Glossary for AI agents and Autopilot customer experience

If you’re evaluating AI agents or “Autopilot CX” platforms, it can feel like everyone is speaking a slightly different language:

• agents, tools, flows, grounding, guardrails, RAG, orchestration.

This glossary is a practical reference for CX, ecommerce, and revenue teams who want to understand the key concepts without getting lost in research papers.

Use this as a go-to reference anytime a vendor, consultant, or engineer introduces a new AI term.

1. Core concepts: Autonomous AI agents, Co-pilots & Autopilot

AI agent

A software “worker” powered by AI that can:

1. Understand user input (natural language processing, voice, etc.)
2. Decide what to do next based on goals and rules.
3. Take actions in your systems (e.g., look up an order, start a return, add to cart).

AI agents can multitask, interact with customers and available external tools, and perform tasks independently or proactively.

They don’t just complete simple tasks: they optimize to complete tasks in line with overarching goals, often automating routine and repetitive tasks to free up human workers.

Advanced agent technology automates complex tasks of problem solving, decision making, and informed decisions using data, context, and past interactions.

Agents can also handle customer issues, identify patterns to improve performance, and are guided by a utility function to evaluate success.

While agents are powerful, some essential tasks or qualities, such as ethical judgment or nuanced human empathy, may still require human input.

Agent types: Agents can be classified by roles and capabilities, such as customer experience agents, product advisors, or account helpdesk agents.

Each type may specialize in specific tasks or services, and agent types can include both autonomous and collaborative roles within multi-agent systems.

Agent technology: Agent technology refers to the architectures and frameworks that enable AI agents to reason, plan, use tools, and interact with users and other agents.

This technology underpins the ability of agents to operate autonomously, coordinate with other agents, and support complex workflows and decision-making processes.

Utility function: The utility function is a core component that guides an agent’s goal-oriented behavior by quantifying success, allowing the agent to evaluate and optimize its actions based on performance metrics.

Key difference from traditional chatbots: Agents don’t just send answers; they can act and change state in your tools.

Bots typically handle simple tasks: basic, rule-based interactions; while other AI agents can perform more complex tasks, multi-step actions and adapt proactively.

Insightful read: AI agents in customer experience: The next CX revolution.

Autonomous agent

An AI agent that can operate for multiple steps without constant human prompts.

Example: An autonomous ecommerce agent can:

• Ask follow-up questions
• Decide if a return is allowed.
• Start the return workflow.
• Send confirmation; all in one continuous flow.

Autonomous agents can automate repetitive and routine tasks and are capable of advanced problem-solving and decision-making.

AI co-pilot

A copilot is an AI that assists a human, not replaces steps:

• Suggests replies to support agents.
• Draft emails for sales reps.
• Summarizes conversations.

The human is still the “pilot” and decides what to send or do.

Autopilot CX

Short for Customer Experience on Autopilot.

It means AI agents handle front-line conversations and workflows (answering questions, running processes) across channels without needing a human on every thread.

Important: Autopilot CX does not mean AI runs your entire business (inventory, logistics, pricing). It focuses on:

• Conversations
• Conversions
• Support workflows

Agent job/role

The specific responsibility assigned to an AI agent.

Examples:

• Order Helpdesk – handles “Where is my order?”
• Returns Desk – manages returns & exchanges.
• Product Advisor – helps shoppers choose products.
• Cart Saver – recovers abandoned carts.
• Account Helpdesk – handles account questions such as billing, dues, and balances.
• Customer Experience Agent – provides support, resolves customer issues, and addresses concerns across channels to enhance the customer experience.

Each job can have its own rules, knowledge, available tools, and flows. Agents in these roles deliver services, understand customer needs for personalized experiences, and resolve customer issues and concerns.

Agent policy

The set of instructions and guardrails an AI agent must follow.

Examples:

• Never issue a refund above $X without approval.
• Follow this tone of voice.
• Never change the shipping address after an order is fulfilled.
• Always propose an exchange before a refund (where allowed).

2. Tools, actions, flows & orchestration

Tool/tool call

A tool is a capability that the AI can use to interact with your systems. A tool call is when the AI decides to use it.

Examples:

• get_order_status(order_id)
• create_return(order_id, line_items)
• add_to_cart(product_id, size)
• create_ticket(contact_id, issue)

Instead of “hallucinating” an answer, the AI calls a tool, gets real customer data, and responds based on that. Tool use is a key capability of AI agents, enabling them to access external systems via APIs and perform tasks effectively.

Defining available tools for each agent is important to enhance their functionality and ensure they can complete specific tasks.

Action

A specific operation performed by an agent using a tool or flow.

Examples:

• “Add this product in size M to the cart.”
• “Cancel order #1234.”
• “Apply coupon SAVE10 if eligible.”
• “Create a follow-up task for sales.”

Think: tools = capabilities; actions = actual steps taken.

Agents perform tasks independently or proactively, automating repetitive and routine tasks to increase efficiency and free up human workers for more complex activities.

Execution flow/flow

A visual or logical sequence that defines what should happen in a multi-step process.

