What is an AI agent?

The short answer

An AI agent is software that can take actions on your behalf, not just answer questions. It uses an AI model (Claude, GPT-5, Gemini) to understand what you want, then uses tools to actually do the work — browsing the web, sending emails, querying databases, running scripts, calling APIs.

The key distinction from a chatbot: agents take actions, chatbots just respond.

How it's different from ChatGPT

ChatGPT can tell you how to draft a follow-up email to a prospect. That's chatbot behaviour. An AI agent drafts the email, finds the prospect's contact details, sends it, and logs the activity in your CRM — without you touching a keyboard. That's agent behaviour.

The underlying language model is often the same. What changes is the layer of tools and permissions wrapped around it. The platform you pick determines what actions the agent can take, what data it can access, and when it has to ask a human first.

How an AI agent actually works

Three layers, every time:

• The model — Claude, GPT-5, Gemini. Does the reasoning. Interprets your request, decides what steps to take, drafts the language.
• The tools / harness — connects the model to the world. CRM access, email, web browser, file system, custom APIs. Lindy, OpenClaw, n8n, Claude Code are different kinds of harnesses.
• The workflow layer — decides when the agent runs, what permissions it has, when to escalate to a human. Some platforms make this explicit (Paperclip's orchestration); others bury it in the agent definition.

Pick a platform that fits your needs at each layer. Most teams need three of those layers wired together; a few need all three plus orchestration.

What AI agents are good at right now

• Customer support ticket triage and first-response drafting
• Sales follow-up sequences and lead enrichment
• Document Q&A against internal knowledge bases
• Meeting prep, recap notes, action item extraction
• Code generation, refactoring, and review (Claude Code, Cursor)
• Voice calls — outbound sales, inbound support, appointment booking (Retell, Vapi)
• Spreadsheet automation and scheduled reports
• Routine ops tasks that previously required a human to click through a workflow

What AI agents are still bad at

Anything requiring consistent fine judgement. AI agents still make errors, take wrong turns, and occasionally hallucinate facts — especially in novel situations. Don't deploy them unsupervised for:

• Client communications with significant stakes
• Financial decisions without a human review step
• Anything where a mistake would be irreversible
• Tasks that depend on real-time emotional intelligence

The pattern that works: AI agents as force multipliers for a human, not autonomous replacements. Handle the first 80% of the task; keep a human in the loop for the final call.

How to pick one

Start with the use case, then pick the platform. The categories that matter:

• No-code SaaS (Lindy, Relevance AI) — fastest to working agent, no developer needed
• Workflow builders (n8n, Stack AI) — flexible, requires technical capacity
• Coding agents (Claude Code, Cursor) — for engineering work
• Voice AI (Retell AI, Vapi) — for phone and web voice conversations
• Open-source harnesses (OpenClaw, Hermes) — for technical operators who want full control

Or just run our five-question Agent Picker — it asks the right questions and matches you to one of 22 platforms we've tested.