Artificial Intelligence in Sembot
Overview — How AI Works in Sembot
Philosophy: AI as Assistant and Executor, Not a Black Box
Sembot treats artificial intelligence as an active partner in marketing work — not as a tool that does things for the user in an opaque and unpredictable way. Every action performed by AI is transparent: the user can see what the agent did, why it did it, and can verify or undo the result at any time.
AI as assistant — answers questions, analyzes data, generates proposals for texts, reports, and strategies. It does not make decisions for the user, but provides reliable information and recommendations.
AI as executor — when the user issues a command (e.g., "update product titles in the feed according to this policy"), the agent can perform the action directly in the system, using its assigned skills and tools. Before performing critical actions, the agent always requests confirmation.
This approach makes AI in Sembot predictable, controlled, and effective.
Three Main Usage Contexts
| Context | Description | Typical Scenario |
|---|---|---|
| Chat (Marketer AI) | Interactive real-time conversation with the agent | Data analysis, answering questions, content generation |
| Workflow | Agent as a step in an automated process | Recurring campaign analysis, reporting, feed updates |
| Background Agents | Autonomous tasks running without user involvement | Campaign monitoring, alerts, overnight data processing |
Relationship Between Agent, Skill, Knowledge Base, and MCP
To fully understand AI in Sembot, it helps to know four key concepts and how they connect:
- Agent — an AI unit with a defined personality, a set of instructions, and a selected language model. On its own, it has no access to any data or tools.
- Skill (Chat Function) — a specific function the agent can perform (e.g., "edit feed", "fetch campaign report"). Without skills, the agent can only converse.
- Knowledge Base — a collection of documents and entries that the agent searches when it needs company-specific information (policies, procedures, product descriptions).
- MCP Server — an external tool connected to the agent (e.g., GitHub, Slack) that extends its capabilities beyond the Sembot ecosystem.
Marketer AI
Chat Interface — Conversation Threads, History, Favorites
Marketer AI is the main entry point to artificial intelligence in Sembot. Accessible from the left navigation bar under the AI assistant icon.
📸 Screenshot: Marketer AI main view — left panel with thread list, center chat area, message input field at the bottom.
Conversation threads — each conversation takes place in a separate thread. A thread has its own history and context; it can be resumed at any time. Threads are visible in the left sidebar, sorted by date of last activity.
History — the full message history in a thread serves as context for the agent. If you mentioned a specific feed or campaign earlier in the thread, the agent remembers it throughout the entire conversation — you do not need to repeat this information with every question.
Favorites — threads you return to frequently can be marked as favorites (star icon next to the thread name). They will appear at the top of the list regardless of the date of last activity.
📸 Screenshot: Sidebar with thread list — a thread marked with a star as a favorite is visible, with the remaining threads sorted chronologically below.
Tip: Instead of starting a new thread every time, continue an existing one when the topic is related to the previous conversation. The agent will have better context and responses will be more accurate.
How to Ask Questions — Good and Bad Prompt Examples
The quality of AI responses depends largely on how you phrase your question. Here are the key principles with examples.
Provide context
| Weak question | Better question |
|---|---|
| "Why isn't the campaign working?" | "The Google Ads campaign for the 'sports shoes' category has a CTR of 0.3% against an industry average of 2%. What could be the cause and where should I start the diagnosis?" |
| "Fix the feed" | "Review the product feed [feed name] and identify products with titles that are too short (under 50 characters) — list them with suggestions for improvement." |
| "Write me a text" | "Write a product description for AlpineX brand trekking boots, aimed at mountain hikers, max. 150 words, in an expert style — no jargon." |
Specify the expected response format
Poor:
"Summarize the campaign results"
Better:
"Summarize campaign results for the last 30 days in a table
with columns: campaign | revenue | ROAS | change vs. previous month.
At the end, add 2-3 sentences of commentary on what requires attention."Iterate instead of starting over — if the response is too general, clarify in the next message:
- "Make it more concise — maximum 5 points."