Example flow: Return request

1. Ask for the order ID.
2. Check if the order is within the return window.
3. Ask for the return reason.
4. If eligible, create a return and share label.
5. Update CRM / ticket system

Flows combine: questions, conditions, tool calls, and messages.

Orchestrator

The “brain” that decides:

• Which agent should respond
• Which tool or flow to call
• When to escalate to a human

In a multi-agent or multi-channel setup, the orchestrator routes work to the right agent or process. Orchestrators use data, context, and past interactions to make informed decisions and optimize performance.

Decision-making is a core agent capability, allowing learning agents to reason, plan, and adapt to new situations.

Multi-agent system

An architecture where multiple specialized AI agents work together.

Agents coordinate with other agents to achieve complex workflows, sharing information and responsibilities.

Example:

• Agent A: authentication & account lookup
• Agent B: product questions & recommendations
• Agent C: returns & exchanges
• Orchestrator decides who handles each step of the conversation.

This coordination enables advanced problem-solving and the ability to handle complex workflows.

Routing

Logic that sends a conversation to the right agent or human based on:

• Intent (billing vs order vs product)
• Customer type (VIP vs guest)
• Channel (chat vs WhatsApp vs email)
• Priority or sentiment

Routing happens at the front door of a conversation.

Handoff/escalation

When an AI agent hands a conversation off:

• To a human support agent
• To another AI agent
• To a different team (e.g., billing vs shipping)

Best practice: hand off with full context so the human doesn’t have to ask the same questions again.

3. Knowledge, data & “Grounding” the AI

Knowledge base (KB)

A structured collection of information that the AI can use:

• FAQs
• Help center articles
• Policy pages
• Product documentation
• Internal SOPs

A good KB is organized and maintained so answers stay accurate.

Grounding

Ensuring the AI’s answers are based on your real data, not its own imagination.

• Fetching relevant docs or records
• Combining them into context
• Answering only using that context

Grounding is one of the main ways to reduce hallucinations.

Retrieval-augmented generation (RAG)

A technique where the AI:

1. Retrieves relevant documents or data fragments
2. Generates a response using the retrieved context

• Accurate policy answers
• Product-specific questions
• Up-to-date information without retraining models

4. CX & support metrics in autopilot world

Ticket Deflection / Containment Rate

The percentage of potential tickets that:

• Are handled by AI or self-service
• Do not become fully human-handled tickets

Example: Out of 1,000 support attempts, 400 are fully handled by AI → 40% deflection.

Insightful read: The cost of doing nothing: Why AI adoption is no longer optional in CX.

AI resolution rate

Within AI-handled conversations:

The percentage that are fully resolved by AI without human escalation.

High resolution rate = AI is actually solving problems, generator, not just handing them off.

First response time (FRT)

Time from the customer’s initial contact to the first reply.

AI agents can reduce FRT to near-instant across channels.

Average handle time (AHT)

Average time taken to resolve an issue or conversation.

AI can:

• Shorten AHT for repeatable issues.
• Free humans to focus on complex ones (which may take longer individually, but fewer in number)

CSAT(Customer satisfaction score)

Usually gathered by post-interaction surveys:

• “How satisfied are you with the support you received?” (1–5 stars or similar)

Useful for measuring how customers feel about AI-led vs human-led journeys.

NPS (Net promoter score)

Measures loyalty with a single question:

“How likely are you to recommend us to a friend or colleague?” (0–10)

While broader than support, AI-driven CX can influence NPS by:

• Faster responses
• Fewer frustrating loops
• More proactive help

Escalation rate

Percentage of AI conversations that get handed off to humans.

You want:

• Escalation for the right cases
• But not because the AI is confused or lacks knowledge

Fallback/“I don’t know” rate

How often the AI:

• Admits it can’t answer
• Or routes directly to human because of low confidence

Some fallback is healthy. Too much means:

• Missing knowledge
• Poor flows or tools
• Or overly cautious configuration

How Skara powers Autopilot CX with intelligent AI Agents

At Salesmate, we’ve transformed customer experience automation with Skara AI Agents: intelligent, autonomous teammates that take action, resolve issues, and move work forward across your entire customer journey.

Meet the AI Agents behind truly Autopilot CX

AI Sales Agent: Engages leads instantly, qualifies prospects, and moves deals faster with personalized, always-on conversations.

AI Support Agent: Resolves up to 70% of customer issues on the spot while delivering warm, human-like support across every channel.

AI Booking Agent: Automates meeting scheduling, rescheduling, and reminders — eliminating back-and-forth and reducing no-shows.

AI Lead Qualification Agent: Chats with website visitors, identifies intent, filters spam, and routes high-value leads directly to your team.

AI eCommerce Agent: Boosts conversions with smart product guidance, order tracking, and abandoned cart recovery.

AI Employee Experience Agent: Supports internal teams with HR, IT, and onboarding queries, improving efficiency and employee satisfaction.

5. Safety, guardrails & quality

Guardrails

Rules that limit what the AI can say or do.

Examples:

• Don’t discuss certain topics (legal, medical, etc.)
• Never override specific policy rules.
• Always get confirmation before sensitive changes (address, payment method)

Guardrails can be:

• Prompt-based (instructions)
• Flow-based (logic paths)
• Policy-based (hard constraints in code)

Some essential tasks or qualities—such as ethical judgment, creativity, or empathy—may require human input, and guardrails help ensure these are handled appropriately.

