Glossary

What is an AI agent?

A clear, modern definition.

An AI agent is a software system that perceives information about its environment, reasons about a goal, uses tools to act on the world, and keeps working across multiple turns until the goal is met or it hands off to a human.

That single sentence covers what separates an AI agent from older tooling. It's also the thing most "AI chatbot" marketing gets wrong.

A chatbot answers a question. An AI agent answers the question, looks up your account, drafts the refund, asks for confirmation, files the ticket in your CRM, sends the customer a WhatsApp message, and tells the support team why it escalated. Same conversation, different category of system.

The shortest possible taxonomy

TermWhat it doesExample
LLM (Large Language Model)Predicts the next token of textGPT-4.5, Claude Opus 4.6, Gemini 2.5
ChatbotFollows scripted rules or treesPre-2022 customer support widgets
Conversational AIFree-form dialog using an LLM, but answers only"Ask our docs anything" search bots
AI agentFree-form dialog plus tool use, memory, multi-step planning, and autonomyA bot that takes the booking, posts to your CRM, and follows up on WhatsApp
Virtual assistantMarketing umbrella covering all of the above"Siri", "Alexa", "your AI assistant"

These differences are not academic. They show up in cost per resolution, in what your team has to do manually, and in how much of a customer journey can run without a human in the loop.

Anatomy of an AI agent

Most modern agents share five components. Whether you build one yourself or buy one off the shelf, this is what's running under the hood.

  1. 01

    Perception

    The inputs the agent can see: a web chat message, an email, a WhatsApp DM, an API call, a sensor reading. The richer the perception layer, the more useful the agent.

  2. 02

    Reasoning

    Usually an LLM that interprets the input, decides what to do next, and chooses which tool (if any) to call. This is where "tool use" or "function calling" lives.

  3. 03

    Memory

    Short-term context for the current conversation, plus optionally long-term memory across sessions: preferences, history, account state.

  4. 04

    Tools

    The actions the agent can take. Look up a customer record. Search a knowledge base. Send an email. Book a Cal.com slot. Update a Salesforce contact. Post a WhatsApp message via the Business API. Tools are what make an agent more than a Q&A bot.

  5. 05

    Evaluation and handoff

    The rules that decide when the agent did its job, when it failed, and when to route to a human. The good ones know their limits and hand off confidently when needed.

What can AI agents actually do today?

Not everything. The honest answer matters here, because the gap between "AI agent demo" and "AI agent in production" is where most projects die.

Things that work well today

  • Customer support resolution that handles 40 to 70 percent of incoming questions without a human, depending on industry. Returns and refunds, order status, account changes, password resets, common how-tos.
  • Lead qualification: the agent interviews a website visitor, scores the lead, pushes the qualified ones to your CRM, and books a demo on the calendar.
  • Appointment booking, especially on WhatsApp where the booking flow is naturally conversational.
  • Multi-channel triage across web chat, WhatsApp, and email, with consistent tone, state, and handoff rules.
  • Drafting and review (internal notes, summaries, suggested replies for human approval) is one of the fastest wins inside a team.

Things that are still hard

  • Long-horizon autonomous work without human checkpoints. Agents drift, especially over hundreds of steps.
  • Anything where a wrong action is unrecoverable: sending money, deleting records, posting publicly without review.
  • Highly nuanced negotiation: pricing edge cases, complex billing disputes, sensitive complaints.

The pragmatic pattern: design the agent for the 60 to 80 percent of work it handles cleanly, define explicit handoff for the rest, and measure both rates honestly.

Real-world use cases

A few concrete examples, drawn from how teams actually deploy agents in 2026.

  • A real estate agency uses an AI agent on its website to qualify a lead, schedule a property viewing on the listing agent's calendar, and update the CRM with the property of interest, all in one conversation.

    See the real estate use case

  • A Shopify store runs an agent on web chat and WhatsApp that answers product questions, handles return requests, and recovers cart abandoners with a personalised follow-up.

    See the e-commerce use case

  • A B2B SaaS runs an agent that triages free-trial users, answers product questions during onboarding, books a call with the customer success team for users who hit complexity, and tags their HubSpot record with the conversation summary.