- "Focus only on campaigns with ROAS below 2."
- "Rewrite the recommendations section — I want specific numbers, not generalizations."
📸 Screenshot: Example conversation in the chat — user's question, followed by a detailed response from the agent in the form of a table with campaign data.
Loading Files and Documents into the Conversation
You can attach a file to any message, which the agent will analyze in the context of the conversation.
Supported formats:
- Documents:
.pdf,.docx,.txt,.md - Data:
.csv,.xlsx,.json - Images:
.png,.jpg,.webp(the agent will describe the content or analyze screenshots)
How to add a file:
- Click the paperclip icon or drag the file onto the chat window.
- The file will appear as a thumbnail above the text field.
- Write a question referring to the file and send.
📸 Screenshot: Message input field with an attached CSV file — file thumbnail visible and cursor in the text field.
Usage examples:
- Upload a Google Ads report in CSV format and ask: "Which products have a ROAS below 1.5? Sort from worst."
- Add a pricing policy file and ask: "Check whether the prices in this feed comply with this policy — identify any discrepancies."
- Upload a screenshot of an error and ask: "What does this error mean and what are the possible causes?"
Note: Files added to a conversation are available only within that thread. If you want the agent to have permanent access to them, add them to the Knowledge Base (see section 11.5).
Sub-threads — When AI Delegates a Task to Another Agent
When a task is too complex or requires specialized permissions, Marketer AI can automatically delegate part of it to a specialized agent, creating a sub-thread.
How it works in practice:
You ask Marketer AI to "update the feed and launch a campaign promoting new products." The agent may:
- Delegate feed analysis to an agent with feed editing permissions.
- Delegate campaign configuration to a Google Ads agent.
- Collect results from both sub-threads and present you with a consolidated summary.
Sub-threads are visible as nested items in the sidebar. You can open each one and see the full work history of that agent.
📸 Screenshot: Sidebar with the main thread visible and a sub-thread indented below — a clear hierarchical nesting.
The agent never delegates tasks without your knowledge — you will always see a notification in the chat that it is creating a sub-thread and to which agent.
Usage Examples
Product feed analysis
User:
"Analyze the feed [name]. Identify the top 20 products with the highest margin that do not yet have an active PPC campaign. Return in table format: product, margin %, reason for no campaign (if known)."
The agent fetches data from the feed, cross-references it with the list of active campaigns, and returns a table of results.
Campaign optimization
User:
"I have a Google Shopping campaign with a ROAS of 1.8 and a budget of $5,000/month. Suggest 3 specific changes that could improve the ROAS to 3.0 — describe each change: what to do, why it will help, what the risk is."
Report interpretation
User (with an attached CSV file):
"Here is the monthly report. Explain to me in plain language which product categories are growing and which are declining — and what this means for my advertising budget next month."
SEO content generation
User:
"Based on this pricing policy [PDF file] and these 10 products [list], generate SEO descriptions for each. Format: H1 title (60 characters), meta description (155 characters), description (200 words). Style: expert but accessible."
AI Agents
What Is an Agent — the Difference Between an Agent and a Regular Chat
A regular chat is a one-off conversation with an AI model — with no persistent context, skills, or permissions to act in the system. Once the conversation ends, the model "forgets" everything.
An agent is a configured AI unit that:
- Has persistent instructions (who it is, how it behaves, what it focuses on, what it avoids).
- Has access to specific skills (can perform actions in the system, not just converse).
- Can use a knowledge base (knows company procedures, policies, and data).
- Has a defined access scope (what it sees, what it has permissions for).
- Maintains consistency between conversations within the same context.
- Can have its own sub-agents — specialized units for specific tasks.
Analogy: A regular chat is talking to a stranger on the street. An agent is talking to your trained employee who knows the company, has access to the appropriate systems, and knows exactly what permissions they have — and what to do when they encounter a question outside their scope.
