-critical content - not yet for auto-generation. Auto-draft + manual editing - normal. Full automation is too early.

Failure 2. Auto-bidding in VK Ads via API

Idea: Claude analyzes CPL by ad group and suggests bid changes. After 3 weeks, it turned out worse than VK’s built-in auto-bidding. VK has its own auction signals that are not available externally. Closed.

Lesson: do not replace the built-in ML engines of advertising platforms with your own. They learn from data you don't have.

Cost and ROI: summary table

30x return on investment in AI stack. This is on one client; Similar agents work for me in other projects.

Where to start if you want to repeat

1. Get n8n on Vercel or DigitalOcean — there are ready-made guides, 1 hour from scratch. Free VPS on DO until the end of the trial.
2. Get Claude API key — Anthropic Console, $5 free credit at launch.
3. Start with the simplest agent — sentiment-monitoring or auto-tagging FAQ. Not immediately with auto-generation of content.
4. Log everything — every request in Notion, including costs. In 2 weeks the real unit-economics will be clear.
5. Don't automate voice-critical content right away — start with classification, filtering, routing. Auto-creative generation - later.

If you want me to analyze your case, write @dipustovalov. Often in a 30-minute call it becomes clear which 3-5 processes in your marketing will pay for the n8n + Claude API in the first month.

Related materials: CPA calculator, KPI dashboard template, a complete guide to performance marketing.