    See the SaaS B2B use case

  • A car dealership uses an agent to answer model questions, capture interest, and book test drives, taking the load off the showroom phone line.

    See the automotive use case

  • A property management firm uses an agent to handle 85 percent of routine tenant and landlord enquiries automatically, capturing maintenance tickets in structured form.

    See the property management use case

Each of these is a measurable outcome, not a "look at our cool chat widget" demo. The shift from chatbot to agent is the shift from impression to outcome.

How AI agents are deployed

There are four common deployment surfaces. Most production agents end up using two or three of them.

01

Web chat widget

A script tag on your website. Fastest channel to ship and the easiest to measure.

02

Messaging apps

WhatsApp Business API is the dominant channel in Europe, MENA, and Latin America. Facebook Messenger, Instagram DMs, and Telegram fill in.

03

Email

Agents that triage and answer support tickets at the inbox layer.

04

API or embedded

The agent runs inside another product or workflow, called via API, with no user-facing chat at all.

The right answer is rarely "one channel". A good agent platform makes the same agent available across surfaces and keeps state coherent. What the customer told the bot on WhatsApp last week is still in context when they open the website chat today.

What to look for when choosing an AI agent platform

If you're past the "what is this thing" stage and into "which one do I buy", here's a short checklist.

  • Tool use is first-class, not an afterthought.

    A platform that only does Q&A on your docs is a chatbot, not an agent.

  • Native channel integrations.

    Especially WhatsApp Business API if you operate outside the US or UK.

  • Memory and CRM sync.

    The agent needs to know who the customer is across sessions, and it needs to write back what it learned.

  • Defined handoff.

    You can configure exactly when the agent stops trying and routes to a human.

  • Compliance posture you can verify.

    For EU teams, this means EU hosting, a real DPA, a deletion concept, and AI Act transparency baked in.

  • Honest pricing.

    Per-message credits make budgeting unpredictable. Flat-rate plans are easier to plan around at SMB scale.

  • Multi-language support.

    And not just translation. Ideally the agent reasons natively in the user's language, not via translate-then-think.

For a side-by-side look at how the major platforms stack up on these criteria, see our comparison hub.

Frequently asked questions

What's the difference between an AI agent and a chatbot?

A chatbot follows pre-defined rules or scripts and answers questions. An AI agent uses an LLM to reason, calls tools to do real actions (look up records, send messages, book appointments, update systems), and continues across multiple turns until a goal is met or it hands off to a human. Chatbots reply. Agents complete tasks.

Are AI agents the same as ChatGPT or Claude?

No. ChatGPT and Claude are large language models. They are the reasoning component of an agent. An AI agent uses an LLM, but adds tool use, memory, integration with your business systems, and goal-directed behaviour. Asking ChatGPT to recommend products is not an agent. A system that asks customers about their needs, looks up your inventory, drafts an order, and pushes it to Shopify is an agent.

Can AI agents replace human support staff?

Not entirely, and the framing is wrong. The right pattern is "agents handle routine and repeatable, humans handle nuance and high-stakes". A well-deployed agent typically resolves 40 to 70 percent of inbound questions without a human, freeing the team to focus on the cases that benefit most from human judgement.

How long does it take to deploy an AI agent?

For a no-code platform like the BitPalm Agent Platform, a working agent can be live in 10 minutes. Knowledge base uploaded, integrations connected, basic escalation rules defined. Production-ready means another few days of iteration on edge cases, persona, and handoff logic. For developer-first platforms like Botpress, expect weeks of engineering work for the equivalent capability.

Are AI agents GDPR-compliant?

It depends entirely on the platform. The relevant questions are: where is the data hosted, is there a real Data Processing Agreement, can you delete user data on request, does the platform meet the EU AI Act's August 2026 transparency requirements (users must be informed they're talking to AI), and is there a Zero-Data-Retention policy with the underlying LLM provider. EU-hosted platforms with explicit DPAs, like BitPalm hosted in Frankfurt, are a safer starting point than US-hosted tools that rely on Standard Contractual Clauses post-Schrems II.

Further reading