Types of Agents by Access Scope
Sembot distinguishes four agent scopes that determine who can use them:
📸 Screenshot: List of agents with visible filtering or scope labels — SYSTEM, USER, WORKSPACE, PROJECT.
System Agents (system)
Created and managed by Sembot. Available to all platform users — regardless of workspace or project. They cannot be deleted, but can be configured within the scope provided by the platform.
Example: Marketer AI — the default marketing assistant available to every logged-in user.
User Agents (user) — Private
You create them yourself and they are visible only to you. Ideal for personal workflows, experimenting with configuration, or specific tasks you do not want to share with the team.
When to use:
- An agent for generating product descriptions in your own writing style.
- An agent for quick analysis of data you work with daily.
- Experimenting with different AI models before deploying to the workspace.
Workspace Agents — Shared Across the Organization
Visible and available to all users within the same organization (workspace). Created by administrators or people with appropriate permissions.
When to use:
- An onboarding agent for new employees (knows HR procedures, company policies).
- An agent that knows global brand communication standards.
- An agent for handling general questions about company tools and processes.
Project Agents (project) — Dedicated to a Project
Assigned to a specific project in Sembot. Visible only to members of that project. They have access to data, feeds, and campaigns associated with the project.
When to use:
- An agent dedicated to serving a specific agency client (knows only that client's products and campaigns).
- An agent with pricing and offer policies specific to a given store.
- An agent analyzing results only from one Google Ads account.
Creating Your Own Agent Step by Step
- Go to Automation → AI Agents in the main menu (or Project → Agents for a project agent).
- Click the "New Agent" button in the upper right corner.
- Complete the configuration form (details below).
- Save the agent — it will appear on the list of available agents.
- Test the agent by clicking "Open Chat" or invoking it from Marketer AI.
📸 Screenshot: New agent creation form — visible fields: name, instructions, model selection, skill selection.
Agent Configuration
Name and Description
Name — a short, descriptive name (maximum 60 characters), e.g., "Feed Analyzer", "Google Ads Assistant", "Product Description Expert". The name is visible in the agent list and in the chat interface.
Description for parent agent (parent_desc) — if your agent will be invoked as a sub-agent by another agent, this description tells the parent agent when and why to delegate tasks to your agent. It should be concise and unambiguous.
Instructions (System Prompt)
The most important part of the configuration. This is the text the agent receives before every conversation and which defines its behavior. Good instructions include:
- Who the agent is and what role it plays
- What it focuses on and what it avoids
- What style it responds in (formal/informal, concise/detailed)
- What information it always includes in its response
- How it reacts to requests outside its scope
Example instructions for a campaign analysis agent:
You are a performance advertising expert (PPC) with 10 years of experience
in Google Ads and Meta Ads. You specialize in data analysis and campaign
optimization for e-commerce.
Always:
- Provide specific numbers and metrics, not generalizations
- Compare results to industry benchmarks (if known)
- End every analysis with a list of priority actions to take (max. 5 points)
Do not answer questions unrelated to marketing and advertising — refer
the user to the appropriate resource or agent.Job instructions (job_instructions) — additional instructions for specific tasks, separate from the main system prompt. Useful when you want to define detailed procedures for specific types of commands.
Choosing the AI Model and Provider
The choice of model affects response quality, speed, and cost.
| Model | When to Use |
|---|---|
| GPT-4 / Claude 3.5+ series models | Complex analyses, content creation, tasks requiring multi-step reasoning |
| Mini / lite models | Quick, routine tasks: classification, summaries, simple questions |
| Models with large context windows | Analysis of long documents, large CSV files, extensive feeds |
Tip: Start with the model recommended by default in Sembot. Change it only when you have a specific reason (e.g., you need faster performance in a workflow or you are processing unusually long documents).
Assigning Skills and Tools
In the interface: the "Skills" or "Chat Functions" section in the agent form. Select from the list of available skills those that the agent needs for its work.
Principle of minimalism: assign only the skills the agent actually needs. The fewer permissions, the safer and less likely the agent is to perform an unintended action.