Hallucination

When an AI confidently generates an answer not based on real data.

Examples:

• Making up a policy that doesn’t exist
• Inventing a product feature
• Guessing an order status

Good systems minimize hallucinations via:

• Grounding
• Guardrails
• Tool usage
• Validation checks

Human-in-the-loop (HITL)

Any setup where humans stay in control:

• Approving certain AI actions (like refunds above $X)
• Reviewing conversations for QA
• Handling escalations for sensitive cases

HITL is crucial for high-risk workflows.

Red teaming

Testing AI systems with:

• Edge cases
• Adversarial prompts
• “Tricky” questions

Goal: find failure modes before customers do.

Evaluation / QA

Systematic ways to measure AI performance:

• Test sets (e.g., 100 typical “Where is my order?” scenarios)
• Scoring on correctness, helpfulness, and tone
• Ongoing regression tests when models/prompts change

Agents use evaluation and feedback from past interactions to identify patterns and continuously improve their performance.

6. Channels, journeys & experience terms

Multichannel vs Omnichannel

• Multichannel: You’re present on many channels (chat, email, WhatsApp), but they’re siloed.
• Omnichannel: Conversations and context are shared between channels.

AI agents work best in omnichannel setups where they can see the whole customer history.

Journey/customer journey

The full path a customer takes:

• Discover → browse → add to cart → checkout → post-purchase → repeat purchase

AI agents can be designed to help at each stage:

• Product advisor during browsing
• Checkout assistant at the payment
• Order helpdesk post-purchase

Touchpoint

Any interaction point:

• Website chat widget
• WhatsApp message
• Email reply
• Instagram DM
• Phone call/voice bot

Autopilot CX aims to standardize the experience across touchpoints.

Session

A period of interaction:

• One visit to the website
• One chat conversation
• One WhatsApp thread (depending on your definition)

Sessions may contain multiple messages and actions; AI agents maintain context within a session.

7. AI & cost basics (Tokens, models, etc.)

Token

A chunk of text (word or part of a word) used in AI processing and billing.

• More tokens = more cost
• Long conversations + large context = more tokens

Good implementations optimize:

• How much history is sent
• How often are tools called?
• When to summarize vs keep raw text

Model/LLM

The underlying AI “brain” (e.g., GPT, Claude, etc.) that:

• Understands language
• Generates text
• Follows instructions

Agent platforms typically manage which model is used for:

• Fast vs complex tasks
• Low vs high-cost scenarios

Large language models (LLMs) power the reasoning, planning, and decision-making capabilities of AI agents.

Generative AI models enable advanced agent capabilities, such as multimodal information processing and natural language conversations.

Some AI agents, especially those used in resource-intensive or large-scale applications, can be computationally expensive to develop and deploy.

Prompt/system prompt

The hidden instruction or “job description” that tells the AI how to behave.

Example system prompt snippet:

“You are an ecommerce support agent. Always check the order status via tools before answering. If you’re unsure, escalate to a human instead of guessing.”

Good prompts are:

• Clear
• Concise
• Consistent with flows and policies

Fine-tuning vs Prompt engineering

• Prompt engineering: Carefully designing instructions and examples without changing the model itself.
• Fine-tuning: Training the model on your own data to change its behavior more deeply.

Many CX use cases can be solved with good prompting + grounding + tools, without full fine-tuning.

8. Implementation & Ops terms

Sandbox/Staging

A safe and dynamic environment where you:

• Test agents and flows
• Connect to test/staging versions of your systems.
• Let internal teams play with the AI

Important: never launch a new agent straight to production without sandbox testing.

Playbook

A documented, reusable pattern for solving a business problem.

Examples:

• “30-day playbook to reduce WISMO tickets with AI”
• “Playbook for automating returns & exchanges”

Playbooks help teams roll out building AI agents systematically instead of improvising.

Runbook

Operational steps for handling specific scenarios:

• What to do if a tool fails
• How to temporarily disable an agent
• How to roll back to human-only handling

Change management

All the work around people and process, not just tech:

• Communicating what AI will and won’t do
• Training agents to work alongside AI
• Adjusting KPIs and roles

Change management often determines whether AI adoption succeeds or stalls.

Planning module

A planning module is a core component of advanced AI agents that enables them to break down user requests or business goals into a sequence of actionable steps.

Unlike simple reflex agents that react to immediate inputs, a planning module allows an intelligent agent to look ahead, consider future states, and organize actions to achieve a desired outcome, even when the path is complex or involves multiple steps.

Wrapping up

You don’t need to become a machine learning expert to use AI agents or Autopilot CX in your ecommerce business.

But understanding the key terms helps you:

• Ask better questions
• Challenge vendors when something sounds vague
• Design better experiences with your own team.
• Align everyone around what “Autopilot CX” really means (and doesn’t mean)

Bookmark this glossary and share it with your support, ecommerce, and ops teams. The more aligned everyone is on language, the smoother your AI projects will go.