“Agent” is used loosely. In practice there are three levels, and most real business value today sits in the first two.
1. An assistant (chat)
You ask, it answers. ChatGPT, Claude, Gemini and Copilot are assistants. They are brilliant for drafting, summarising and explaining, but they only act when you ask and only do what you then copy and paste.
2. A workflow with AI steps
A fixed sequence of steps, such as “when an email arrives, extract the details, add them to the CRM, draft a reply”, where AI handles the reading or writing. The path is decided in advance by you. Tools like Zapier, Make and n8n build these. They are predictable and easy to test.
3. An agent
The AI decides which steps to take, uses tools (search, a spreadsheet, your calendar) and keeps going until the task is done. Anthropic’s guidance describes agents as systems where the model “dynamically directs” its own process and tool use, compared with workflows that follow predefined paths. Agents are more flexible but also less predictable, slower and more costly, so they need tighter limits and checks.
Which do you need?
- The steps are the same every time: build a workflow.
- The steps change depending on what it finds, and mistakes are cheap to fix: an agent may be worth trying.
- Mistakes are costly (money, customers, regulation): keep a person approving each action, whatever you build.
Anthropic’s own advice is to start with the simplest solution and only add complexity when it clearly improves results. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027 because of costs, unclear business value or weak risk controls. That’s why every entry in our library includes a way to measure return.
Sources
- Anthropic, "Building effective agents" Vendor guidance
- Gartner press release, 25 June 2025: over 40% of agentic AI projects predicted to be cancelled by end of 2027 Analyst prediction
- Zapier AI (product page) Vendor page