AI Agents

AI agents are autonomous workers that handle outreach, follow-ups, meeting booking, and lead qualification on behalf of your sales team. Each agent operates independently, working leads 24/7 based on the persona and rules you define.

Key Concepts

Templates vs. Agents

ScendCore uses a template-based system:

• Templates are blueprints that define personality, goals, channels, and autonomy settings. Templates do not work leads directly.
• Agents are instances created from a template and assigned to a specific team member. Each agent works that rep’s leads independently.

This means you define your playbook once in a template, then create individual agents for each rep who inherits those settings.

Agent Types

Each agent type comes with a default name and persona. You can rename agents to match your team’s preferences.

Channels

Agents can operate across multiple channels:

• Voice — Handles phone calls (inbound and outbound)
• Conversation — Email, SMS, WhatsApp, and chat
• Both — Voice + Conversation

Creating an Agent

Creating a Template

1. Go to AI Employees in the left sidebar
2. Click + New Template
3. Configure the template (see Configuration below)
4. Save your changes

Creating an Agent from a Template

1. On the AI Employees page, find your template under Master Templates
2. Click + Create Agent
3. Select the team member this agent will belong to
4. Click Create Agent

The new agent inherits all settings from the template and is assigned to the selected rep.

Configuration

Each agent (or template) has several configuration tabs:

Overview

Basic information about the agent:

• Name — The agent’s display name
• Type — SDR, Receptionist, Recovery, etc.
• Channel — Voice, Conversation, or Both
• Objective — What this agent is trying to achieve

Configuration

Core personality and behavior settings:

• Business description — What your company does (gives the AI context)
• Custom instructions — Specific rules or guidelines for the AI to follow
• Greeting — The opening message or script for voice calls
• Qualification framework — How the AI qualifies leads (e.g., BANT)
• Booking URL — Calendar link for meeting booking (Calendly, Cal.com, etc.)
• Escalation triggers — Keywords or situations that trigger escalation to a human
• Voice — The TTS voice used for phone calls (Neural2 voices)

Knowledge Base

Upload documents that the agent can reference during conversations — product sheets, FAQs, pricing guides, etc.

Channels

Assign phone numbers, SMS accounts, WhatsApp accounts, and email sender profiles to the agent.

Skills

Enable or disable specific capabilities:

• Schedule management (book meetings, check availability)
• CRM writes (create/update contacts, log activities)
• Sequence enrollment
• Document generation (PDF, DOCX, PPTX)
• Web search and competitive intelligence
• LinkedIn lookup
• SMS sending

MCP Connections

Connect external tools via the Model Context Protocol (MCP), such as HubSpot CRM integration.

Autonomy Levels

Autonomy controls how much independence your AI agent has. There are four levels:

How Risk Scoring Works

Every job the AI creates gets a risk score from 0 to 1. The risk score is compared against the agent’s risk threshold to determine if approval is needed:

• Risk below threshold — Job auto-sends (at Controlled or Full autonomy)
• Risk at or above threshold — Job goes to the approval queue

The risk threshold is dynamic and adjusts based on the agent’s confidence score, which is calculated from:

Increasing Autonomy

To increase an agent’s autonomy level:

• The confidence score must be at least 50/100
• The agent must not be in email warmup mode
• The agent cannot exceed level 3 (Full Autonomy)

Tip: Start new agents at the “Supervised” level. As the AI learns from your approvals and edits, its confidence score increases, and you can gradually raise the autonomy level.

The Kill Switch

Owners can activate the Kill Switch from the AI Employees page. When active, it immediately halts all AI agent activity across the workspace — no messages are sent, no calls are made. This is a safety mechanism for emergencies.

Click the Kill Switch button again to resume normal operations.

Job Lifecycle

Every action an agent takes follows a strict lifecycle:

Pending --> Approval Required --> Queued --> Executing --> Completed
                |                    |           |
                v                    v           v
            Cancelled            Cancelled    Failed / Cancelled

1. Pending — Job is created and risk-scored
2. Approval Required — Risk exceeds threshold; waiting for human review
3. Queued — Approved (or auto-approved); waiting for the worker to pick it up
4. Executing — Worker is actively sending the email/SMS/call
5. Completed — Successfully delivered
6. Failed — A system or provider error occurred
7. Cancelled — Rejected by a human, or halted by the kill switch

No job can skip states or go backwards. This ensures a complete audit trail for every AI action.

Self-Learning AI

Every AI agent learns from outcomes automatically. There is no manual training step — the system observes what works and adjusts future behavior across five mechanisms.

Cross-Call Memory

At the start of every voice session, the agent receives a summary of the last 10 calls with the same contact. This gives continuity — the AI remembers what was discussed, what objections were raised, and where the conversation left off.

Cross-call memory is always active for voice-enabled agents. No configuration is needed.

Contact Persona Evolution

After each interaction (email draft or voice call), ScendCore extracts communication signals from the conversation. These are stored on the contact’s AI persona profile and build up over time. After two or more interactions, the persona is injected into future drafts so the AI adapts its tone and approach to each individual.

Strategy Weight Injection

ScendCore tracks which persuasion approaches (strategy tags) lead to wins across your team. Before drafting any message, the system queries win rates by strategy tag and injects the top-performing strategies into the prompt. Your agents collectively benefit from every successful conversation.

Draft-Time RAG

Before generating an email or SMS draft, ScendCore searches your historical conversation outcomes using vector similarity. It retrieves similar past wins — what messages, angles, and CTAs worked for comparable contacts. The AI uses these as context and inspiration, not as templates to copy.

Outcome Attribution

Every completed conversation is scored: meeting booked, deal won, reply received, no response, and so on. These outcomes feed back into strategy weights and vector embeddings, closing the learning loop. Over time, your agents get measurably better at converting leads.

Tip: Self-learning is fully automatic. The more conversations your agents handle, the better they perform. Monitor improvement on the Performance tab.

Performance Analytics

Each agent has a Performance tab that surfaces key metrics and call quality data. Navigate to AI Employees, click an agent, and select the Performance tab.

Date Range

Filter all metrics by time period: 7 days, 30 days, 90 days, or all time.

Metrics

The Performance tab displays:

• Meetings booked — Total meetings scheduled by this agent
• Conversion rate — Percentage of leads that converted to meetings or deals
• Lead quality distribution — Breakdown by BANT qualification level
• Total calls — Number of voice sessions handled
• Average call duration — Mean length of voice conversations
• Sentiment — Overall sentiment analysis across conversations

Call Log

The Performance tab includes a table of recent voice sessions showing:

• Contact name
• Call duration
• BANT qualification state
• Timestamp
• Quality badges

Tip: Use the 7-day view to spot regressions quickly. If conversion rate drops, check recent calls for quality issues and adjust the agent’s custom instructions.