Example: an agent for analyzing feeds only needs the "Feed Read" skill. An agent for editing feeds needs the "Feed Edit" skill. Do not give the agent the editing skill if it is only supposed to analyze.
Connecting the Knowledge Base
In the agent form, the "Knowledge Base" section — select the document collections the agent can use during conversations. The agent automatically searches the knowledge base every time a question may relate to its content.
You can connect multiple knowledge bases simultaneously — e.g., a global company base and a project base. The agent will search both and merge the results.
Sub-agents
The list of agents to which your agent can delegate tasks, along with a description of when to do so. The "Can delegate to sub-agents" option must be enabled for the agent to use sub-agents at all.
Example sub-agent configuration:
{
"id": 42,
"description": "Invoke this agent when the user requests editing or updating the product feed."
}When to Use Which Agent — Scenario Examples
| Scenario | Recommended Agent Type |
|---|---|
| Quick question about advertising strategy | Marketer AI (system) |
| Feed analysis for a specific agency client | Project agent with access to the client's feed |
| Generating descriptions in your own individual style | Private user agent |
| Onboarding a new employee (questions about company procedures) | Workspace agent with HR knowledge base |
| Recurring campaign analysis (weekly) | Project agent in a workflow step |
| Verifying price compliance with policy | Project agent with pricing policy in the knowledge base |
Skills (Chat Functions)
What Is a Skill — an Agent Without Skills Is Just a Conversation
A skill (also known as a Chat Function) is a specific executable capability of an agent — the ability to take a real action in the system, not just provide a text response. Without skills, an agent is an intelligent conversationalist, but cannot change anything in Sembot.
Analogy: An agent is an employee with knowledge and good judgment. A skill is access to a specific tool — the ability to edit a spreadsheet, send an email, update a database. Without tools, the employee can only advise.
Built-in System Skills and What They Do
Sembot provides a set of ready-made system skills:
| Skill | What it does | Requires confirmation |
|---|---|---|
| Feed Read | Fetches and analyzes the product feed (titles, prices, availability, attributes) | No |
| Feed Edit | Updates selected fields in the feed (titles, descriptions, categories, prices) | Yes |
| Campaign Report | Fetches campaign results (Google Ads, Meta Ads) — costs, revenue, ROAS, CTR | No |
| Campaign Management | Enables/disables campaigns, changes budgets, edits targeting settings | Yes |
| Knowledge Base Search | Searches connected knowledge bases to provide answers | No |
| Report Generation | Creates and exports a report to PDF or CSV | No |
| Notification | Sends a notification to the user or a specified channel | No |
📸 Screenshot: List of available skills in the agent configuration form — checkboxes or toggles visible next to each skill.
How to Assign a Skill to an Agent
- Go to the agent configuration (Automation → AI Agents → [agent name] → Edit).
- Open the "Skills" or "Chat Functions" section.
- Check the skills you want the agent to have available.
- Optionally configure the skill parameters (e.g., data scope, edit permissions).
- Save changes by clicking "Save".
📸 Screenshot: Skills section in agent editing — a list with checked and unchecked skills visible, Save button.
Action Confirmation — When the Agent Asks Before Acting
For skills that make changes to the system (feed editing, campaign modification, data deletion), action confirmation is enabled by default. Before acting, the agent displays a summary of the planned change and waits for your decision.
Example confirmation for the feed editing skill:
The agent is planning to perform the following action:
Skill: Feed Edit
Scope: 47 products from the "Sports Shoes" category
Change: Adding the word "New" at the beginning of each title
Do you confirm this action?
[Confirm] [Cancel] [Show Details]📸 Screenshot: Agent action confirmation dialog — summary of the change scope visible and Confirm / Cancel buttons.
The agent waits for your response and only then performs the action or cancels it. If cancelled, you can refine the command and try again.
Confirmation can be disabled for specific skills — recommended only for low-risk skills (e.g., report generation, data reading that does not modify anything in the system).
