A chatbot answers questions. An agent does things: it reads a database, calls a webhook, raises an invoice. That difference is where the value is, and where the risk is too.
Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, due to rising costs, unclear value or weak risk controls. It also warns of "agent washing", old chatbots and automation tools renamed as agents (Gartner, 2025).
1. Start with a workflow, not an agent
Anthropic separates workflows, where code decides the steps and the model fills them in, from agents, where the model decides its own steps and tools. Its advice is to find the simplest solution and add complexity only when needed (Anthropic, 2024).
| Level | Who picks the next step | Good for |
|---|---|---|
| Chatbot | The user | Answering from known documents |
| Workflow | Your code | Known steps: classify, extract, route |
| Agent | The model | Varied tasks where steps can't be listed in advance |
2. Give it tools through one standard door
The Model Context Protocol is an open standard for connecting AI applications to external systems (MCP, 2026). Wrapping each database or API as a small MCP server keeps every capability in one reviewable place. Our free MCP server templates show the shape.
3. Grant the least it needs
OWASP lists excessive agency as a top risk for LLM applications. Its fixes: limit the tools an agent can call, limit their permissions, and require a person to approve high-impact actions (OWASP, 2025). Read-only by default; payments and deletions always need a human.
4. Log every action like a transaction
Record what the agent called, with which inputs, and what changed. If you can't replay a day's actions, you can't explain them to an auditor or a customer.
Is your agent ready for production?
Score each 0 (no), 1 (partly) or 2 (yes):
- Need: a simpler workflow was tried and couldn't do the job.
- Tools: each tool is listed, owned and reviewed.
- Permissions: read-only unless a write is essential.
- Approval: money, deletions and messages to customers need a person.
- Audit: every action is logged and replayable.
| Score | Next move |
|---|---|
| 9–10 | Go live on one workflow with a daily review. |
| 6–8 | Close permission and approval gaps first. |
| 0–5 | Build it as a fixed workflow instead. |
Your next step
Write down every action your planned agent could take. Mark each "read", "write" or "irreversible". Ship the read actions first, and put a human approval on every irreversible one.
Sources
- Gartner (2025). Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. 25 June 2025.
- Anthropic (2024). Building effective agents. 19 December 2024.
- MCP (2026). What is the Model Context Protocol?
- OWASP (2025). LLM06:2025 Excessive Agency. OWASP Top 10 for LLM Applications.