Every SaaS demo in 2026 has the word "agent" in it. Most of those products are chatbots wearing a costume. Here's the honest distinction.
The shorthand
Chatbot: input → response → input → response. The conversation is the product.
Agent: input → plan → tool calls → file writes → API calls → response. The output of work is the product. The conversation is just the steering wheel.
The five things only agents do
- Spawn other processes. An agent can run a shell command, install a package, call an API, hit a database. A chatbot cannot.
- Persist memory across sessions. An agent remembers what you set up last week. A chatbot starts every session blank.
- Run on a schedule. An agent can be told "every Monday at 9am, do X" and it does. A chatbot needs a human to push the button.
- Coordinate with other agents. An agent can dispatch sub-agents in parallel — content scout, email manager, PR reviewer — and synthesize their outputs.
- Write to your real systems. An agent updates your CRM, fills your spreadsheet, books your appointment, sends your email. A chatbot just tells you how you would do that.
Why this matters for small business
The labor savings only show up when software starts doing work, not describing work.
If your "AI assistant" can write you a great cold email but you have to copy-paste it into your CRM, you saved maybe 10% of the work. If your "AI agent" reads the prospect's LinkedIn, drafts the email in your tone, sends it via your Gmail, logs the activity in your CRM, and follows up in 3 days if there's no response — that's 80% of the work.
Both could be called "AI" in a marketing deck. Only one is actually replacing labor.
How to tell which one you're being sold
Three questions to ask any vendor:
- "Show me the agent doing one task end-to-end without a human in the middle." If the demo always has a human approving at every step, it's a chatbot with extra prompts.
- "What does it have access to?" If the answer is "the chat," it's a chatbot. If the answer includes "your filesystem, your CRM, your calendar, the internet, my MCP servers" — it's an agent.
- "What does it do when I'm asleep?" If the answer is "nothing," it's a chatbot. If the answer is "follows up on yesterday's emails, drafts tomorrow's content, runs the morning report" — it's an agent.
The trade-off you should know about
Agents that actually do things have a higher blast radius when they screw up. A chatbot that hallucinates a wrong answer wastes 30 seconds. An agent that hallucinates a wrong answer might book a wrong meeting, send a wrong email, or charge a wrong card.
This is why every agent on ClawHQ has rental restrictions, pre-action approval gates, and per-domain permission scoping. It's also why we explicitly mark which container tier each agent runs on — so you know its capability ceiling before you turn it loose.