Skills with Sub-agents — Automatic Task Escalation
Some skills are too complex for a single agent. In such cases, a skill can automatically launch a specialized sub-agent to handle a specific part of the task.
Example: the "Full Campaign Optimization" skill might launch:
- A sub-agent to analyze the feed (checks titles, prices, availability).
- A sub-agent to analyze ad results (fetches ROAS, CTR, budgets).
- A sub-agent to generate recommendations (combines both data sources).
Results return to the main agent, which synthesizes them and presents you with a complete summary with action proposals.
Skill Examples in Practice
Campaign analysis skill
User: "How are my Google Ads campaigns doing this week?"
The agent runs the "Campaign Report" skill, fetches data from the last 7 days, and responds:
"Your 12 campaigns spent $8,340 this week with a combined ROAS of 2.4.
Best performer: Sports Shoes (ROAS 5.1, revenue $12,200) — deserves a budget increase. Watch out: Winter Accessories (ROAS 0.8, cost $1,100) — losing money, requires intervention or pausing."
Feed editing skill
User: "Find all products without an image and mark them as inactive."
Agent [displays confirmation]: "I found 23 products without an assigned image. Do you want to set their status to inactive?"
After user confirmation: the agent updates 23 products and reports: "23 products deactivated. You can view the list here: [link to filtered feed view]."
Knowledge Base
What the Knowledge Base Is For — Supplying Agents with Company Knowledge
An AI model by itself knows nothing about your company, products, pricing policies, or procedures. The knowledge base is a way to pass this company-specific knowledge to the agent in a structured and searchable form.
With the knowledge base, the agent can:
- Answer questions in accordance with your internal rules (it does not make up answers).
- Verify data (e.g., "Is this price consistent with the minimum price policy?").
- Create content consistent with the brand voice and company terminology.
- Follow procedures described in documents, not template responses.
Structure: Folders and Entries
The knowledge base is organized into a structure of folders and entries:
📚 Knowledge Base
├── 📁 Policies and Procedures
│ ├── 📄 Pricing Policy 2024
│ ├── 📄 Promotions and Discount Rules
│ └── 📄 Returns Handling Procedure
├── 📁 Products and Categories
│ ├── 📄 Category Descriptions
│ └── 📄 Technical Specifications
└── 📁 Marketing and Advertising
├── 📄 Brand Communication Tone
└── 📄 Prohibited Keywords (Google Ads)Folders are for organization — they do not affect how the agent searches, but make management easier, especially when the knowledge base grows to dozens of entries.
Entries are specific documents or knowledge snippets. Each entry has a title and content, optionally tags.
📸 Screenshot: Knowledge base list view — folder hierarchy on the left, entry list on the right with titles and last edit dates.
Adding Entries — Text, Documents, Files
Text entry — created directly in the Sembot editor. Ideal for short procedures, rules, lists, and FAQs.
Document — upload a .pdf, .docx, or .txt file. Sembot will automatically process the content and make it searchable by the agent.
Data file — upload .csv or .xlsx with data (e.g., price list, product list, category map). The agent will answer questions about specific values contained in the file.
How to add an entry:
- Go to Knowledge Base in the main menu or project settings.
- Select a folder or create a new one by clicking the "+" icon next to the directory name.
- Click "New Entry" and choose the type: text / file.
- Give it a title, add content, or upload a file.
- Optionally add tags to facilitate searching.
- Click "Save" — the entry is immediately available to connected agents.
📸 Screenshot: New entry creation form — title field, text editor with sample content, tags field, Save button.
Knowledge Base Scopes
Like agents, knowledge bases have their own visibility scopes:
| Scope | Who has access | Typical use |
|---|---|---|
| User | Only you | Personal notes, private templates, your own procedures |
| Project | All project members | Client specifics, policies for a given store, product descriptions |
| Workspace | The entire organization | Global company policies, communication standards, HR procedures |
When configuring an agent, you can connect knowledge bases from different scopes simultaneously. Example: connect the workspace base (global rules) and the project base (client details) to a project agent — the agent will know both.
How the Agent Searches the Knowledge Base (Semantic Search — What This Means in Practice)
Sembot uses semantic search, which means the agent does not look for identical keywords, but understands the meaning and context of the question.
Real-world example:
User asks: "Can I give this customer a 30% discount?"
The agent searches the knowledge base and finds the "Pricing Policy" entry containing the sentence: "The maximum discount for individual customers is 25%. Any exception requires approval from the sales manager." — even though the user wrote "discount" and the document says "discount" (a synonym), the agent understands these refer to the same concept and responds correctly.
What this means in practice:
- You do not need to use exact keywords — write naturally.
- The agent can combine knowledge from several different entries at once.
- The more specific and unambiguous the entry, the more accurate and confident the agent's answers.
- The agent indicates in the response which entry the information comes from — you can verify this.
Best Practices — How to Write Entries for AI to Use Effectively
Write in declarative form, not as questions
| Poor | Good |
|---|---|
| "Can discounts be given?" | "The maximum discount for retail customers is 25%. Discounts of 15%–25% require manager approval." |
| "How does the returns policy work?" | "The customer has 30 days to return a product without giving a reason. The return must be submitted through the customer panel or by email to [email protected]." |
Provide specific values and conditions
Poor:
"Prices should be competitive and market-based"
Good:
"Minimum product price = net purchase price × 1.15 (15% margin).
Exception: seasonal products in final clearance may have a 5% margin.
Prices below purchase cost are strictly prohibited."One topic = one entry — do not combine multiple unrelated topics in one long document. Sembot searches at the entry level; short, topic-specific entries work better than one 50-page document.
Use headings in longer entries — even if an entry is long, H2/H3 headings help the agent locate the relevant section faster.
Update entries — outdated information in the knowledge base can lead to incorrect responses. Treat the knowledge base as a living document — review it quarterly.
Knowledge Base Content Examples
Advertising procedure
# Google Shopping Campaign Creation Rules
## Launch Conditions
Every new campaign requires:
1. A defined goal: target ROAS or target CPA
2. An approved monthly budget in the CRM
3. A minimum of 50 active products in the feed
## Budgeting
- Daily budget = monthly budget ÷ 30 (rounded up)
- Maximum daily budget without manager approval: $500
- Test campaigns (first 14 days): no more than 20% of the planned budget
## Review
- Every campaign reviewed every 14 days
- Campaigns with ROAS < 1 are automatically paused after 7 daysPricing policy
# Minimum Margin by Category
| Category | Minimum Margin | Exception |
|---|---|---|
| Electronics | 12% | Seasonal promotions: 8% |
| Clothing | 20% | Final clearance: 5% |
| Accessories | 30% | No exceptions |
Promotional prices below the minimum margin require CRM system approval
before activation in the feed. Approval waiting time: up to 24 hours.Brand communication tone
# AlpineX Brand Voice
## Style
Expert but accessible. We avoid technical jargon without explanation.
We write like an experienced colleague with a passion for the mountains — not like a user manual.
## We always highlight
- Materials and their properties (breathability, waterproofing, weight)
- Certifications (EN 13781, GORE-TEX, bluesign)
- Conditions of use (up to what temperature, for what trails)
## We avoid
- Comparisons with specific competitors by name
- Superlatives without justification ("best", "revolutionary")
- Wording that promises a safety guaranteeTool Integrations (MCP Servers)
What MCP Servers Are — External Tools for Agents
MCP (Model Context Protocol) is an open protocol that allows AI agents to connect to external systems and tools. Thanks to MCP Servers, an agent can not only converse and use Sembot data, but also read and write data in external services.
Analogy: If the agent is an employee, then skills are its internal tools (access to the company CRM, product database). MCP Servers are access to external applications — GitHub, Slack, Google Sheets, client databases.
Each MCP Server has:
- A name and description (what it is and what it does).
- Configuration — tokens, API keys, project identifiers (stored in encrypted form).
- Status — can be temporarily disabled without removing the configuration.
Available Integrations
📸 Screenshot: List of available MCP Server templates in Sembot — cards or list with service logos (Slack, GitHub, Google Sheets, etc.).
| Category | Example Integrations |
|---|---|
| Communication | Slack, Microsoft Teams |
| Code Management | GitHub, GitLab |
| Data and Analytics | Google Sheets, Airtable, external SQL databases |
| CRM and Sales | HubSpot, Salesforce |
| Task Management | Jira, Linear, Notion |
MCP Server Configuration
MCP configuration is done directly in the agent editing form. Each integration has its own set of fields displayed as an accordion — click the integration name to expand the configuration fields.
Access token / API key — generated on the external service side. Once saved, it is masked (displayed as ••••••••) — the value is never visible again, but you can overwrite it with a new one.
Project / workspace identifier — e.g., Jira project ID, Slack channel ID, Google Sheet ID.
Special field: Sembot API Token — if the MCP server requires access to the Sembot API (e.g., to read feed data through an external tool), the form displays a dedicated component for token management. You can:
- Generate a new token with selected permission scopes.
- Extend the validity of an existing token.
- Change permission scopes.
📸 Screenshot: MCP Server configuration section in the agent form — accordion with integration name, expanded with text fields (e.g., "GitHub Token", "Repository URL"), Save button.
How to configure an MCP server:
- In the agent form, find the section with the integration name (e.g., "GitHub", "Slack").
- Click the section header to expand the configuration fields.
- Fill in the required fields (marked with an asterisk
*). - Click "Save" — the configuration will be verified and saved.
- If the configuration is invalid (e.g., incorrect token), the system will display an error message.
To remove an MCP server configuration, click the trash icon next to the section name and confirm deletion in the dialog.
Assigning MCP to an Agent
An MCP Server is configured directly in the given agent's form — there is no separate MCP management list. This means that the same type of integration (e.g., GitHub) can be configured differently for each agent (different tokens, different repositories).
The "MCP Access" option at the agent level must be enabled for the MCP configuration section to be active and for the agent to use connected servers.
When This Is Useful — Usage Examples
Agent with Slack access
User: "Send a summary of this week's campaign results to the #marketing channel."
The agent generates a summary from Sembot data and sends the message directly to the specified Slack channel — no need for manual copying and pasting.
Agent with GitHub access
User: "Check whether there are open issues related to parsing errors in the feed integration repository."
The agent fetches the list of open issues, filters them by tags and key phrases, and returns a concise summary with links.
Agent with Google Sheets access
User: "Update the report spreadsheet with sales data from the last 7 days."
The agent fetches data from Sembot and writes it to the specified Google Sheet — automatic report update without manual export.
Agent with external database access
User: "How many orders were placed for products from this campaign in the ERP system this month?"
The agent executes a query to the ERP database and returns the result without needing to log into the external system.
AI in Automation (Workflow)
Using an Agent as a Workflow Step
A workflow in Sembot is a sequence of automated steps triggered on a schedule or by a triggering event. An AI agent can be one of these steps — it can analyze data, generate content, make decisions, and pass structured results to subsequent steps.
A step with an agent differs from a regular automation step in that instead of a rigid IF→THEN rule, it performs intelligent analysis: it can handle exceptions, interpret ambiguous data, and adapt the output to context.
📸 Screenshot: Workflow editor with step diagram — an AI agent step with an agent icon embedded in the sequence between the "Fetch Data" step and the "Send Report" step.
Passing Data to the Agent and Receiving Results
In the AI step configuration in a workflow, you define three things:
1. Agent — select which agent will handle this step. You can use a system, project, or your own agent — any agent you have access to.
2. User prompt — an instruction for the agent on what to do. You can use variables from previous workflow steps in syntax:
Analyze the following campaign data from the last 7 days:
{{campaign_data}}
Identify campaigns with ROAS below 2.0 and suggest specific actions.
Return the response in JSON format according to the provided schema.3. Response parsing + JSON schema — if you want the agent's output to be structured and available as variables for subsequent steps, enable parsing and provide a schema:
{
"type": "object",
"properties": {
"campaigns_to_optimize": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": { "type": "string" },
"current_roas": { "type": "number" },
"recommendation": { "type": "string" }
}
}
},
"issue_count": { "type": "number" }
}
}If parsing is disabled, the agent's output goes to the next step as raw text.
📸 Screenshot: AI step configuration form in a workflow — "Select Agent" field (dropdown), "Prompt" field with sample text containing variables, "Parse Response" toggle, JSON schema field.
Example: A Workflow That Analyzes Campaign Results and Suggests Changes
Below is a complete example of a workflow triggered automatically every Monday at 08:00:
[START — trigger: Monday 08:00]
│
▼
┌─────────────────────────────────────────────────┐
│ Step 1: Fetch Campaign Data │
│ Type: Campaign Report (skill) │
│ Scope: Last 7 days, all active campaigns │
│ Output → variable: {{campaign_data}} │
└─────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────┐
│ Step 2: AI Analysis │
│ Type: AI Step (agent) │
│ Agent: PPC Campaign Expert │
│ Prompt: │
│ "Data from the last 7 days: {{campaign_data}}│
│ Identify: │
│ 1. Campaigns ROAS < 2 → actions │
│ 2. Campaigns ROAS > 4 → budget increase │
│ 3. CTR < 0.5% → A/B test proposal │
│ Return JSON with list of recommendations." │
│ JSON Schema: { recommendations: [...], n: int }│
│ Output → variable: {{ai_analysis}} │
└─────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────┐
│ Step 3: Condition │
│ IF {{ai_analysis.n}} > 0 │
│ THEN → Step 4 │
│ ELSE → End (no recommendations) │
└─────────────────────────────────────────────────┘
│ (if n > 0)
▼
┌─────────────────────────────────────────────────┐
│ Step 4: Send Report to Slack │
│ Type: MCP Slack │
│ Channel: #marketing-weekly │
│ Content: "AI Report — {{ai_analysis.n}} │
│ recommendations. Details: {{ai_analysis}}"│
└─────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────┐
│ Step 5: User Notification │
│ Type: Notification (requires confirmation) │
│ Question: "Should budget changes be │
│ applied automatically?" │
│ YES → Step 6 (campaign edit) │
│ NO → End │
└─────────────────────────────────────────────────┘
│ (if YES)
▼
┌─────────────────────────────────────────────────┐
│ Step 6: Update Campaigns │
│ Type: Campaign Management (skill) │
│ Data: {{ai_analysis.recommendations}} │
│ Action: Change budgets per recommendations │
└─────────────────────────────────────────────────┘
│
▼
[END]📸 Screenshot: Workflow diagram in Sembot's visual editor — step rectangles connected by arrows, AI step highlighted with a color or agent icon.
Best Practices When Using an Agent in a Workflow
Always define the JSON output schema — an agent in a workflow should return data in a predictable format so that subsequent steps can reliably process it. An undefined format leads to parsing errors and unstable automations.
Add a human confirmation step for risky actions — for workflows that make changes (feed editing, budget changes), it is worth adding a step requiring approval before the final action. Especially important at the start, when the workflow is new and unverified.
Monitor logs for the first few weeks — every workflow run with an AI step is logged. Check logs regularly for the first 2–3 weeks to make sure the agent is responding correctly and the JSON schema is always filled in correctly.
Test at small scale — before running the workflow on all campaigns or the entire feed, test it with a parameter that limits the data scope (e.g., only one campaign, only 10 products). A configuration error in the agent will be detected safely before it affects the entire database.
Use descriptive variable names — variables passed between steps should be unambiguous: is better than . Your future self or another team member will more easily understand what the workflow does without reading every step individually.
Sembot Documentation — Chapter 11: Artificial IntelligenceVersion: 1.0 | Date: 2026-05-19